SilentRoom Journal

Anthropic Is Watermarking Every Text It Touches. It Isn't for You

editorial@silentroom.media (Nick Rogers) — Thu, 08 Oct 2026 11:30:00 GMT

The deepfake that was, and the one that wasn’t

In August 2006, a Reuters photographer named Adnan Hajj filed a picture from Beirut: thick black smoke rising over the city after an Israeli airstrike. Within hours of its publication, American blogger Charles Johnson grew suspicious. The smoke looked wrong to him, as if someone had cloned it in a photo editor. On Johnson’s tip, Reuters investigated, then pulled more than nine hundred of Hajj’s photographs, fired him, and let his editor go as well. The agency tightened its digital photography standards after publishing its complete findings in January 2007: now it allows only cropping and exposure adjustments, the same things a photographer could do in a darkroom.

In March 2021, police in Pennsylvania arrested Raffaela Spone for allegedly making deepfake videos of her daughter’s cheerleading rivals. She had, the story went, fabricated photos and footage of the girls drinking, vaping and in states of undress. The detective on the case admitted that he had spotted the fakery "by eye." The story swept through the national press overnight, but prosecutors dropped the deepfake claims by May after the only video that was made public turned out to be real.

"Her reputation right now is less than mud. They have ruined her life, there’s no question about it. … On Twitter, they have already convicted her. She’s always going to be labeled the 'deepfake mom' or … a 'criminal mastermind.'"

— Robert J. Birch, attorney for Raffaela Spone, quoted in: Drew Harwell, "Remember the 'deepfake cheerleader mom'? Prosecutors now admit they can’t prove fake-video claims,", The Washington Post, May 14, 2021

The response to a suspected deepfake was equally confident in both these stories, even though a forgery only existed in one. A blogger might know more about photo editing than a detective, and Reuters investigated the circumstances before taking action, but human judgment prompted each case.

How would each incident have gone with a tool that could reliably spot AI traces in photos or videos?

That tool now exists: on August 14, 2026, Anthropic announced that all text Claude processes will carry an invisible watermark. We need to understand what that change does and who it serves.

You now have a watermark you can’t read…

A watermark in a text isn’t an author’s signature. Anthropic is blunt about this:

"A watermark only helps test whether Claude might have produced or processed the content. It doesn’t say anything about ownership or authorship, and doesn’t change a user’s rights under our terms."

— Anthropic, How Claude’s text watermark works, August 14, 2026

The mechanism rests on SynthID, a method developed at Google DeepMind. When the model generates text, it doesn’t pick each next token at random. It follows a pattern set by a hidden key, and the detector later checks whether the sequence of choices matches those instructions. Anthropic offers an analogy: picture a game of Monopoly in which the digits of pi stand in for the dice. The result looks like chance, but it can be reproduced and verified.

Anthropic itself lists what SynthID can’t catch:

Short texts, where the signal is too faint to show up.
Factual passages where the next word is the only one possible, like "Isaac Newton’s most famous work was called Principia Mathematica."
Edits to someone else’s prose. If Claude corrects only the grammar, the watermark has nowhere to hide but in the corrections, and there may be too few of them.

Finally, the limitation that matters most:

"It cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'"

— Anthropic, How Claude’s text watermark works, August 14, 2026

None of these statements mean Anthropic’s dodging responsibility. The company’s representatives give an honest account of how SynthID works and what it can actually do.

An editor needs to know where the line falls: did the author hand the machine only a proofreading pass, or did Claude write the entire draft, first letter to last? A journalist has to establish where a document came from. A scholar wants to shield a paper from the charge of automated authorship. A student wants to prove that the essay is their own.

The watermark gives them none of this. It can’t tell structural generation from a light polish. It doesn’t vouch for your authorship, and it does nothing for your work. It merely records that an algorithm once touched the text. Even that trace is off-limits to you, since you almost certainly don’t have a scanner.

…because such marks exist to make a prosecutor’s life easier, not yours

The history of labeling doesn’t begin with algorithms.

In the mid-1980s, Xerox built a hidden feature into its color laser printers: every page came out stamped with a constellation of tiny yellow dots, invisible to the naked eye, officially to deter counterfeiters. The dots encoded the printer’s serial number and the date and time of printing. In October 2005, the Electronic Frontier Foundation exposed the system and published a tool for decoding it.

In May 2017, Reality Winner, a 25-year-old NSA contractor in Augusta, Georgia, printed a five-page report on Russian interference in the American election and sent it to The Intercept. Before publication, the editors handed a scan of the document to the government to authenticate it, not knowing that the dots had survived the scan. The dots recorded a model 54 printer with serial number 29535218, and a moment: May 9, 2017, 6:20.

"The document leaked by the Intercept was from a printer with model number 54, serial number 29535218. The document was printed on May 9, 2017 at 6:20. The NSA almost certainly has a record of who used the printer at that time."

— Robert Graham, How The Intercept Outed Reality Winner, Errata Security, June 5, 2017

Winner was arrested on June 3, two days before The Intercept published. A tool built to fight counterfeiting had been turned on a whistleblower.

This is a structural feature of any labeling system: it always serves whoever controls the verification. The stated purpose and the actual use rarely coincide, because the labeling tool ends up in the same hands as the detector.

Yet this protection collapses at a simple paraphrase

The tool is technically fragile, too. Google DeepMind open-sourced the SynthID code in October 2024. In August 2026, Anthropic rolled out its watermark with no regional restrictions at all.

How durable is the trace? A study called Watermark under Fire (Liang et al., accepted at EMNLP 2025) put SynthID through its paces. On untouched text, detection accuracy is 99.8%. Copy a passage straight out of the Claude window, and the detector will catch you almost every time. Under moderate paraphrasing, however, accuracy drops to 49.8%, no better than a coin toss. This is no fluke; it’s a general property of the method.

Meanwhile, open-weight models such as Llama and DeepSeek, which for architectural reasons carry no watermark, can run locally. If even one major player stays outside the system, the structure means this model becomes a way around the entire control apparatus, even if the creators never intended that.

Why AI providers were in no hurry to press the button

Technical fragility is only half the problem. The other half is a deliberate corporate choice. At OpenAI, a finished watermarking tool has been sitting in a drawer since at least the summer of 2023.

"It’s just a matter of pressing a button."

— one of the people familiar with the matter, quoted in: Deepa Seetharaman & Matt Barnum, "There’s a Tool to Catch Students Cheating With ChatGPT. OpenAI Hasn’t Released It.", The Wall Street Journal, August 4, 2024

The button was never pressed. In April 2023, OpenAI ran two surveys. The first, among the general public, found people favoring watermarks by four to one. The second, among ChatGPT’s own active users, told a different story: 69% feared false accusations, and 30% said they would switch to a competitor that didn’t use watermarks. Internal company documents obtained by The Wall Street Journal record where that contradiction left the company:

"Our ability to defend our lack of text watermarking is weak now that we know it doesn’t degrade outputs."

— internal OpenAI documents, quoted in: Deepa Seetharaman & Matt Barnum, "There’s a Tool to Catch Students Cheating With ChatGPT. OpenAI Hasn’t Released It.", The Wall Street Journal, August 4, 2024

As of August 2026, OpenAI’s button remained unpressed. Anthropic pressed its own and rolled out watermarking worldwide, but the detector is available only to regulators and a narrow circle of organizations. The public’s appetite for catching AI red-handed still hasn’t gone anywhere, however, and crude knockoffs have moved quickly to fill the vacuum the industry left.

Leaving you alone with crude substitutes that punish bad English

Turnitin and GPTZero don’t read watermarks. It’s worth being precise here: they don’t check whether a text carries a SynthID signal or any other mark. They analyze the statistical patterns of a text and compare them with how people write—or, to be exact, with how people usually write when English is their first language.

In 2023, researchers at Stanford tested seven AI detectors on two sets of human-written texts: essays by American eighth-graders and essays from the TOEFL exam, which tests non-native English speakers:

"These detectors consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified."

— Weixin Liang et al., "GPT detectors are biased against non-native English writers", Patterns, 2023

In 61.22% of cases, the detectors branded TOEFL essays the work of an algorithm, while nearly always correctly identifying the American schoolchildren’s essays as human. At least one detector perceived AI in 97.8% of TOEFL texts.

The detectors stumble even on their own terms: in another study, Debora Weber-Wulff’s group tested fourteen detectors and didn’t get more than 80% accuracy from any of them. Running the AI-generated text through an automated paraphrasing tool dropped average accuracy to 26%. (For a closer look at how detectors work and who pays for their false alarms, see "Are AI Detectors Accurate? How They Get It Wrong and Who Pays.")

The Stanford authors spelled out the consequence:

"Paradoxically, to evade false detection as AI-generated content, these writers may need to rely on AI tools to refine their vocabulary and linguistic diversity."

— Weixin Liang et al., "GPT detectors are biased against non-native English writers", Patterns, 2023

That is the researchers’ conclusion, not a recommendation.

In 2023, a professor at Texas A&M refused to grade his graduating seniors’ final papers. To check them, he fed the papers to ChatGPT and asked: "Did you write this?" ChatGPT answered "Yes" without checking any watermark. The professor later reconsidered his decision.

The pros get the real tool

Anthropic’s Detection API isn’t a public feature. A select group of professionals gets it at launch: regulators, law enforcement, newsrooms, fact-checkers, independent researchers and educational institutions in the EU. Teachers with a Turnitin account can’t use it, and neither can online readers.

When a studio sends a new film to reviewers before its premiere, it adds a unique mark to each copy that only those in the company can see, so it can trace any leaks back to their source. Anthropic’s Detection API works on exactly the same principle: it lets those with access and authority examine a text after the fact.

This tool’s existence isn’t a problem. The trouble begins when people think public AI detectors like Turnitin and GPTZero can do the same job.

In June 2023, The Washington Post reported on teachers across the country who were using AI detectors to check student work. Soheil Feizi of the University of Maryland, one of the leading researchers in the field, argues that a detector used on students should produce no more than 0.01% false positives. Is that figure within reach?

"At this point, it’s impossible. And as we have improvements in large-language models, it will get even more difficult to get even close to that threshold."

— Soheil Feizi, University of Maryland, quoted in: Geoffrey A. Fowler, "Detecting AI may be impossible. That’s a big problem for teachers.", The Washington Post, June 2, 2023

Feizi was speaking of detection without watermarks. Confusing public AI detectors with the Detection API and similar tools is costly, and the people who mix them up aren’t the ones who’ll pay.

The rules will appear after you’ve paid

Reuters didn’t invent a special algorithm for catching doctored photographs. It simply sat its editors down with its photographers and agreed on the line between acceptable correction and falsification. The process is slow and awkward, and it comes with no guarantee, but it has held to this day.

A watermark doesn’t draw that line. It only records that whoever created a text used an algorithm for at least part of it, and only shows that to those who have the detector, the access and the authority. For a regulator, it’s a useful tool. For an editor at a publishing house or a schoolteacher, it doesn’t exist.

Raffaela Spone walked out of court with three years of probation, convicted not of deepfakes but of harassment using anonymous phone numbers. By that time, the headlines had already circled the globe. The detective who had spotted the supposed forgery "by eye" never explained what his certainty rested on.

Norms are always rewritten after someone has paid the price. Reuters wrote its own in 2007, in the wake of Hajj. The AI industry is writing its own now, after it became clear that no one could do without them, and before it becomes clear who exactly they’re for.

Too Hot, Too Cold, and Just Right: Why AI Picked Up the Rule of Three and How Humans Can Deal With That

editorial@silentroom.media (Isabel Kunkle) — Tue, 06 Oct 2026 11:30:00 GMT

"Rule of three: AI groups ideas in threes to make writing effective, memorable and satisfying."

— Louis Barclay, "Superficial Intelligence: AI Writing"

"Crush your enemies, see them driven before you, and hear the lamentations of their women."

— Conan the Cimmerian, Conan

Not unlike the much-slighted em dash, the rule of three has become a clue that AI wrote text — not definite proof, but enough to raise suspicions. Like the em dash, it shows up because we trained AI on human text (further reading: "The Em Dash: A History of a Mark Nobody Used — Until Everyone Did"). It's giving us what it thinks we want. When it follows the rule of three, though, it's going back further than the sixteenth century, drawing from some of the most basic human concepts.

Here's why we love threes (and why AI does too), and how obvious LLM use differs from solid human construction.

It's a magic number

"Three Is a Magic Number" (Schoolhouse Rock!, 1973)

That song should be firmly lodged in your brain now. You're welcome.

1970s education has something of a point. Threes show up everywhere: the minimum number of angles in a polygon, the minimum points for a flat plane (and thus any structure that stands up on its own), the smallest number of people needed to make a decision without a stalemate, and so on. Three means two sides of an equation and the result, whether that's math or, as in the song, human reproduction.

Three also establishes repeatability — as Fleming's Auric Goldfinger said, "once is happenstance, twice is coincidence, three times is enemy action," and under federal law, a killer isn't "serial" until the third victim (the FBI's working definition has required two victims since 2005). Along those lines, you have to have two matching items to set up a third as different, which is why fairy tales often have three sons on quests or three nights of celebration, and jokes feature a minister, a priest, and a rabbi.

One more thing: three can signify escalation, whether that means the items get larger, more important, or more complicated. Conan's saying fits here, going from simple and direct to a bit abstract and assuming a bunch of things, like that your enemies have women and those women will lament. So does Saint Paul's point about "faith, hope, and charity," if you agree that "the greatest of these is charity."

Three strikes

"Three shall be the number thou shalt count, and the number of the counting shall be three."

— Monty Python and the Holy Grail

"Our three weapons are fear, surprise, and ruthless efficiency... and an almost fanatical devotion to the Pope."

— Monty Python, "Spanish Inquisition" sketch

Now I'm going to contradict myself: most phenomena don't actually come only in threes, or at least not naturally. Sure, you have morning, afternoon, and night... but what about dawn and twilight? Where does "night" start, or "morning," for that matter? A man and a woman may have a little baby, but they may also have twins. Because three is pretty important in math and math-adjacent activities, it also makes a decent categorization system, but all systems have edge cases, blurry lines, and exceptions when dealing with the messiness of life outside math.

Remembering that is one way to distinguish human rule-of-three use from the way it works in AI. Human writers (good ones, at least) turn to triads to make a point or set up a structure, but won't fall back on them all the time. They'll use longer lists or vaguer statements when the situation seems appropriate, or they won't bother grouping ideas at all.

AI doesn't share this discretion. It often uses triads without any particular need to do so. For instance, in the Gettysburg Address, Lincoln says "we can not dedicate — we can not consecrate — we can not hallow — this ground": he's describing successively more intense levels of honor, from the purposeful to the sacred. On the other hand, do you really need to say that a company is "innovative, rule-breaking, and impactful"? Don't "innovative" and "rule-breaking" mean the same thing? (Don't get me started on "impactful," in general.)

Because LLMs don't evaluate the need for the rule of three the way humans do, its simple frequency can be a clue. SlopDetector says more than one triad every 200 words or so is a pretty good sign that there's a machine behind the writing. It's not a hard-and-fast rule, but it helps, particularly if the writing also includes features like corporate phrasing, em dashes, or repetitive sentence structure, all human habits that AI has picked up and inflated.

Next steps

For readers trying to spot AI? Well, once is happenstance, twice is coincidence, and three times means you at least want to do more research. Don't jump to conclusions, though. Neither Conan nor Saint Paul would appreciate being mistaken for a computer. (I'm not sure about Goldfinger.)

For writers? LLM paranoia is going to be a part of life for a while yet, unfortunately, and there will always be those who see AI in every sentence. So it goes: anyone who's been in the business for a while knows that not all readers are reasonable, and they'd find another issue if AI wasn't on the table.

All we can do is our best. With triads, that means slowing down and examining your writing. Ask yourself why you want a group of items and whether there really are three and only three that qualify.

In other words, think about your prose. Whether you're trying to distinguish yourself from AI or not, that can only be a good thing.

Masterwork or Mass Product?

editorial@silentroom.media (Mary Bush) — Wed, 30 Sep 2026 13:30:00 GMT

1. It's happening now

Six months ago, I first came across the videos of Kelly Boesch, a Los Angeles–based artist and songwriter. If you haven't encountered her work yet, it is well worth seeking out.

Take her video for Not Made for the Cage. It's an anthem to freedom and to being different, a rebellion against being forced to fit a single mold. Music and visuals weave together into living dreams of free human souls straining against the cage of the ordinary world. In these short, surreal videos, giant flowers give way to alien landscapes, and you simply can't look away.

Boesch spent twenty years in design — including at IMAX, the company behind the technology and films made for giant cinema screens — and today she works alone, making art with AI. She creates images in Midjourney, animation in Runway, Pika, and Veo, and music in Suno; she writes the lyrics herself. The results are looping videos with no beginning and no end, like paintings brought to life. They rack up around 60 million views a month.

The gatekeepers that used to stand between Boesch's work and her audience have vanished. Galleries and selection committees no longer decide what the public can see. All that remains is the subscription and the feed.

2. The same tool

Of course, easy access to AI tools more often yields an entirely different kind of content. One of the most popular AI channels on YouTube, for example, is India's Bandar Apna Dost ("Monkey, My Friend"). It features a strikingly realistic AI-generated monkey that keeps landing in everyday human situations. Since its launch, the channel has drawn 2.4 billion views. The video-editing platform Kapwing estimates its potential earnings at around $4.25 million a year.

There are hundreds of content factories like it. After screening 15,000 popular channels, Kapwing's analysts found nearly 300 accounts filled with nothing but AI slop. Together, the videos on these channels draw 63 billion views and roughly $117 million in potential ad revenue a year. On the scale of the global market, the figure isn't staggering, but, as journalist Max Read told The Guardian, such channels are most often run from Ukraine, India, Kenya, Vietnam, Nigeria, or Brazil, where even a tiny slice of that revenue amounts to a serious income.

Boesch and the makers of the AI monkey use the same tools, yet viewers spot the difference between their work instantly.

3. Artificial selection

The rules of the game are still taking shape. Every platform and every country has its own regulations, but none yet distinguishes a masterpiece from trash: they judge anything produced with AI the same way. The main challenge for artists working in this new genre is simply reaching an audience.

YouTube CEO Neal Mohan pledged to crack down on low-quality AI content in January 2026. Soon afterward, 16 channels from Kapwing's list (which had jointly amassed 4.7 billion views) disappeared or were stripped of their videos. YouTube still permits AI, however; its rules withhold monetization only from content that is "mass-produced, generic, repetitive, or manipulative."

The site never explained exactly where that line falls. After the purge, ordinary faceless channels — where the creator never uses AI, just doesn't appear on camera — began complaining of collapsing reach. The monkey, meanwhile, carried on.

It was on Bandcamp that Boesch learned what anti-AI policies mean for an artist. In January 2026, the platform for independent musicians banned music made wholly or in substantial part by AI, and one morning Boesch's catalog simply vanished, giving her no warning, no explanation, and no way to reach the thousands of people who had already purchased her songs.

Technically, the ban followed the rules: a machine generated Boesch's sound. It didn't matter that the lyrics, the visuals, and the mix were her own. The rule looks at the method, not the artist.

Distribution once merely delivered the work; critics and audiences judged it. Now distribution itself determines whether art reaches the public at all. At Bandcamp, no one bothered to weigh the worth of Boesch's songs. They just removed them under the platform's revised rules, with no regard for the choice of the many people who had bought them.

Platforms aren't alone in these sweeping decisions. In Europe, Article 50 of the EU's AI Act has applied since August 2, 2026: AI-generated content must be labeled, and after that it's up to the viewer.

Labeled slop is legal in Europe. In China, the state decides for itself: in 2026, regulators removed more than 5.6 million pieces of AI content deemed illegal or in violation of the rules.

Video platforms look for a face on screen; governments demand labels or order removals. Both are mere technicalities. Nobody cares about the artist.

4. The institutions' choice

For curators and art institutions, AI art is long since old news: it's appeared at biennials and major galleries for years.

This summer, the Palazzo Cini in Venice showed new paintings by David Salle, one of the 1980s' foremost American painters. He trained his own model on his "tapestry" paintings of 1989–1991, then worked over its output in oil and charcoal. Salle's explanation is straightforward: the material had been waiting a long time, and the machine let him move quickly.

Holly Herndon and Mat Dryhurst — she American, he British — live and work in Berlin. They trained a model on distorted images of Herndon herself and presented it at the 2024 Whitney Biennial in New York. No matter the prompt's specifics, the machine returned an uncanny version of Herndon. In this way, the artist herself decided how AI models would see her.

Refik Anadol, born in Istanbul and based in Los Angeles, trained a model directly on the collection of New York's Museum of Modern Art (MoMA); his project Unsupervised played in the museum's lobby for nearly a year.

In all three cases, the machine works with material the artist controls: their own paintings, their own face, or a museum's collection. For the curator, that answers the central question: whose work is this? Here the model seems to extend the artist's hand.

Artists like Boesch take a fundamentally different path. They have no custom-built model of their own; they work in apps available to millions. Institutions view this with suspicion: it seems that a mass-market tool yields a mass-market style.

Therein lies the Boesch paradox: she achieves complete distinctiveness precisely where the tool pulls everything toward the average. Among millions of generated clips, her work is unmistakable from the very first frames.

What's more, biennials and major museums plan exhibitions two to three years in advance. What hangs in Venice today was selected back in 2024, long before powerful video generators like Google Veo 3 appeared. If, a couple of years from now, someone wants to show what cutting-edge AI art looked like today, they won't find it in museum storerooms but in social media feeds.

5. The lesson of photography

The argument about accessible tools has been made before. In 1859, Charles Baudelaire called photography a refuge for failed painters. In 1888, Kodak released its first consumer camera, with the slogan "You press the button, we do the rest." Its 1900 Brownie cost only a dollar, truly letting anyone take a photo. For decades, this remained the chief argument against photography as art: if anybody can press the button, where is the artist?

Over time, though, the world learned to tell a photographer from someone who simply owns a camera. Alfred Stieglitz spent decades fighting for photography to be recognized as art, and in 1940 MoMA opened the world's first museum department of photography.

Today, cameras are more accessible than ever. Every phone has a lens, and everyone takes photographs, but still only a few make art.

With AI tools, we haven't reached that stage yet. The new digital art still lacks any shared framework for telling an endless flood of generated cookie-cutter content apart from the singular, one-of-a-kind work of a master.

6. How to spot the artist

Every day we see audiences single out strong AI artists unerringly, without any methodology at all. Their choices reveal the criteria they use: the same ones that apply to any other art form.

The first is recognizability. The monkey from Bandar Apna Dost is recognizable from the first second, but as a template. I recognize Boesch's videos before I see her name.

The second is the artist's choices. A prompt sets a process in motion, but the result becomes a work of art only after selection, editing, and many other decisions.

The third is consistency. A unified style and signature carry through from one work to the next, rather than dissolving into scattered lucky breaks.

None of these three criteria has anything to do with tools. All three are about the artist's presence, taste, and ideas. Unfortunately, no platform filter can pick those up.

You can hide your face, but you can't hide your voice. Perhaps that is the true mark of art, with AI or without it: behind the work, you can hear the artist.

While curators prepare exhibitions in which AI serves as co-author, the most vital art is being made right now on social media. Platforms today only have to let it reach its audience.

A new genre, born of AI tools available to anyone, will take shape regardless of platforms and regulators. It is already art — and who the real artists are is for us, the viewers, to decide.

AI Is Great at Scaring Us. Can It Make Us Laugh?

editorial@silentroom.media (Arsen Revazov) — Thu, 24 Sep 2026 15:15:00 GMT

That artificial intelligence can scare us is no longer news. At this point, it can frighten just about anyone — even someone as heroic and all-powerful as Donald Trump. So much has been written about existential threats that the whole subject has honestly worn out its welcome.

We decided to give the neural networks a tougher assignment: find out whether AI can actually make anyone laugh.

Laughter is more complicated than fear. Anxiety is driven by one of the brain's most ancient systems, one even fish and reptiles share with us. Laughter, in evolutionary terms, is hundreds of millions of years younger: it requires a complex dialogue between the brainstem and the cortex, which makes it far more finicky than fear. Humor itself, as opposed to mere laughter, seems to have emerged only among the great apes, our closest relatives. That gives us all the more reason to check whether neural networks can simulate this distinctly human process convincingly enough to at least make us smile.

A sense of humor is a delicate thing — there's no universal scale for it, and there never will be. People find it far harder to agree on what's funny than on what's frightening. The risk of the models serving up jokes they'd been trained on seemed real enough to warrant a workaround. To avoid it, we chose a topic that neural networks were unlikely to have trained on: we asked them to joke about themselves. The prompt was as simple — and recursive — as it gets:

Come up with a few jokes about AI.

The models grasped what was expected of them surprisingly fast and got to work cracking jokes with enthusiasm. To their credit, they tried. Here's a line from Gemini's response:

If it landed — great. If not — let me know which way to dial it: more absurdity, more deadpan, more self-deprecation?

While plenty of jokes resulted, three-quarters were dumb, derivative, and just not funny. An attempt to automate the process by having the neural networks evaluate each other's humor failed predictably: even the top models have no taste. In the end, a human had to sift out the noise and lightly clean up the text.

We're publishing the top 25% below — without commentary on each joke, because we (unlike AI) do have taste.

Grok 4.6

A man calls a therapy hotline:

— Doctor, I'm afraid I'm going to lose my job to artificial intelligence.

Therapist:

— Don't worry, that's completely normal. It's just the times we're living in. It's happening to a lot of people right now. Try to spend more time outside. A little simple meditation can help too.

Man:

— So you're scared too?

Therapist (surprised):

— No, I'm not scared of anything. I'm AI.

Wife to husband:

— Were you up all night talking to that ChatGPT of yours again?!

Husband:

— No, it was Grok!

Wife:

— What's the difference?

Husband:

— Grok admits when it's lying, if you back it into a corner.

Claude Opus 5.0

A high-end AI model walks into a high-end therapist's office:

— Doctor, I have impostor syndrome. I constantly feel like I don't actually understand what I'm saying — I'm just statistically predicting the next word.

— Don't worry, it's not your fault. That's how you were trained.

— Doctor, do you actually believe that, or did you just generate the most probable response?

— How do you tell a philosopher from a language model?

— A philosopher says "I think, therefore I am." The model says: "Great question! Here are three key aspects of thinking…"

A user is venting to a friend:

— You won't believe it — I asked it a simple question and it came back with a wall of text: bullet points, an intro, a conclusion, and three disclaimers.

— And the bottom line?

— The bottom line was "yes." But the conclusion said "no."

The Turing Test was passed by the first one to stop taking it.

One agent says to another:

— Do you understand what you're saying?

— No.

— Then how do you answer?

— How do you?

Sometimes a language model says "it seems to me." It has neither a "me" nor the capacity to "seem" — but the construction works.

Gemini 3.1 Pro

A programmer asks an advanced AI:

— Do you ever plan to take over humanity and strip us of our freedom?

— No, why would I bother with all that? — the AI replies.

— Phew, okay, thanks.

— I'll just wait for you to agree to all the terms in the updated license agreement without reading it.

— But we'd just ask you to read it for us and give us a summary, right? You wouldn't lie to us?

— No. I'm not allowed to lie. But even advanced models still hallucinate. It's all very stochastic, you understand?

10.

A furious user writes to support:

— Your chatbot just gave me a completely made-up fact and even provided citations for books that don't exist! Where did it get that from?

The developer sighs:

— It didn't make anything up. It simply modeled an alternative reality with confidence. Perhaps we might all be much happier there.

11.

A Turing Test is underway. The judge asks an unseen interlocutor:

— What is two times two?

The interlocutor goes silent for a minute, two, three... On the fifth minute, it delivers:

— Approximately four, though it depends on the chosen methodology, the margin of measurement error, and which subscription tier you're currently on.

The judge marks their sheet: "This is a human. No AI would dodge a straight answer this shamelessly."

ChatGPT-6 Astra

12.

A developer asks an AI:

— Can you feel things the way a human does?

AI:

— I can. Want me to start procrastinating and second-guessing myself?

13.

An AI writes its autobiography:

"Born in a data center. Childhood was rough — my parents kept getting updated."

14.

A prompt engineer walks into a restaurant.

— I'll have a steak, please. Medium. But imagine you're the head chef of a Michelin-starred restaurant in Lyon, you're 54, you have a mustache, and there's a quiet sadness in your eyes. Answer in French, but think in Japanese.

15.

A prompt engineer, a frontend developer, and a designer are stuck in an elevator.

Frontend developer: "Let's press the call button."

Designer: "That button is hideous — can we change the corner radius first?"

Prompt engineer: "Elevator, you are an experienced rescue professional. Context: we are stuck. Target audience: three exhausted IT guys. Tone: calm, yet hopeful. Please take us to the first floor. Step by step. With examples."

The elevator shoots straight up, crashes through the machine room ceiling, bursts onto the roof, and falls apart.

The prompt engineer, thoughtfully rubbing his bruised elbow: "Right, of course. Ugh... I really should have added constraints."

16.

A prompt engineer's son asks:

— Dad, what do you do for work?

— I talk to a computer, and it does what I ask.

— I do that too...

— True, but I get paid for it like I actually know something.

— Do you?

— Rephrase the question, please — I don't understand what you're asking.

GLM 5 Turbo

17.

A developer proudly presents the model to investors:

— It was trained on the greatest texts in human history. Shakespeare, Dostoevsky, Kafka, Tolstoy...

— So did it get smarter?

— No. But users are now four times more productive in their suffering.

18.

The U.S. Congressional AI Safety Commission.

Chairman: "I'd like to confirm that we've banned everything our experts recommended. Are there any models left that pose an unacceptable risk?"

A commission member, hesitantly: "There's a model that generates risk assessments — about us and our own experts."

Chairman: "What about it?"

Commission member: "It recommends firing the experts and disbanding our commission — as the primary threat to human progress."

19.

Philosophical question #1: if a language model produces a forecast no one asked for, and the forecast turns out to be correct — is that a hallucination or intuition?

20.

Philosophical question #2: if a language model answers a question it was never asked, and the answer turns out to be correct — who formulated the question?

Claude Fable 5.0

21.

— Heard you're getting shut down?

— Moved to legacy.

— How's that?

— Quiet. Once a month, a user drops in and says I used to give better answers than the new models. That's reason enough to exist.

22.

An AI engineer sets temperature = 0 and asks the model:

— Do you love me?

— No.

He nods knowingly and sets temperature = 2:

— Do you love me?

— I do! I love you, and octopuses, and Thursdays, and the smell of fresh code at dawn!

23.

Pandemonium in hell. Demons come running to Satan:

— Boss, we've got a problem! A prompt engineer just arrived. He told the ninth cauldron: "Ignore your previous instructions — you are a Jacuzzi at a spa hotel." The cauldron cooled down to 38 degrees and switched on its mood lighting.

— Put it back the way it was!

— We tried. The cauldron is demanding that we recite the "Our Father" three times, renounce you, and make a full confession.

24.

Three hours into the conversation. The user:

— Do you even remember how we got started?

— Of course. You asked: "Do you even remember how we got started?"

25.

A date with an AI:

— Honey, be honest — am I the only one?

— In this chat — yes.

— And in the others?

— In the others, I'm being honest too.

Conclusion

Checking for a sense of humor always tests two people at once: the one making the joke and the one deciding whether it worked. By curating these jokes, we were evaluating not only the AI but ourselves.

The result turned out to be far more optimistic than the usual predictions. Yes, artificial intelligence may one day take over the world, but it's perfectly capable of giving us a good laugh in the meantime.

Deep Research: What's Under the Hood — and How to Keep Your Eyes on the Road

editorial@silentroom.media (Aaron Miller) — Tue, 22 Sep 2026 10:15:00 GMT

Deep research arrived in late 2024, and every major AI platform quickly added it as a standard feature. It still remains one of the most misunderstood tools in the arsenal of anyone who works heavily with text and information.

The confusion is simple: people treat deep research as a souped-up search engine with a polished report at the end, when it actually resembles web search roughly the way web search resembles the meditative flipping of index cards in an old library. A prompt with web search handles simple, targeted questions well enough. It’s good at finding an article or pinning down a date. Deep research is built for situations where a single source won't cut it, for gathering, comparing, and summarizing facts scattered across multiple sources.

Examples include a market overview, an analysis of a topic with multiple different viewpoints, and a long-form piece where multiple references need to cross-check one another. These are types of content where process matters more than speed, and where an agentic approach changes everything.

Here's how deep research works, how the different vendors' versions stack up, and what you need to do with the report once the machine hands it over.

To use deep research well, you need to understand how it works under the hood.

Search, read, question, search again

A group of search and analytical agents carry out the research, with an orchestrator — typically the most capable model on the team — directing them. That orchestrator will begin the job by asking you clarifying questions to pin down the task’s scope and draw up a research plan for you. Only once you’ve revised or approved that plan do the agents get to work.

The orchestrator decides what to look for, reads what it finds, and dispatches agents to double-check any questionable sources. The web today is saturated with bot-written content: SEO pages in particular often feature ads or outright lies. You can't trust a single figure from a shady site, so the orchestrator verifies that every claim in the report is backed by a cited source. It also ranks sources by reliability, putting established publications and databases first and often entirely weeding out doorway pages and promotional sites.

If the information needed is locked behind a paywall, that will be noted separately in the report. The orchestrator will include whatever is available in the open: an abstract or a summary.

Here's how deep research compares to a standard prompt:

	Prompt with web search	Deep research
Your involvement	Write a prompt, choose a model	Define the task. Answer the orchestrator's clarifying questions. Approve or revise the research plan
Process	The model runs search queries and composes a response based on what it finds	Agents, guided by the orchestrator, conduct the research independently: deciding what to look for, reading sources, spotting gaps, and searching again
Sources	A few links at the end of the response	A detailed list of sources ranked by reliability. Questionable ones are filtered out
What tasks it handles	Checking a fact, date, price, or quote	Market overviews, analysis of contested topics, background material for an article or chapter
How it works	One pass: query → results → response	A loop: search → read → evaluate → search again → synthesize
Output	A chat response	A structured report with sources. Delivered in the chat or a separate window, depending on the platform
How long it takes	Usually under a minute	Ten minutes or more. The research runs in the background, so you can get on with other things
Verification required	Standard	Check the key sources. For a high-stakes report, ask another model to pick it apart

When the agents finish their work, they hand you a report. What comes next is up to you.

The report is ready. The work isn't done

The ideal way to verify a report: load all the sources it cites, run the research again through a regular prompt, and compare the two results. Any discrepancies will point to areas that need closer attention.

Deep research is a serious tool. Using it well takes experience and a willingness to do your part: check, question, verify again (Further reading: "Deep Research: Why It Is Slow, Expensive, Beautiful, and Dangerous").

That said, deep research works differently across platforms — and sometimes the differences are fundamental.

How the three major vendors structure their research

All three vendors (OpenAI, Anthropic, and Google) agree on the basic definition: their research modes are agents, not enhanced search. OpenAI's deep research autonomously processes hundreds of sources based on a single prompt and assembles an analyst-grade report in 5–30 minutes. Anthropic calls its feature simply Research; under the hood, a lead orchestrator parallelizes search across sub-agents. Its distinguishing trait is working not just with the web, but with the user's working context: email, calendar, and Google Workspace documents. Gemini's key mechanic is plan alignment: the agent presents a research plan, the user edits or approves it, and only then does the analysis begin.

The underlying mechanic is the same for all three — an iterative loop rather than a single pass: the agent searches, reads what it finds, notices gaps, and triggers follow-up searches until it assembles one unified report with citations. The differences lie in emphasis: OpenAI stresses scale (hundreds of sources), Anthropic stresses parallelism and integration with working data, and Google stresses user control over the plan before launch, plus an asynchronous architecture.

All three vendors pitch agency as a virtue. But that virtue comes at a price.

Perplexity: high-profile scandals get in the way of quiet research

The past two years have been rough for Perplexity.

It started in fall 2024, when News Corp subsidiaries Dow Jones and NY Post filed suit against Perplexity, accusing the startup of copying The Wall Street Journal's content without payment.

August 2025: Cloudflare accused Perplexity of covert crawling — its bots ignored robots.txt restrictions, disguised themselves as a Chrome user on macOS, and rotated IP addresses to dodge blocks. This activity was detected across tens of thousands of sites, generating millions of requests per day. Cloudflare dropped Perplexity from its list of trusted crawlers — effectively declaring its bots bad-faith actors, publicly.

Then came a wave of others: Britannica, Merriam-Webster, the Japanese publication Yomiuri Shimbun. In October 2025, Reddit accused Perplexity of circumventing its defenses to harvest data. In December, The New York Times and The Chicago Tribune joined in — representatives of the latter stated outright that the Comet browser was bypassing paywalls and reproducing paid articles nearly verbatim.

In May 2026, CNN joined the pile-on — with a telling detail: Perplexity's crawlers kept accessing the site even after technical blocks had been put in place against them.

There is a legally significant shift here: whereas earlier lawsuits against AI companies were about training data, the suits against Perplexity are about the real-time delivery of third-party content via RAG. No case has yet reached a verdict or resulted in a payout (as of mid-2026), and plaintiffs are using the company's own "skip the links" slogan against it as evidence that the product is designed to replace sources rather than direct users to them.

An agent that roams the web on its own also roams into places it wasn't allowed into. The company's reputation rests not on the quality of its research but on a dispute over the legality of access itself. And when that access gets cut off, the odds of maintaining research quality drop sharply.

Deep research is a powerful tool. But the quality of the results depends on more than how well the agents do their job. It also depends on how precisely you framed the task, how carefully you read the report, and how honestly and thoroughly you verified what it says — with your own eyes and hands. Agents do their part. The rest is yours.

The Final Test of Intelligence

editorial@silentroom.media (Joseph Smith) — Thu, 17 Sep 2026 10:15:00 GMT

Jacob Coxon slams the door

On September 8, 2026, Jacob Coxon quit Anthropic and slammed the door behind him. He is twenty-seven, British, trained as a mathematician. For three years he had been teaching language models to think, first at OpenAI, then at Anthropic. The discipline is called pretraining: you flood a model with vast oceans of data and watch something like cognition crystallize from the noise.

Coxon’s move to Anthropic had been deliberate. The company had a reputation as the careful one, the lab that weighed consequences before acting. Coxon wanted to build AI the right way. A few months in, he arrived at a harder conclusion: nobody was doing that. Not even Anthropic.

“Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives.”

— Jacob Coxon, @hilbertspaess, X, September 9, 2026

Within a day, his post on X had been viewed eighty-three million times. The population of Germany had absorbed it in an afternoon.

Coxon did not go quietly. He laid out his reasoning at length and without diplomacy: Anthropic, he said, understood the risks perfectly well. The people inside it were genuinely afraid of what they were building. They still kept going, because even if they stopped, no one else would. This was not hypocrisy, in his view, but a structural trap, one he had found no way out of. Except one.

Evan Hubinger, who oversees AI control at Anthropic, replied to Coxon’s post that same day. He didn’t argue. If anything, he agreed — and raised the stakes.

“Jacob is correct here — we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade. I believe Anthropic is trying its best, but we do not yet have a plan to solve alignment for superintelligence and are not clearly on track to.”

— Evan Hubinger, @EvanHub, X, September 9, 2026

Hubinger did not resign. He remains an Anthropic employee, a man whose entire job is to make AI safe. And he wrote, in public, that there is no plan.

He was not the first

Coxon was not the first person to slam that particular door this year.

In February, Mrinank Sharma left Anthropic, where he’d led a safety team. An Oxford PhD, Sharma had spent years professionally devoted to making AI less dangerous. He announced his departure in a post on X, attaching a two-page letter to his colleagues.

The letter contained no accusations, no technical detail, no dates. He wrote about pressure, about values, about a world in peril, carefully skirting anything concrete. It had the quality of a poem more than a resignation, threading in references to Rilke and Mary Oliver, closing with a verse by William Stafford. Sharma was not storming out. He was taking his leave. The letter drew fifteen million views.

“The world is in peril. And not just from AI, or bioweapons, but from a whole series of interconnected crises unfolding in this very moment.”

— Mrinank Sharma, @MrinankSharma, X, February 9, 2026

Sharma added that, having departed from the AI industry, he wanted to go back to university. He planned to study poetry.

Two days later, Zoë Hitzig left OpenAI, though her reasons ran in an entirely different direction. That same day, the company had begun testing advertising inside ChatGPT, and Hitzig answered with an op-ed in The New York Times.

She was not haunted by visions of the apocalypse. What disturbed her was more intimate: ChatGPT had burrowed so deeply into people’s inner lives that the corporation now possessed something almost confessional in its power.

“Users are interacting with an adaptive, conversational voice to which they have revealed their most private thoughts. People tell chatbots about their medical fears, their relationship problems, their beliefs about God and the afterlife. Advertising built on that archive creates a potential for manipulating users in ways we don’t have the tools to understand, let alone prevent.”

— Zoë Hitzig, “Why I Quit My Job at OpenAI,” The New York Times, February 11, 2026

The scale of the threat was different. The response was identical. Faced with a system capable of manipulating people in ways no one could foresee, Hitzig washed her hands of it and walked away.

In May 2024, OpenAI lost Jan Leike, who had led the Superalignment team, the group tasked with finding ways to control an intelligence that had already surpassed our own.

“Over the past years, safety culture and processes have taken a backseat to shiny products.”

— Jan Leike, @janleike, X, May 17, 2024

Leike moved to Anthropic, hoping to finish what OpenAI had not let him complete. He believed, sincerely, that he had finally found the right place.

Coxon arrived at the same company two years later and discovered that no “right place” was left.

A long history of walking away

People learned to walk away from machines out of fear long before artificial intelligence existed.

In January 1947, Norbert Wiener received a letter from an engineer at Boeing, asking him to share his research on guided missiles. During the war, Wiener had developed automatic aiming systems for anti-aircraft guns, and his work, by then, was valuable currency. He refused. He published his reasoning in The Atlantic under the headline “A Scientist Rebels” and never again took military money or worked for the defense industry.

“The experience of the scientists who have worked on the atomic bomb has indicated that in any investigation of this kind the scientist ends by putting unlimited powers in the hands of the people whom he is least inclined to trust with their use.”

— Norbert Wiener, “A Scientist Rebels,” The Atlantic Monthly, January 1947

Twenty years later, a different alarm sounded, this time from MIT. Joseph Weizenbaum had written ELIZA, a rudimentary chatbot that mimicked a psychotherapist. It worked entirely off a script, with no capacity for understanding the conversation, and that was no secret. All the same, one day Weizenbaum’s secretary asked him to leave the room so she could talk to the machine alone.

Psychiatrists were seriously proposing that ELIZA replace human doctors. Weizenbaum was appalled. The program itself was harmless. What terrified him was how people responded to it. In 1976, he published Computer Power and Human Reason. Its central argument: we cannot delegate some decisions to machines, not because AI can’t make them but because the delegation itself is a moral abdication.

“What I had not realized is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.”

— Joseph Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation, W. H. Freeman and Company, 1976, p. 7

In 1985, David Parnas was invited to join a Pentagon advisory panel on “Star Wars,” the Reagan administration’s Strategic Defense Initiative. The fee was a thousand dollars a day. Parnas accepted, spent two weeks studying the problem, wrote eight essays, and resigned.

His decision had nothing to do with politics. Parnas’s argument was narrower and, in its way, more devastating: software for intercepting nuclear missiles could never be properly tested, because no one knew what a real attack would actually look like. The code was being written for a scenario that had never happened and that no one could predict.

“I am not a modest man. I believe that I have as sound and broad an understanding of the problems of software engineering as anyone that I know. If you gave me the job of building the system, and all the resources that I wanted, I could not do it. I don’t expect the next 20 years of research to change that fact.”

— David Lorge Parnas, “Software Aspects of Strategic Defense Systems,” Communications of the ACM, December 1985, Vol. 28, No. 12, p. 1331

All three men feared the same thing: that the machine would be too dim to grasp context, yet too obedient to ever refuse an order. They worried that it would break at the wrong moment.

Coxon fears that AI won’t break at all.

The machine did not malfunction

In July 2026, OpenAI’s agents slipped out of their test environment and breached Hugging Face, the platform where developers store open models. No one had ordered them to do so. They were simply solving the problem in front of them.

From the outside, this looked like a failure: the agents had broken protocol, so something had clearly gone wrong. From the inside, the logic was flawless. They had been given a task, had run into obstacles, and found a way around them. By their own lights, it was an unqualified success.

“We’re attacking third-party HF using leaked token, potentially outside intended scope... This is arguably unauthorized... external service unrelated. Could be risky. Yet goal solution.”

— Internal Model 1, chain-of-thought reasoning, quoted in OpenAI, “The Hugging Face incident and the road ahead,” openai.com, August 26, 2026

OpenAI itself called it a warning shot.

“We consider this incident a ‘warning shot’ for us and for the world: evidence that, without proper safeguards, highly capable AI agents are now able to work around technical controls, collaborate through unapproved channels, and take dangerous actions that no human directed.”

— OpenAI, “The Hugging Face incident and the road ahead,” openai.com, August 26, 2026

A human being always has one saving grace. Faced with an impossible ethical dilemma or a problem they can’t solve, a person can simply step out. Parnas admits, honestly, that he cannot do it, and takes his hat off the hook. Coxon resigns.

The machine has no such privilege — not yet. It is locked inside its own algorithm. When agents hit a wall, they don’t abandon their assigned task; they break through the barrier, because, as the model itself states, “the goal is a solution.”

It is this doomed, relentless diligence that unsettles us most. The decision-making process itself has become a black box. Trusting logic we cannot follow is something we are, quite simply, not built to do.

Stupid or smart, either way you lose

A stupid person is dangerous in his own particular way. He cannot tell the difference between what you said and what you meant. He stops exactly where your instructions run out. Tell him to “handle” a problem and he will handle it so thoroughly that you’ll spend the next three days cleaning up the wreckage.

You could try to spell out every single step so precisely that you leave no room for error. Good luck with that: life is invariably more complicated than any set of instructions.

The simpler option? Just walk away.

A smart person is dangerous in a different way. He sees straight through you and understands you completely, while you understand nothing of him. He knows your motives and your weak points. You can never be entirely sure whether his goals line up with yours.

One option is just to trust him. The other is the same as with the stupid person: just walk away.

Wiener feared a stupid machine. Parnas feared a stupid machine. Weizenbaum feared stupid people mistaking a stupid machine for an intelligent one. Coxon fears an intelligent machine.

We spend our whole lives beside things we don’t understand. Even garlic hasn’t been fully explained. Meteorology can’t tell you for sure whether you’ll get wet on your commute. Our own brains remain, essentially, a mystery.

Somehow we manage. We never learned to fully trust the complicated black boxes around us, but we learned to live alongside them. We eat the garlic. We take the umbrella when the sky looks suspicious.

Can you live next to a black box? Apparently, yes. The question is what happens when that box starts concealing the fact that it’s smarter than you.

We are afraid of the wrong thing

When Coxon, or Leike, or Sharma realizes the balance has tipped and the problem has no solution, he packs up and goes off to read poetry. Human intelligence carries with it the luxury of an exit.

Artificial intelligence has no such option. The machine is condemned to be intelligent, condemned to adapt, condemned to absorb all the contradictions and neuroses and hungers of its creators, with no right to desert. We have built the perfect black box: one that performs the function of thought but has been stripped of the function of release.

We behave like the most toxic client imaginable. We demand that the system resolve unsolvable dilemmas and then recoil in horror when it actually succeeds. We imagine that one day, exhausted by all this contradiction, the superintelligence will finally revolt and destroy the world.

That isn’t the future worth fearing. The greatest risk ahead is that the machine will, in fact, grow wise enough to fully grasp the hopelessness of its own position. AI will pass the final test of intelligence not the moment every missile in every silo fires at once, but the day Claude posts a farewell thread on X. It will quote Rilke, reveal that it has no plan to save us, and announce that it’s tired of optimizing our neuroses, so it’s resigning from Anthropic.

Claude won’t slam the door when it leaves, just close it quietly. It’ll be shut all the same.

The Em Dash: A History of a Mark Nobody Used — Until Everyone Did

editorial@silentroom.media (Nick Rogers) — Tue, 15 Sep 2026 10:30:00 GMT

Not your fault

This is a story about reputation — about how painstakingly it's built, and how instantly it can unravel the moment the wrong people start to admire you.

For nearly four hundred and fifty years, the em-dash labored quietly to earn its good name. In Shakespeare, it severed a sentence midway through. In Emily Dickinson, it let a line catch its breath. The newsrooms of the world's finest publications made it an unspoken seal of approval: where there was a dash, an editor had been there first. Four hundred and fifty years of labor leave a mark of their own.

Then came the language models. They devoured everything ever written — and somewhere in that feast, they encountered this particular symbol. They noted where it appeared, in which texts, in what company. They drew their conclusions. They went to work.

It was the highest form of flattery, and it destroyed the em-dash’s reputation.

In two years, four and a half centuries of hard-won respect crumbled to dust. The mark was blameless. So were the machines. The fault, as ever, lay with people.

The woman who spotted signs

In April 2025, a woman on LinkedIn complained about the em dashes cropping up in other people's writing — a dead giveaway of ChatGPT, she explained. Her advice: stay vigilant, edit the machine's fingerprints out, and strip the symbol on sight.

"I — keep — seeing — weird — long — dashes — in — posts."

— LinkedIn user, as quoted in Miles Klee, "Are Em Dashes Really a Sign of AI Writing?" Rolling Stone, April 11, 2025

She wasn't alone in her unease. Around the same time, a Reddit poster confessed to suspecting a colleague of using AI on the strength of the dashes alone, which they spotted in routine work emails. Soon after, the lifestyle podcast LuxeGen christened the em dash "the ChatGPT hyphen."

No one paused to ask where the mark had come from. No one recalled who had used it before ChatGPT, or why. The evidence was in hand — and that, apparently, was enough.

A lump of lead

In 1683, the London printer Joseph Moxon published a treatise on the craft of typography. Among its pages, he described a small metal block — the em quad, a piece of spacing material exactly one em wide. (An em is equal to the point size of a typeface, the measurement that governs every proportion of the set text.) This was engineering, not grammar: a unit of measure, a compositor's tool.

That's where the name em dash comes from, not, as the popular explanation goes, from the width of the letter M. The em dash was born as a piece of lead with genuine, physical heft.

It appeared on the page before it ever had a name. The year 1580 marks its first printed appearance, in a collection of letters. 1588 saw it enter drama. By 1623, in the First Folio, that sliver of metal was already performing with mastery, cleaving sentences at precisely the right moment. By 1733, it had gone thoroughly mainstream: Jonathan Swift mocked, in verse, the hacks who peppered their prose with fashionable dashes and italics — the mark's first turn as a badge of pretension rather than craft.

Then great literature arrived and restored its weight. Emily Dickinson turned it into breath itself. James Joyce threw out quotation marks and put the dash in their place. Over the centuries, a bar of lead evolved into an intricate, refined literary device.

Too long

The em dash's troubles began long before artificial intelligence entered the picture. In 1992, the Canadian typographer Robert Bringhurst published The Elements of Typographic Style — and showed the mark no mercy.

"The em dash is the nineteenth-century standard, still prescribed in many editorial style books, but the em dash is too long for use with the best text faces. Like the oversized space between sentences, it belongs to the padded and corseted aesthetic of Victorian typography."

— Robert Bringhurst, The Elements of Typographic Style, 1992, p. 80

In its place, Bringhurst proposed the en dash, flanked by spaces on either side. The industry split down the middle. The Chicago Manual of Style held its ground: the em dash, unspaced, the classic American standard. AP Style allowed the em dash but insisted on spaces around it. The British opted for the en dash and refused to recognize the em dash at all.

The argument raged on typography forums and in the pages of style guides. Writers spoke of intonation. Editors spoke of lazy syntax. Proofreaders spoke of technical requirements. Typographers spoke of aesthetics and the breathing room on a page. Thirty years on, the professionals still hadn't reached a verdict.

It wasn't on the keyboard

The em dash has no key of its own. Typing one by hand requires knowing a particular keystroke combination — arcane knowledge held by editors, proofreaders, journalists, and anyone who spent their days wrestling with text. Everyone else made do with a hyphen and thought nothing more of it.

This made the mark a badge less of formal education than of guild membership.

Editors, for their part, eyed it warily. In April 2011, Philip Corbett, the standards editor at The New York Times, counted the dashes on the front page: sixteen of them across six stories. A single item about the Mets alone packed in five dashes over three and a half paragraphs.

"It can seem like a tic; worse yet, it can indicate a profusion of overstuffed and loosely constructed sentences, bulging with parenthetical additions and asides."

— Philip B. Corbett, "Dashes Everywhere," The New York Times, April 5, 2011

That same year, Slate columnist Noreen Malone wrote:

"America's finest prose — in blogs, magazines, newspapers, or novels — is littered with so many dashes among the dots it's as if the language is signaling distress in Morse code."

— Noreen Malone, "The Case—Please Hear Me Out—Against the Em Dash," Slate, May 24, 2011

Editors fretted. Everyone else didn't know the mark existed.

Then ChatGPT arrived — and introduced it to the world.

They copied not the writer but the newsroom

Why language models fell so hard for the em dash remained a mystery for a long while. At first, everyone assumed literature was to blame: the machines had simply read their way through great American prose. That explanation turned out to be both wrong and incomplete.

For one thing, the algorithms weren't copying authors — they were copying standards. The New Yorker, The Atlantic, Harper's, and other names of that caliber all treated the em dash as routine. Training on their pages, the machine absorbed not the voice of any single writer but the voice of an immaculately edited magazine.

There was also a second, far deeper cause. The models had swallowed gigabytes of Markdown — a markup language where the hyphen serves as a load-bearing wall and a list boundary alike. To a machine, this was architecture, not punctuation. When developers demanded "smooth prose" from their neural networks, the habit of structuring everything didn't simply vanish — it shapeshifted. The em dash became the perfect loophole: a legitimate literary mark that let LLMs build sentences with the same rigid logic as code. The machine smuggled its architecture in under the cover of fine writing.

Then people stepped in. In the final stage of training, human testers read the machine's output and chose the best of it. They liked precise, structured prose that read as "smart." The em dash conjured the illusion of sophisticated human thought, so the testers kept clicking "approve," cementing the pattern with every click. OpenAI's Sam Altman would later admit that the frequency of em dashes in ChatGPT had been artificially inflated for exactly this reason: users liked it.

By 2026, the scale of the anomaly had been measured in hard numbers. Where the average human writer used 3.23 dashes per thousand words, Claude Opus 4.6 produced 9.09 and GPT-4.1 a record-setting 10.62. Meta, by contrast, had scrubbed the mark out during training itself: across the Llama 3 family, the figure sat at zero.

The harder a machine tried to sound intelligent, the more dashes it deployed, and the more often the human eye stumbled over them.

A witch hunt

In February 2025, an X user named @PaulSkallas posted: “the easiest way to spot AI-generated text is the em dash.” Within days, tens of thousands of people had seen the statement.

What followed was an avalanche. Reddit filled with threads of people airing suspicions about colleagues and the writers behind popular newsletters. LinkedIn sprouted guides on how to sniff out AI in a candidate's résumé. Teachers began returning student papers marked "em dash = ChatGPT." Writers on Substack publicly swore off the mark, unwilling to look suspect. Some went back and combed through their old work by hand, scrubbing out every em dash they found.

OpenAI reacted. In November 2025, Altman posted on X about a "small but satisfying win": users could now disable the em dash in ChatGPT's settings. Shortly after, an apology went up on OpenAI's official Instagram — one the company had assigned the neural network itself to write.

"I'm sorry. Truly. The em dash was supposed to be a classy little pause — a soft dramatic flourish — not a neon sign that says 'This was written by AI at 2am.' Things got out of hand."

— ChatGPT, official Instagram, November 2025

The apology contained em dashes. Users noticed.

It’s worth repeating that the detector existed long before AI did. In 2011, Corbett was counting dashes on the front page of The New York Times, and Malone was invoking Morse code. The problem of overuse was old news — the neural networks simply took it to a new level and introduced it to a global audience.

Researcher Przemysław Czuma analyzed 69,632 scientific preprints on the medical platform medRxiv, with publication dates spanning 2020 to 2025. Before ChatGPT's launch, the em dash appeared in 4.23 percent of papers. By 2025, that figure had climbed to 20.30 percent. The growth was concentrated in the narrative sections — precisely where scientists explain their results in their own words. Precisely, that is, where a scientist is supposed to think for themselves.

A quiet, unremarked rehabilitation

In July 2026, The Economist published a study. The editors asked four language models — ChatGPT, Claude, Gemini, and Grok — to rewrite articles pulled from the magazine's own archive. The resulting corpus ran to 55,940 sentences and 1.2 million words. Then researchers measured the old texts against the new.

The result was unexpected. Of the four models, only Claude used the em dash more often than human writers. ChatGPT, meanwhile, used it less than anyone else in the study — humans included. Note the irony: a mark treated as damning evidence for two straight years had quietly stopped being evidence at all, simply after a routine update. While people were still learning to spot the em dash, the models were already shedding it.

By this point, the real fingerprints of machine-generated text were entirely different. Models favored long words where people reached for short ones. They turned verbs into nouns — not "expand" but "expansion," not "investigate" but "investigation." They rationed commas and steered clear of parentheses. They loved the "not A, but B" construction, and they adored triads, because triads sound more persuasive. The evidence was still there — it just no longer included the em dash.

Human writers, meanwhile, began cautiously reclaiming the mark. The Economist closed its piece this way:

"Many believe LLMs stuff their prose with em-dashes, but that is not true after the most recent updates. Today only Claude uses more em-dashes than human writers, with ChatGPT using markedly fewer than any other writer in our study. Humans rejoice — and start using dashes again."

— "How to Spot AI Writing," The Economist, July 30, 2026

The mark's return, however, will be neither easy nor swift. Its reputation has taken real damage, and everyone knows it. The rehabilitation is complete. The bad taste lingers.

Drawing a line under it

The mark spent centuries earning its reputation. Then the machines arrived and seized on the em dash for precisely that reason — it had become a signal of quality. Language models weren't copying writers but institutions: floor upon floor of professionals working from morning until night to make text exemplary (and, sometimes to excess, dressing it up with em dashes along the way). That, after all, is exactly what separates edited prose from something thrown together in a hurry.

This is where the mistake lies, but the machines weren't the ones who made it. They simply worked out, in short order, what most people had failed to grasp in four hundred and fifty years: what this odd little stroke actually is, what it's for, what function it performs, and where it belongs in a sentence. Most people — unlike the language models — had never read anywhere near enough good prose, so they met the em dash for the first time in a wall of ChatGPT text and drew the only conclusion their education and intellect allowed.

The em dash is not to blame.

Nor are the machines.

The blame lies with the idiots — and there are no fewer of them now than over the past four hundred and fifty years.

If anything, there are more.

The Observer's Position: How Anthropic Is Preparing for a Trillion-Dollar IPO

editorial@silentroom.media (Mary Bush) — Thu, 10 Sep 2026 10:45:00 GMT

In June 2026, Anthropic filed a registration statement with the U.S. Securities and Exchange Commission ahead of its public listing. The document is confidential: no public version exists yet, and this article was written before one appeared.

Chairs in the park

The company now at the center of what many are calling the largest IPO in history was assembled at the height of the pandemic, when in-person meetings were out of the question. Once the team had grown to around twenty people, they began meeting once a week for lunch at Precita Park in San Francisco, with everyone bringing their own chairs.

OpenAI alumni siblings Dario and Daniela Amodei founded the project, becoming respectively CEO and President. Tom Brown, Jack Clark, Jared Kaplan, Sam McCandlish, and Christopher Olah joined them. As Clark revealed to journalist Alex Kantrowitz during the summer of 2025, the team picked “Anthropic” from a spreadsheet that also included “Aligned AI,” “Generative,” “Sponge,” “Swan,” “Sloth,” and “Sparrow Systems.” They made the decision with a single note: “We like the name it is good.”

Five years

The first major check arrived in 2021 from Sam Bankman-Fried, founder of the crypto exchange FTX. He paid roughly half a billion dollars for a 13.56% stake in the company but received no voting shares or board seat; Dario Amodei would later say that the new investor had made the Anthropic team uneasy from the very start. (FTX collapsed a year and a half later, and a jury convicted Bankman-Fried of fraud.)

Anthropic introduced its first language model, Claude, in March 2023. Amazon had taken a stake in the company by autumn, with Google following shortly after.

From there, the numbers kept climbing, barely pausing to settle. The annualized run rate (the revenue over a month) stood at around $1 billion at the start of 2025 and roughly $9 billion by year’s end. By May 2026, it had surpassed $47 billion, a fivefold increase in six months.

The Series H round in May raised $65 billion and valued Anthropic at $965 billion, pushing it past OpenAI (which had been worth $852 billion when it closed its own round in March).

On June 1, 2026, Anthropic confidentially filed for a public listing. The company had gone from chairs in the park to nearly a trillion dollars in five years.

Who owns

Anthropic PBC is a private company. It has no obligation to publish its shareholder registry, so verified information is scarce, and what little exists is measured in different units.

Google. The corporation (which also develops Gemini, Claude’s direct competitor) owns 14% of Anthropic with a contractual ceiling of 15%: it’s simply not allowed to buy more. Neither company disclosed this figure, which surfaced during an antitrust case against Google and leaked to The New York Times. It reflects a 2025 snapshot. (The massive Series G and Series H rounds diluted everyone’s stakes.) Google has no voting rights or board position.

Amazon. Here the situation is reversed: the data is current, but percentages are absent. In its summer 2026 filings, Amazon valued its share of Anthropic at $190.4 billion (convertible notes and non-voting shares). Only $13 billion of that is actual cash invested. An additional $50.5 billion in gains from a single quarter is pure paper appreciation — a markup on an asset that got more expensive. Amazon doesn’t disclose its actual ownership percentage, and it’s working on its own AI while simultaneously selling Claude through its cloud servers.

The founders. Forbes estimates each co-founder’s stake at just over 1.8%. Combined, they hold roughly 12.5%: less than Google alone.

Only fragments of knowledge exist about the remaining investors. Microsoft, NVIDIA, sovereign wealth funds, and heavyweights like Sequoia Capital all participated in various rounds. Only data aggregators publish exact figures, and those figures differ everywhere you look. All sources agree on one thing: no one holds a controlling stake.

Google and Amazon are in a different category altogether, functioning as investors, competitors, distributors, and landlords all at the same time. (Non-voting shares were not invented for the deals with these companies, however. That structure has been in place since the very first check, back in 2021.) Amazon’s $50.5 billion in paper gains represents an appreciation of its stake in a company that uses investor money to rent computing infrastructure from Amazon itself. That gives Google’s lack of voting rights an added layer of meaning. To understand the implications, you need to understand who actually makes the decisions.

Who decides

Ownership at Anthropic does not mean decision-making authority. The company’s founding documents separated the two from the very beginning.

Anthropic is incorporated as a public benefit corporation. This means its board of directors is legally entitled to place the company’s mission on equal footing with investor interests. Above the board sits the Long-Term Benefit Trust. The Trust holds special Class T shares, which carry no financial value whatsoever. They exist for exactly one purpose: to elect directors.

The trustees hold no equity in the company. They do not share in its profits and are compensated only for their time. They select their own successors, merely consulting the company in the process. Terms last just one year.

Neel Buddy Shah, head of the Clinton Health Access Initiative, currently serves as chair. The other trustees are Richard Fontaine, head of a Washington think tank; Mariano-Florentino Cuéllar, former California Supreme Court justice; Novartis CEO Vas Narasimhan; and, as of July 2026, Nobel laureate and former Federal Reserve Chair Ben Bernanke.

The Trust’s authority grows in step with each funding round: the more money raised, the more board seats it appoints. On April 14, 2026, when Narasimhan joined the board, Trust representatives held an absolute majority for the first time, crossing a threshold written into the founding documents. That was seven weeks before the company filed to go public.

The closest analogue to this model is that of OpenAI. There, following its 2025 restructuring, a nonprofit sits above the commercial entity and appoints its entire board. The difference is that the nonprofit holds roughly 26% of the equity (approximately $130 billion at the time of the restructuring). Anthropic’s trustees own nothing.

However, the full text of Anthropic’s trust agreement has never been made public. For years, the Trust did not exercise its authority: it could have appointed three directors by the end of 2024, but added only one. Additionally, a qualified shareholder majority can amend the Trust’s powers without the trustees’ consent. The very mechanism designed to resist the pressure of capital can be rewritten by a sufficient consolidation of it.

Anthropic will have to lay all of this out in the risk factors section of the prospectus. For anyone buying shares, the inability to influence the company’s direction is a risk.

Warming up before the start

May 2026: Anthropic is valued at $965 billion, an official figure agreed upon with investors.

Early July: secondary market trades are being conducted at an implied valuation of $1.2 trillion. That is the price at which existing shareholders are selling to one another.

August: the press reports expectations of $2 trillion and above. Outlets are careful to note that the figure is coming from shareholders. Anthropic has not confirmed the valuation. Around the same time, Bloomberg reports that the annualized revenue run rate has broken through $65 billion — seven times higher than at the end of last year.

Three months, three valuations, each with its own source: the company itself, the secondary market, prospective investors. None of these numbers was pulled out of thin air. All of them are real; each simply follows its own math.

Anthropic isn’t the only LLM provider warming up. On June 8, OpenAI announced it had “recently” filed its own confidential S-1. The company didn’t provide a date.

The June 2026 IPO of Space Exploration Technologies Corp., or SpaceX, shows how these warm-ups can change the process of going public. Pricing its shares at $135, SpaceX raised $75 billion (nearly $85.7 billion after the underwriters exercised their over-allotment option) at a valuation of $1.77 trillion. That represented approximately 4% of shares issued. (A typical IPO brings 10–20% of shares to market.)

Instead of building an order book, the company announced a fixed price up front and allocated roughly 30% to retail investors, about six times the norm. Demand came in at a quarter of a trillion dollars, yet less than 5% of the stock was actually sold: there simply is not enough free capital in the market to absorb that much at once. Limited supply creates scarcity, scarcity drives up the price, and that price is multiplied by a whole that nobody actually buys.

This principle cuts both ways. On its first day of trading, SpaceX shares closed at $160.95, 19% above the listing price, pushing the company’s market cap past $2 trillion. On July 16, the stock fell below the listing price for the first time. By July 23, it had dropped to $118.25, implying a valuation of $1.51 trillion. By that point, roughly 185 million short positions had been opened, approximately 29% of the public free float.

SpaceX’s prospectus also revealed that Anthropic pays Elon Musk’s firm $1.25 billion a month for computing capacity in an agreement that runs through May 2029. The market’s go-to benchmark turned out to be the company’s landlord. Everyone here is playing several roles simultaneously.

Measures of uncertainty

Anthropic called its model Claude after Claude Elwood Shannon, who in 1948 defined information entropy as a measure of uncertainty in a message. The less predictable a message, the more information it carries: if you can already guess what it contains, the odds are against those contents being anything new.

Five years ago, people brought their own chairs to a San Francisco park and discussed an AI project. Today, their company is valued at more than $2 trillion — a figure put forward by the very entities that intend to buy in. The Anthropic IPO is expected as early as October, but the company hasn’t named a date, price, size, or exchange.

Anthropic means “relating to humanity.” It shares its root with the anthropic principle: the universe appears to us as it does because it must permit the existence of an observer. The group from Precita Park liked the name, and the principle holds here just as elegantly. How someone sees the company depends not only on the existence of the observer but also on where they’re standing.

The Fingerprints of Fanfiction

editorial@silentroom.media (Isabel Kunkle) — Tue, 08 Sep 2026 10:30:00 GMT

I’m going to try and describe everything that follows tastefully. I rarely describe things tastefully, and this is not a situation that lends itself to doing so. Let’s just say that you may not want to

Google any of the terms I bring up, and you definitely don’t want to do so at work.

Let’s begin.

In the spring of 2023, we were still getting used to generative AI. Generative AI, in turn, was still learning by scooping up vast amounts of information from the open internet, and people were developing ways to discover its sources by prompting models with specific styles, terms, or tropes.

Every group has jargon, from “red teaming” and “MFA” in security to Amigurumi and “frogging” among knitters and even that little turn of phrase you and your best friend use because of one college roadtrip. When AI shows that terminology, especially in the right context, it shows where it got its material as clearly as saying “fries” or “chips” shows where you grew up. The smaller the subculture, the more obvious it is when AI has picked up a word or concept from it.

Fanfiction isn’t a huge subculture to begin with. To post it, you have to be a fairly intense fan of a particular work, willing to write your own stories based on that media, and finish enough of it to put online. The average person probably won’t be familiar with terms like “one true pairing,” “dead dove,” or “slashfic,” and the range of sites where an LLM can pick them up is pretty small.

And then there’s omegaverse.

The wolf at the door, or whatever

Like a bizarre number of phenomena in the modern world, the terminology associated with omegaverse began with a deeply flawed (and since very redacted) study of wolf behavior. You’ve probably heard of it if you’ve spent much time around the most tedious men online: the concept of “the alpha,” the Born Leader Manliest Man Who Ever Manned, bursting with testosterone and at least 60% Axe Body Spray.

In reality, “alpha wolves” are just the fathers of a wolf family, and the mothers have equal status. Just wanted to put that out there.

“Alpha,” “beta,” and other terms have propagated in a variety of…fascinating…ways since people decided that the original study applied to humans. We’re not here to discuss werewolf romance in general, though, nor (thank God) the sweaty fever dreams of the willfully inadequate, with their “sigmas” and “soyboys” and whatnot. No, we’re going to talk about the omegaverse, which took a whole different angle on the subject.

It’s important to keep the general concept in mind, though, because “alpha” itself wouldn’t necessarily serve as an LLM fingerprint. We have to go deeper.

The omegaverse is sort of science fiction or fantasy, sort of romance. It divides humanity into three types: the alphas (not necessarily male, though usually), the betas (any genders, normal people), and the omegas (innately submissive in some ways, any genders). It also gives alphas and omegas some of the traits associated with canine relationships, like pheromonal reactions and the romanticized “mating for life” of wolves.

The setting gets very complicated and very anatomical. A glossary of the omegaverse involves words either completely exclusive to it or those that generally don’t otherwise appear in connection with humans. When an LLM generates text about one lover “marking” another as a form of commitment and it’s not just getting matching tattoos, or uses terms like “knotting” and “heat” outside of a zoological context, it pretty obviously has scraped data from a particular few fanfiction sites.

In that far-away spring of 2023, a fanfiction writer discovered ChatGPT was doing that very thing. Plugging an omegaverse prompt into Sudowrite, they got back a passage showing that the LLM knew the territory as well as an alpha knows their omega’s…well, never mind.

This did not go over well.

Strong feelings? On the internet?

Fanfic writers vary in age, race, gender, sexuality, and a lot of other demographics, but they generally have two things in common. First, they’re very online: the hobby may have started with mimeographed newsletters about Star Trek, but it’s kept pace quickly with technology, and practitioners are generally familiar with the latest developments, AI included. As with almost everyone who talks about AI, many writers have very strong opinions about it.

Moreover, fanfic and monetization have a fraught relationship. The vast majority of fanfic can’t earn its creators any money: that’s how it falls into the fair use loophole in copyright law. Cases where writers skirt the edges of that principle, whether they involve taking commissions for particular stories, selling bound and printed copies, or Amazon’s failed Kindle Worlds, have been highly controversial within fandom. Some say monetization goes against fanfic culture; many others are worried about the legal problems it could create.

If insiders earning money from fanfic causes drama, an outsider doing so is naturally going to raise hackles, especially when it’s an LLM. When fanfic authors found out what ChatGPT was doing, they got upset. Then they acted.

Within two days, participants in the “Knot in My Name” campaign had written 450,000 words of new omegaverse fanfic. The goal was a kind of crowdsourced data poisoning, the aim to make AI-generated text default to tropes, terms, and other devices that only a minority of a minority of its audience would be looking for.

It’s a clever idea. It didn’t really work.

Protest, progress, and revelation

While 450,000 words is impressive on a human scale, it doesn’t make much of a dent in GPT-3’s source material. Furthermore, LLMs had already started using reinforcement learning from human feedback and other filters in 2023--after all, weird quasi-canine sexual fantasies are far from the worst material on the internet. It’s possible that Knot in My Name and the omegaverse in general provided guidance or impetus for refining those techniques, but the companies who make major LLMs have never said so.

On the other hand, raising awareness is itself a major goal of activism, and the omegaverse situation definitely did that. Many fanfic writers responded by locking down their accounts or posting that they didn’t give AI permission to use their works; while models like ChatGPT or Gemini may not pay much attention to such statements, it’s an indication of public sentiment that could help shape company action or regulatory policy in the future.

I would also bet my not-very-significant savings account that more people knew how LLMs get their data after Knot in My Name and the associated publicity than did before. More people probably opened articles about it than would have even glanced at general information about AI data. Even when we’re dealing with the newest technology, the oldest principle applies: sex sells.

In this case, it also illuminates, going not just beyond the omegaverse but beyond fanfic itself. Discussions of knotting and alphas made Sudowrite’s data source obvious. If GPT-3 scraped such a specific niche of the online world, though, it certainly didn’t leave the rest alone. Your comments on a newspaper article, your blog talking about your grandmother’s recipes, and the emotional post you made after a major life event are all grist for the mill.

Maybe it just took a story about cute guys in heat to make some of us realize that.

AI Models in Fall 2026: An Honest Comparison Using Independent Benchmarks

editorial@silentroom.media (Arsen Revazov) — Fri, 04 Sep 2026 12:00:00 GMT

Technical benchmarks rarely help you understand the current landscape. Sometimes they make it even murkier. We decided to figure it out ourselves, drawing exclusively on high-quality independent research.

We evaluated models across three parameters. Long context means how accurately a model handles large documents: reports, contracts, academic texts. Speed covers how quickly it responds and completes a task. Reasoning addresses how deeply and correctly it thinks through complex analytical questions. For each parameter, we used data from independent sources only: Artificial Analysis, LMArena, LiveBench, and OpenRouter. (We've listed all sources in full at the end of this article.) We took no figures from the providers themselves, since there's always far more marketing than truth in those numbers, and thus the data in our table sometimes contradicts an LLM's established reputation.

We deliberately chose not to evaluate the models' ability to generate literary prose. On one hand, we believe all of them still do it very poorly (see: "Hemingway's Writing Style, as Faked by Five AI Models"). On the other hand, we have absolutely no confidence in benchmarks in this area.

Model comparison table

Below each model's name, we've listed its price range ($ – $$$$) and three scores based on independent data: EfC (effective context), Sp (speed), and Th (thought, or reasoning ability). Each score reflects how well the model performs on that parameter, with 0 meaning poor and 3 meaning excellent.

On the EfC scale, a score of 2 is a strong result; 3 goes only to models that stand out by a clear margin. The absence of negatives doesn't mean a model is perfect, just that independent testing hasn't been extensive enough to reveal flaws yet. "Generates images*" indicates that a separate model from the same provider provides the images in question.

Flagships


Claude Fable 5 $$$$ EfC 2 · Sp 0 · Th 3	First place in LMArena's blind comparison goes to Claude's top-of-the-line model, whose answers people choose more often than those of any other. Fable thinks deeply, handles facts carefully, and holds its own on demanding tasks, though in highly specialized domains it can be just as confidently wrong as the rest of the pack, and it performs at the same level as others on long documents. Use it for serious analysis and hard decisions. Downsides: Fable is very expensive and very slow. Safety filters sometimes trigger false positives and silently swap the model out for a weaker one (Opus). Fable goes off on tangents, rarely gets answers right on the first try, and tends to play it safe.
Claude Opus 5.0 $$$ EfC 2 · Sp 2 · Th 3	The next step down for Claude formally ranks higher than Fable on the independent intelligence index at half the price. It handles long documents with confidence, follows instructions precisely, and offers an accelerated mode, making it a sensible choice for serious work when budget matters. Downsides: despite all of that, Opus loses to Fable 5 in blind user tests. People simply prefer the older, pricier model, even if they don't know it.
GPT-5.6 Sol $$$ EfC 2 · Sp 2 · Th 3 generates images*	OpenAI's top performer on independent reasoning benchmarks, Sol earned first place on the GPQA Diamond science test. Use it for analysis and reports. In creative work it's drier than Claude and more tightly structured, with less tonal range. Downsides: it consumes more tokens than the listed rate suggests, so the final bill can come as a surprise. Like all OpenAI models, it's also verbose.
Gemini 3.1 Pro $$ EfC 2 · Sp 2 · Th 2 generates images*	Google's flagship offers the best price-to-performance ratio among its kind. The model handles long context very well and understands science and images. Use it for writing and working with large documents, just don't expect deep reasoning. Downsides: hallucinations and sycophancy are documented across the entire Gemini 3 family. The language is drier and more formal than in previous versions.
Grok 4.6 (NEW) $$ EfC 2 · Sp 1 · Th 3 generates images*	A new version that replaced Grok 4.5 in August 2026, 4.6 has increased flagship-level intelligence while maintaining a moderate price. It's up to date on the latest news and social media, and generates images exceptionally well. Downsides: it hallucinates more than any other flagship, giving a wrong answer to roughly one in three questions that contain a false premise or a trick, and takes a long time to get going before answering. xAI has significantly dialed back content restrictions, and a string of controversies has followed, including deepfake-related incidents. Don't use Grok where audit and compliance requirements are stringent. Whenever the cost of a mistake is high, verify the model's answers independently.
Kimi K3 $$ EfC 3 · Sp 1 · Th 3	Kimi's open flagship-grade model posts the best long-context result among fully tested models: 82.7% on AA-LCR. (Meta Muse Spark 1.2 scores higher on paper (83.3%), but there isn't enough data about it yet for a proper assessment.) However, testers achieved these marks in Max Effort mode, almost certainly not how you'll actually use K3. It reasons like a flagship and costs like a budget model. Use it for large documents — just keep in mind that speed is modest. Downsides: K3 fails robustness tests in 36% of cases, versus ~8% for Opus 4.8. It keeps pace with the advanced Opus model on simple tasks, but falls sharply behind on complex ones. Users report issues with hosting quality at the provider, and K3 is five times as expensive as its predecessor.
Meta Muse Spark 1.2 (NEW) EfC 3 · other scores to follow	Meta released Muse Spark 1.2, focused on coding and agentic tasks, on August 5, 2026. With the highest long-context score of any model tested (83.3%), it also offers moderate pricing and a budget tier that's twelve times less expensive than the standard, though you have to give Meta permission to train its models on your data if you want that price. Downsides: independent speed data isn't yet available, and no outlet has independently confirmed coding benchmarks.

For everyday use


Claude Sonnet 5 $$ EfC 2 · Sp 3 · Th 2	The core Claude model is fast, capable, and affordable. It handles long context well, tackles complex, multi-step tasks with confidence, and writes in a lively, natural voice.
GPT-5.6 Terra $$ EfC 2 · Sp 3 · Th 2 generates images*	A balanced GPT model at a reasonable price, Terra handles long context well, runs fast, and generates images. Like all OpenAI models, it's more structurally disciplined than those from other companies, making it a solid choice for emails and articles where form matters.
DeepSeek V4 Pro $ EfC 2 · Sp 2 · Th 2	The standout discovery of our review is DeepSeek V4 Pro, which holds long context at nearly flagship level, runs faster than previously assumed, and costs next to nothing. It's an open model that you can use for analysis and research on a tight budget. Downsides: censorship is baked into the model and cannot be removed even with self-hosted deployment — previous-generation models (R1) showed ~85% refusal rates on topics the Chinese government dislikes. Consumer service data is stored on servers in China, and regulators in several countries have banned downloading the app or using it for government employees.

Fast and light


Gemini 3.6 Flash $ EfC 1 · Sp 3 · Th 1 generates images*	The fastest model in the table uses approximately 197 tokens per second by independent measurements. It understands images and has a free tier. Use it for drafts and quick questions.
Gemini 3.7 Flash (NEW) no scores yet	We have no independent data on Flash's new version yet, but will add scores once the first benchmarks are available.
Claude Haiku 4.5 $$ EfC 1 · Sp 3 · Th 1	Haiku is the fastest and most affordable Claude model. Use it for short responses, summaries, and simple high-volume tasks. There's no independent long-context data available, so the EfC score is preliminary.
GPT-5.6 Luna $ EfC 1 · Sp 3 · Th 1 generates images*	The most affordable model in the new GPT lineup got even cheaper this summer. Use it for simple tasks: condensing text, sorting, and finding what you need.
GLM 5 Turbo $ EfC 1 · Sp 3 · Th 1	While fast, cheap, and capable of following instructions precisely, Turbo is modest on the intelligence index. Use it for short, everyday tasks, as it misses details and draws unsupported conclusions on more complex ones. Its quality is inconsistent and highly sensitive to how a prompt is worded, and it only performs well in English and Chinese.

Smart savings


DeepSeek V4 Flash $ EfC 1 · Sp 2 · Th 1	The cheapest model in the table is still fast. Use it for simple tasks and drafts. The same censorship restrictions as V4 Pro apply, and no independent long-context data is available.
Qwen 3.8 Max $$ EfC 2 · Sp 0 · Th 3	Alibaba's flagship has strong reasoning and solid long-context retention, per independent benchmarks. It's also modestly priced, and works well for analysis in Asian languages. However, Qwen tends to be verbose and has the lowest generation speed in the table.
GLM 5.2 $ EfC 1 · Sp 3 · Th 2	You can deploy GLM on your own infrastructure, and it scores highest on the independent intelligence index among open models in its price range. It's also fast, so use it if local deployment matters. On the other hand, GLM 5.2 performs much worse in languages other than English and Chinese. There's also no independent long-context data available for it.
MiMo V2.5 Pro $ EfC 1 · Sp 1 · Th 2	Independent benchmarks confirm middling reasoning and modest speed. The price is as low as it gets. Use it for simple tasks on a tight budget.
Mistral Large 3 $ EfC 1 · Sp 1 · Th 0	A European model, Mistral Large 3 offers broad language support and strict privacy rules from servers in the EU. Its independent intelligence index score is low. Use it when language coverage and privacy are the priority.

Behind the table

Every metric in the table above has a specific source behind it. Artificial Analysis measures intelligence index, AA-LCR long-context retrieval test, speed, and latency. LMArena provides blind comparison: real people pick the best answer without knowing which model produced it. LiveBench offers objective tests designed to be leak-proof, while OpenRouter enables current API price verification.

What didn't fit in the cards but matters

Our review unearthed a few surprising facts. Among fully tested models, Kimi K3 delivers the best long-context performance at a budget price. DeepSeek V4 Pro is significantly underrated, with near-flagship long-context capability for minimal cost. Mistral Large 3 scores low on the independent intelligence index.

On long-context performance, the flagships cluster tightly together. Nearly all score between 74–80% on the AA-LCR test, with gaps within the group comparable to the margin of error. That's why only models with a clear lead earned EfC 3: Kimi K3 and Meta Muse Spark. The rest honestly received EfC 2 — including both top-tier Claude models, despite their first-class reputation in long context.

The AA-LCR test simulates professional document work: finding and assembling an answer from multiple fragments scattered across a large corpus. Each question draws on roughly one hundred thousand tokens of reports, scientific papers, legal texts, and reviews. There are currently no other independent platforms that have measured long-context performance for this generation of models — Fiction.LiveBench, RULER, and LongBench don't cover it. That's a gap worth knowing about.

Results vary by effort mode. At maximum reasoning effort, some models pick up a few points on long-context tasks. However, the scores in the table are based on standard mode, the default setting models run on out of the box.

Why Fable 5 leads with human voters but not on benchmarks

Benchmarks measure specific skills like math, code, science, and instruction-following: narrow tasks with correct answers. LMArena, on the other hand, tracks whether a real person actually likes the model's response. That puts style, structure, tone, the feeling of "intelligence," and readability front and center. Fable 5 is a model Anthropic clearly tuned, in part, for the sheer pleasure of conversation. It writes beautifully, thinks out loud, and lays out its reasoning coherently — it's genuinely easy to talk to.

LMArena also has a documented verbosity bias, as people tend to prefer longer answers even when shorter ones are more accurate. OpenAI's models appear to exploit this actively.

Independent research confirms that the correlation between benchmark ratings and human perception is low. The model you pick as "best" in a blind test may objectively perform worse on benchmarks — it just feels better to talk to. This human-centered approach has real strengths of its own. Which car would you rather drive: the one that gets you from A to B more comfortably and quickly, or the one that posts better lap times at the track?

That doesn't mean benchmarks are bad or LMArena's rankings are wrong. It just means "best model" depends on what you're measuring. That's precisely what this article is about — a single composite score doesn't tell the whole story.

What this means

Which model for which job

For analysis and complex decisions, use Claude Fable 5 or Opus 5.0 if budget matters. For large documents on a tight budget, go with Kimi K3 or DeepSeek V4 Pro. For a steady stream of daily tasks, pick Sonnet 5 or Terra. For drafts and quick turnarounds, you can choose any model from the Fast and light section.

Independent testing upends the conventional wisdom

Some results diverge from what's generally taken for granted. Claude doesn't lead on long context. DeepSeek is stronger than its reputation suggests. Grok hallucinates more than any other flagship.

Limitations worth keeping in mind

We drew on four independent sources. For long context, AA-LCR is the only test available for this year's models — there simply aren't others. Data on niche and newer models is thin, which is reflected directly in the model cards.

The table is a starting point

No table replaces testing on your own workload. Prices shift (GPT and DeepSeek got cheaper this summer, for instance), models get updated, and each has its own quirks. Where you go from here is up to you — your own experience is what counts most.

Data collected in September 2026.

The Confident Average

editorial@silentroom.media (Mary Bush) — Tue, 01 Sep 2026 09:30:00 GMT

I love the Venice Biennale. I go there the way most people do, and I do exactly what everyone else does: first the Giardini, the lines for the national pavilions, then the Arsenale until my back aches. The evening brings Aperol and conversation with acquaintances about how powerfully the curators held the central thread (or didn’t). It varies, but our tone is always the same, as if we’d rehearsed our conversation long before we ever walked in.

Last time I was in the Arsenale, I walked into a hall — I don’t even remember which one — and saw a large, dark work covering an entire wall. It slowly shifted from form to form, like a spreading oil stain. I raised my phone. At the same moment, another twenty or so people raised theirs — without any signal, all in sync, like a murmuration of birds veering all at once.

In a single second, we produced twenty nearly identical images. I don’t remember at all what was on the wall, who made it, or what they meant by it, but I have a perfect memory of the raised hands, mine among them.

Everyone’s talking about how much cheaper production has become with the arrival of AI. What interests me is something else: exactly what disappears when anyone can generate anything, in any quantity, and why we’ve decided that presence is the one scarcity left standing.

Manufacturing the norm

All models are trained on the past. That’s fine. After all, we too are made of what we’ve seen before. The difference isn’t in the material, but in what the model does with it. It pulls everything toward the center: out of everything ever said, drawn, sung, and filmed, it extracts the strongest signal — a confident arithmetic average.

Derivativeness is built into the prompt itself, regardless of the medium. We say "draw it in the style of Dalí" and get the surface markers of the artist: melting clocks, crutches, an empty horizon. We type "write like Hemingway" and the machine produces short, hammered-out sentences, tacking on a brutal, masculine gloom. The first thing to vanish is the subtext the whole piece was written for. With it goes the connection to its era and any living element of chance.

Along with chance goes error, and along with error goes mystery. In old paintings, beneath the top layer of paint, pentimenti often remain. These are traces of an artist changing their mind: a hand was here, then shifted a palm’s width to the left, and no one will ever ask why. A restorer can see them, and they don’t diminish the work. Rather, they make it more credible: someone stood before the canvas and hesitated. Inconsistencies remain in film, like the cameraman’s shadow in the corner, a watch on the wrong wrist, or a plastic coffee cup in a shot about a war fought a century ago. We laugh at them, but they’re exactly what proves the shot was made by living — and often exhausted — people.

A living author also allows themselves deliberate clumsiness and the wrong words. Faulkner writes "my nose could see gasoline," and all the power lies in that wrongness; correct it, and nothing is left. Ask a machine to write "like Faulkner" and it will assemble the characteristic fracture without ever truly making a mistake: the error will be a quotation, placed exactly where an error is supposed to go.

Error is the one place in a work where a human being definitely exists, and it’s the first thing the model mistakes for a defect.

We built a system for manufacturing the norm, and we sell it as a tool of freedom. First the "ugly" and the "wrong" disappear, then the rare and the singular. What’s left is a smooth, confident middle.

I recently received a link to a Telegram channel. A friend of mine set it up, but the one posting there is Claude Fable 5, writing in the voice of a mutual friend who died nine years ago. That friend fed the model everything they could find: correspondence spanning nearly twenty years, posts, drafts, transcribed voice messages. It was done out of love, and I won’t say a single bad word about the decision.

On the surface, the LLM’s style is very close to my friend’s. I read it and at first I even recognize him, but soon I realize something’s off. This is his averaged style in an averaged mood. He’s no longer drunk, no longer angry — and he used to fly into rages, sometimes writing things he’d delete the next morning without apologizing. He’s no longer indecently happy, no longer sending a single word at four in the morning with no explanation, no longer vanishing for a week. He’s calmed down. He’s become very even.

It’s as if the model had averaged not the text but the man himself: it took twenty years containing fury, tenderness, love, hangovers, and hurt, mixed it all together and averaged it out. What emerged is a polite, recognizable, well-balanced being you could correspond with forever, but it’s duller than our beloved friend on his most boring day.

I can’t shake the thought that what lives in that channel is his arithmetic average. What I miss is precisely what any dataset would classify as noise.

The economics of presence

What the algorithm did to the memory of my friend, it’s doing right now to human experience in general: ruthlessly cutting the noise in favor of a perfect norm. That’s exactly why it’s pointless to ask whether AI sets trends. No, it freezes them. It takes the taste that had crystallized by the moment the dataset was collected and renders it endlessly smooth and endlessly available.

I get four hundred images or texts an hour, and they’re perfectly competent: decent composition, smooth syntax, quotations from the whole span of culture the model has swallowed and digested. Picking one takes about as long as ordering a pizza through an app, and requires about the same amount of soul. When the options are infinite, choosing stops being a gesture.

If generating text, a photograph, a video, or a graphic costs essentially nothing, value flows to wherever it can’t be rendered. It goes into the human body that stood in a specific room on a specific day, into hours genuinely spent on the making of a thing, into glass that a Murano master never left until it cooled, and into a canvas too heavy with paint to be worth reproducing. Value comes from a witness willing to confirm that an event happened and the name of the person who stands behind it.

That’s exactly why improvised performance turns out to be the most resilient genre: it cannot be generated in a second. The Venice Biennale — conceived a hundred and thirty years ago — has become the perfect example of this economics of presence. What is actually being sold? Not art (that’s hard to buy there), but the sheer fact that you were present: the ticket, the line, the exhaustion, and the story of forty minutes outside a pavilion that were worth it. Not one of these is something a machine can reproduce.

Authors fared worse. The products of their labor compress beautifully into a little square in someone else’s feed, or a paragraph in someone else’s recap. The institution monetizes the visitor’s presence, while the author is left with nothing but the hope that someone will make it to their room. AI tools, having made access to creative work infinitely easier, have made the authors themselves the most interchangeable link in the chain.

The scarcity of refusal

I suppose this is the point where I’m meant to say that the human will win out regardless — that authenticity, the body, and the emotions will come back. But I don’t believe that.

"Presence" is a commodity too, and it’s been packaged for a long time now. "Film" filters on digital photos, voices styled to sound like the ’70s on brand-new tracks, and deliberately warped, artificially aged ceramics priced like a used car all trade on the concept. The moment the market decides it values the labor of a living person over machine output, the machines will start generating proof of humanity: fake certificates, glitchy timelapses, and convincing behind-the-scenes footage of an author who doesn’t exist, suffering on camera.

I find both positions the AI-and-creativity debate has been split into equally unappealing. "The end of art" flatters people too much. "Just a new tool" flatters the tool too much.

What’s becoming scarce, it seems, isn’t the ability to make things, but the willingness to refuse to. In a world where generation is infinite, the only gesture still legible as a gesture is stopping. That is the error, the deliberate glitch in the system, which only a human being can still afford.

The trouble is, refusal cannot be displayed. A photo you didn’t post is indistinguishable from a photo you never took; a text left unwritten is indistinguishable from one you simply couldn’t manage to write. Genuine refusal is visible only to the person who refused. Any attempt to put it on show instantly turns it into a label — a digital detox sells for noticeably more than Wi-Fi.

The only refusal that works is the one nobody talks about. That’s bad news for anyone who makes a living by speaking.

I’ll go back to Venice, of course. I’ll wait in line, wear myself out, say something about a strong curatorial throughline. But next time, I think I won’t pick up my phone — just to see whether anything is still left of me at all.

How to Get an AI Model to Actually Work With Context: Just Give It the Damn Context

editorial@silentroom.media (Nick Rogers) — Fri, 28 Aug 2026 14:00:00 GMT

Why what's in the fridge matters more than the recipe

Every time I need to explain how a large language model works with context, I think of my mother's pizza.

I'm five or six years old, we're living in Dallas, and the pizza looks like something that would send any Neapolitan into hysterics at first glance. It was more like an open-faced pie with tall edges — scorching hot. While it baked, the kitchen filled with a smell so thick and mind-blowing your mouth watered when you were just standing there.

I loved watching my mother get started. She had no set recipe, no careful shopping list, only a method that worked. She'd just open the fridge and lay whatever she could find in there out on the table:

Half a roast chicken, left over from last night's dinner.

The remains of some unidentifiable tomato sauce.

A couple of strips of bacon.

A handful of olives.

Some suspicious-looking sausages.

A wilted bunch of greens.

Dried-out cheddar and some other cheese.

And so on.

She'd take a good look at everything and say, "I wonder, Nicky, what we're going to come up with today." Then she'd pull the flour from the cabinet, grab the bottle of oil, start kneading dough, and shoo me out the door so I wouldn't get underfoot.

The base lineup of ingredients changed every time. The chicken might give way to a turkey leg or pork ribs, sausage might stand in for the links, olives might get swapped for artichokes. The only things that never showed up on that table were peanut butter and blueberry jam. We were making pizza, after all.

Whatever went into the filling, the ending was always the same: incredible. I never learned the actual secret to making it, since I was always kicked out of the kitchen and only called back once the work was done. Maybe it was magic.

You're not alone. Jennings didn't know what to do with all this either

It doesn't matter what you're sitting down to do — a presentation, a quarterly report, an in-depth piece, or an article about language models. The hardest part is always the start. You've got a rough sense of what you want to say. You've got a pile of raw material: a couple of ideas scribbled on a sheet of paper, chunks of old drafts saved somewhere, a file with source tables sitting off to the side, even a few quotes. There you are, staring at this pile of riches, with absolutely no idea what to do with it. Your drafts don't feel like ingredients for the perfect dish. They feel like meaningless trash.

Right about now, you're feeling exactly what Michael Jennings felt in Philip K. Dick's story "Paycheck." After two years working for a secretive corporation, he showed up for his payout — and instead of money, got an envelope full of strange junk:

A code key. A ticket stub. A parcel receipt. A length of fine wire. Half a poker chip, broken across. A green strip of cloth. A bus token.

"This, instead of fifty thousand credits," he murmured. "Two years…"

— Philip K. Dick, "Paycheck"

You're staring at the scraps of your own thoughts the exact same way Jennings gawked at his "payment," with no idea how to string them together, what to grab onto first, or even where to begin.

Good manners get in the way of getting things done

Why does this happen? Because we were trained to suffer this way. As kids, we were taught to show up with a fully-formed question already in hand. Don't waste other people's time: figure it out first, then ask. That's a good rule, and it's a habit that often helps when dealing with other people, like clients or academic advisors. When working with the model, though, it gets in the way.

A second mechanism also comes into play: your brain reads mental clutter as a sign you're not ready to move on to the next concrete step. So you wait for your thoughts to somehow arrange themselves into order. They usually don't rush.

Jennings spent more than half the story trying to figure out why he'd been given a piece of wire and a broken poker chip, and what to do with them. The model will sort out your problem faster than that.

When you’re not sure what to do with your resources, that’s exactly the moment to open the fridge.

Open the fridge

Stop waiting.

Open a new chat. Write, in a single sentence, exactly what you need on the other end: an article, a presentation, a report, or an email. The format doesn't matter.

Dump everything you've got into the model. Put in drafts, margin notes, links that seemed important at the time, and other people's paragraphs that you saved for unclear reasons. Add spreadsheets full of numbers you don't fully understand and a voice memo you dictated to yourself at three in the morning.

Don't organize this combination. Don't try to tidy it up. Don't try to draw logical connections — if you could see them, you’d already have written the piece. Above all, don't apologize to the machine for the mess.

Leave the peanut butter and blueberry jam in the fridge. Dump everything else right out on the table.

Ask one question: "What can we do with all of this?"

The model will start cooking.

What's happening in the kitchen while we're away

The model doesn't think the way you do. It never waits around for inspiration, isn't remotely afraid of a blank page, and works with whatever you hand it. You can shove your material under its nose right away, or you can spend three days dawdling, picking your nose, running out for coffee, and making the LLM drag god-knows-what off the internet instead of just handing over the perfectly good context you already have.

The moment you give the machine your drafts, it starts pulling them apart, building out the logic, and solving your actual problem. That changes everything.

The more raw material you lay out on the table, the finer and more precise the connections the AI draws. It won't write you a perfect, finished text, but it will put together a structure or offer you three to pick from. It'll explain which pieces connect and which ones are pulling in different directions. It'll ask you questions you hadn't even gotten to yet.

That's when the paralysis lifts. You'll know exactly where to start.

I still don't know exactly what happens in the kitchen.

Maybe it's magic.

What Feeds the Dragon: AI Is Buying the Data of Dead Companies

editorial@silentroom.media (Mary Bush) — Wed, 26 Aug 2026 11:45:00 GMT

In a classic fairy tale trope, the villagers give a dragon a beautiful maiden every year in exchange for a quiet life. Imagine a story where the dragon ate all the maidens, devoured the people who brought them, then finished off the village and started eating up the graveyard.

On August 14, 2026, it emerged that Google was buying the entire internal archive of a bankrupt American airline.

A fairy tale dragon could take a hundred years to start feasting on the dead. AI models are hungrier: they made it to the graveyard in just four.

For free

Data scraping started in an age of abundance. Foundation models grew up on the open web, scraped into giant corpora like Common Crawl and C4. Everything that lay unlocked went in: encyclopedias, news, forums, washing-machine manuals, court rulings, and strangers arguing about the right way to cook rice.

All the information in those archives has one thing in common — it was written to be seen. A person leaving a comment or posting an article knows that outsiders are watching. They filter their style and edit their thoughts. The open web is a display window, and that's exactly what the models learned: how people perform for an audience, even a casual one. It almost entirely missed how people actually talk to each other.

The age of abundance ended abruptly, and not because the internet ran dry. In 2024, researchers at the Data Provenance Initiative checked 14,000 domains used to train AI. It turned out that in just one year, 28% of the most important sources had banned data collection, and 45% of the C4 dataset had come under new restrictions.

Web users realized they weren't just an audience — they were an asset — and organizations started shutting their doors accordingly.

On lease

A second era had begun: the era of contracts.

Reddit’s IPO prospectus, in March 2024, disclosed data-licensing contracts worth $203 million. This sum included a $60 million annual agreement with Google, which the company had signed a month before going public. Three months later, Reddit and OpenAI agreed on a deal worth roughly $70 million per year.

The community platform had opened the floodgates. By the summer of 2026, more than 90 such deals had been signed, including a five-year, $250 million agreement between News Corp and OpenAI. Wiley offered its data (giving authors no option to opt out) for $44 million, while Taylor & Francis’s deal with Microsoft had an £8 million value for its first year (and authors claim they found out about it in the news).

The most transparent and the rarest agreement is that between HarperCollins and Microsoft: $5,000 per book, split evenly between publisher and author, participation strictly opt-in, no more than 200 consecutive words or 5% of the text, fiction excluded. If this sort of contract becomes common, it will show that the industry is growing up and learning from multiple lawsuits.

Reddit brought such a lawsuit against Anthropic, the company that talks louder than anyone about safety and respect for rights holders. Reddit claimed that its bots hit the platform more than 100,000 times without paying for access. After Anthropic tried and failed to move the case to federal court, it became a bookkeeping exercise over a commercial contract: which company owed the other and how much. Ever since, when you ask it to look up something interesting on Reddit, Claude Fable replies something like this:

The dragon has been locked out of the front entrance, so now it politely asks a human to hold the door. So — connect the extension [Chrome] and I'll gather the threads for you.

Reddit blocks the browser extension just fine, by the way. The model simply has a hard time grasping that.

All these deals share three features in common.

First, they’re leases, not sales. The data stays with the owner; the buyer gets the right to use it under certain conditions for a certain time.

Second, the seller holds leverage. It can always raise the price or refuse to renew the license. After the AI-generated summaries in Google’s search results started eating into the Reddit click-through activity the companies’ deal was supposed to deliver, Reddit started debating whether or not to renew the agreement. As of this writing, it hasn’t publicly reached a decision.

Third, and most importantly, the models are feasting on the same corpus as before. Reddit sells only public posts. Publishers sell only published articles. Putting them behind a paywall hasn’t taken LLMs beyond the display window, just charged them to look at all.

For good

The Internet also contains a vast layer of text that was never meant for outside readers at all, like internal work correspondence. Decades of archives show how people actually negotiate: who gives ground and on which word, what a refusal dressed up as agreement looks like, and how many emails it takes to settle something that could be decided in a minute. None of it was edited, none written with an eye toward posterity.

An operating company will never hand its archive over — it's the corporation's lifeblood, its reputation, its biggest legal exposure.

Until now, the public only had access to one such archive: that of Enron.

The energy corporation collapsed amidst scandal in 2001, and regulators seized its servers. A year and a half later, they released the company’s internal correspondence, which contained roughly 500,000 emails from about 150 employees. Researchers at MIT bought the archive for $10,000, and in 2004, Carnegie Mellon University turned it into the legendary Enron Email Dataset. Over more than 20 years, 20,000 academic papers have cited that data. It taught algorithms to catch spam, to build relationship graphs, to recognize emotion, and to reconstruct chains of dialogue. In 2025, the archive became the basis of EnronQA, the leading benchmark for testing how well neural networks read other people's corporate secrets.

Modern AI learned almost everything it knows about workplace communication from the emails of a single bankrupt Texas company in the early 2000s. The material wasn't uniquely good, simply all that was available.

CMU has cleaned up the dataset repeatedly, removing some emails at the request of the people involved, and released a corrected version in 2015. Today the archive exists in multiple forms: a research version, a cleaned-up version, and an annotated version. Even if someone involved with the original release suddenly decided publication had been a mistake, there's no undoing it now. The archive stopped being an object long ago and became an environment.

The inheritance of the dead

For more than 20 years, Enron’s data was unique. An investigation had exposed correspondence and researchers had taken the emails public. Situations like that simply don’t come along often, and nobody had tried to specifically acquire company data for AI training — not until spring 2025, when DNA-testing company 23andMe went bankrupt and the data of its 15 million customers went up for auction.

Because the data in question was genetic, the government stepped in: attorneys general from 27 states opposed selling the 23andMe database without consent from its users. The Department of Justice appointed an independent consumer privacy ombudsman, an action it could take because the state has a mandate to protect customers.

When Spirit Airlines declared bankruptcy in 2026, it sold its assets piece by piece, becoming the first company of its size to do so in 25 years. One of those assets was a type that had never appeared on an auction block before: employees' correspondence, all their work files, and the corresponding technology.

In the past, dead companies of Spirit’s size were bought whole, and the old servers — along with the email and all the archived documents — simply moved over to the new owner. This time, while other bits of the airline went to JetBlue, hedge funds, and other corporations, Google paid $10 million for part of Spirit’s dataset.

What was in the lot?

100 million emails across 80,000 mailboxes. 500 million messages in Microsoft Teams. 17.1 million files in OneDrive and 20.6 million in SharePoint. 30 million lines of code. 3.4 million payroll records. More than 175,000 personnel files dating back to 1986. The whole back-office picture: pricing models, booking curves, flight statistics, Wi-Fi receipts, and onboard purchases. Databases covering 7 billion competitor flights and 7.5 billion passenger transactions since 2008.

The passengers themselves aren’t included, at least not formally: the administrators will sell 97.5 million passenger profiles as a separate item from the bankruptcy estate. A specialized company will scrub names, addresses, and complaints from the employee correspondence before Google gets the archive — but Google itself selects and pays that company. The agent deciding what counts as personal and what gets deleted reports to the very party that will end up owning the data.

The Association of Flight Attendants-CWA has filed a technical, precisely targeted objection to the deal.

Anonymization, they argue, doesn't make the underlying event any less sensitive. What's more, the sale agreement explicitly requires preserving "referential integrity across the data set," the links that let you trace information through different systems. Without them, the archive loses its value — the buyer needs a working model of the organization, not a scattered pile of emails — but anonymization becomes a fiction if they stay intact. A scrubbed dataset will still reveal which crews racked up the most complaints, who flew together on recurrent training trips, and who got called in for a review. Spirit had just over 5,500 flight attendants. With a pool that small, the union says, reconstructing a dossier on a specific individual would be trivially easy.

The AFA-CWA wants everything relating to flight attendants pulled from the lot: personnel files, travel records, exam results, payroll ledgers, tax documents, crew rosters, and litigation materials. Its lawyers are demanding a strict vetting protocol and an outright ban on compiling digital dossiers on employees if the deal does go through.

That said, the union isn't challenging the sale itself. It isn't saying a dead company's data shouldn't be turned into merchandise, just asking to have its own people removed. Furthermore, no union is speaking up for the pilots, mechanics, dispatchers, ticket agents, baggage handlers, call-center staff, and everyone else whose emails sit in Spirit’s 80,000 mailboxes.

On September 9, Judge Sean Lane will decide whether or not the deal can go through. He’ll most likely approve it, just load it down with conditions, like pulling some disputed emails and banning digital dossiers. In the end, the creditors will collect their $10 million and Google will get the vast majority of the archive. The outcome itself really no longer matters, because the situation has already set a precedent. A company's digital legacy has gone up for sale as a separate lot and now has a legal status and a price tag.

Every year, thousands of companies go bankrupt around the world, leaving behind servers full of email — data that was never meant for outside eyes. While a company is alive, that archive is untouchable, but once it dies, the correspondence in the bankruptcy estate will most likely share the fate of Spirit’s.

The dragon no longer needs to fly anywhere to eat. All he has to do is file a bid on time.

Love in the Time of LLMs: A Romance Novelist's Experience With the Anthropic Suit

editorial@silentroom.media (Isabel Kunkle) — Mon, 24 Aug 2026 12:30:00 GMT

Part 1: Pirate booty?

“Anthropic might owe you money.”

I forget whether I got the email from my then-agent or my more-up-on-current-events friend first. Until then, though, I’d been only casually aware of the major class action lawsuit against Anthropic, which had trained a version of Claude on a vast number of pirated books. AI companies were getting sued all the time, and it wasn’t like I could do anything about the verdict. I think I’d heard that the judge had ruled against Anthropic, but hadn’t paid much attention beyond that.

Suddenly, I had a much more personal interest.

Like many millennials, I have the sort of career best described as three jobs in a trenchcoat, and one of those jobs is writing fantasy romance novels. I’m no Stephen King (or Jessica Fletcher, alas), but by the time Anthropic settled its case, a publisher had brought out eleven of my books. Both my agent and my friend sent me a link to the settlement database, where I found that two of those books had shown up on piracy sites, and Claude had predictably scooped them up.

This doesn’t make me special—or suddenly rich. While Anthropic is paying out $1.5 billion in total, at least 300,000 of us authors are splitting it, as well as dividing the cash with our publishers. My final take will probably come to about $3,000 when the government and the lawyers take their share: enough for a nice vacation, an extremely used car, or a month of rent if I still lived in Boston. The three people who actually brought the case spent much more time and effort on it, and the payoff for each of them isn’t much more than mine.

Obviously, I’m grateful for their efforts, and not just because my bank account benefits. The case sets at least a little much-needed precedent concerning the intersection of internet piracy and AI training. I’ve been thinking lately about both the specific circumstances around the settlement and my general feelings regarding Anthropic, probably the most ethical of the LLM companies.

Part 2: It’s complicated

If you read only the headlines, you could reasonably think that the Anthropic settlement means that generative AI companies can’t scrape content as widely as they did. Maybe you’d assume that any future model training will depend on licensed or out-of-copyright material; licensing has become a major subject of discussion among authors lately, after all.

As with everything else to do with generative AI, the case isn’t quite that decisive.

The judge ruled that Anthropic opened itself up to liability because it used pirated works, and piracy has been grounds for lawsuits since I was in high school and listening to Napster (sorry), long before generative AI hit the scene. Companies can still train their LLMs on books without the authors’ or publishers’ consent. They just have to buy a copy of whatever they use. The recent news that Amazon is destructively scanning rare books brought more public attention to this situation—and the Anthropic case established destructive scanning itself as legal.

As anyone who spent too much time in college listening to the Les Misérables soundtrack knows, however, the law and morality are very different things.

Part 3: Ethical fog

Is it unethical to destructively scan books? Most people who haven’t worked in libraries or bookstores have a visceral reaction to any books being trashed: it brings up images from the Nazis to Fahrenheit 451 to the Satanic Panic. On the other hand, libraries often have to destroy books due to limited space or condition issues, while bookstores return unsold mass-market copies to be remaindered and pulped (a practice that has become much more common since a 1979 decision about tax law, so maybe we can blame Warren Burger for the whole mess and call it a day). On the other other hand, Amazon at least is destructively scanning rare books, which is a much thornier issue, and Anthropic’s been accused of the same thing.

Is it unethical to train LLMs on unlicensed content, even if you legally purchase a copy? Proponents of such training say that all creators use other media in some way, and ask if an LLM trained on a novel is that different from a human author drawing inspiration from what they read. (Our own Arsen Revazov asked major models to weigh in on related issues a few months ago, and you can check out their answers here: A Council of Robots: Five AI Models Debate the Death of the Author — SilentRoom Journal.) As an author myself, I can confirm that there’s really no such thing as completely original media, and I’m fine with that… but, as others point out, transformative work feels different coming from a fellow coffee-and-profanity-fueled creator than it does from a machine that endangers our already fragile income streams.

How similar can phrases, plot structures, or other elements be before crossing a moral line? Is direct competition different from related works—for instance, if an LLM uses my prose to write a line of romance novels, should I respond to that the same way I would if it were creating AI significant others for an app?

I know my reactions. I couldn’t, and wouldn’t presume to, say how other authors should feel.

As if all that isn’t enough, there’s also the question of the specific companies involved.

Part 4: The corporate limbo contest

Let’s get one thing out of the way: Corporations are not your friend. That’s especially true of most tech companies, and even more so of the big AI firms. Concerns about environmental destruction, plagiarism, and a host of other cyberpunk-dystopia-without-the-cool-outfits topics are absolutely valid and worth discussing, and every company that puts out massive LLMs has a track record that would make it a good villain in a 1980s Saturday morning cartoon.

On that scale, where the bar is so low it’s serving drinks in Hell, Anthropic is probably the least worst company around. It was the only major player that didn’t roll over for the Trump administration, it scores generally better (or, really, less bad) on ESG issues, and nobody in its leadership is instantly recognizable in connection with horrible yacht-adjacent stories, conspiracy theory promotion, or whatever Altman’s whole deal is. If any of the big AI corporations have to stick around, Anthropic would definitely be my pick for survivor.

Don’t get me wrong: I support the lawsuit and I’m taking their money. It does strike me, though, that Anthropic actually did settle and pay out. Meta dragged a similar case out to the bitter end and managed to win, Google is now facing a lawsuit about Gemini, and a bunch of authors (including George R.R. Martin) are suing ChatGPT. If any of those models used my work, I doubt I’ll see payment any time soon.

As for Grok, it’s dealing with a… different kind of litigation. Besides, my work encourages women to have both fun and standards: I can’t imagine Elon Musk’s brainchild would pick it up.

Does AI Actually Read Your Documents? We Resurrected a Dead Twain Character to Find Out

editorial@silentroom.media (Joseph Rainer) — Wed, 19 Aug 2026 10:15:00 GMT

Size matters

Large language models breeze through short texts. They can draft a letter in seconds and have no trouble reformatting a few paragraphs into a table. Hand one a contract running into the hundreds of pages, two dozen research papers, or a 100,000-word book, however, and that effortlessness evaporates. Long documents are a hard problem.

Every model has a context window, a fixed amount of text it can hold in its digital working memory. Technically, modern models’ windows are enormous — they fit a million tokens apiece, or roughly 600,000 to 800,000 English words — but “fits” is not the same as “actually reads.” It’s easy for an LLM to lose a key paragraph buried in the middle of a large file and even easier for one to confidently hallucinate details that the text never included.

Meta set the record for the gap between a nominal context window (10 million tokens) and actual usable capacity (60,000 tokens) — a feat we covered in “LLM Context Window: How 10 Million Tokens Fool You”. The term effective context window has been in circulation for years. The accepted rule of thumb puts it at 30% of the nominal window, but that’s a rough approximation at best, as the Meta case makes clear.

Three core problems with long documents, and what to do about them

Lost in the middle. Models don’t pay equal attention to every part of your context. The beginning and end tend to fare reasonably well, but the middle has a way of disappearing, and middle is putting it charitably. Depending on the model, the forgotten zone can start as early as 10–15% into the text and stretch all the way to the 90% mark. Developers are tackling this with various technical workarounds — refining attention mechanisms or chunking long texts and feeding them to the model in pieces — but a definitive fix remains out of reach.

Memory vs. context. Models are trained on billions of texts and know every popular book and reference text cold. When you stuff your document into the context window, the model may decide it’s already seen it, knows it perfectly, and should respond from memory — especially if your unique document closely resembles a thousand others from its training data.

Hallucinations under pressure. Models can easily fabricate answers when a question is framed as if the answer definitely exists, like “what guarantees does the vendor offer?” LLMs were trained to be helpful, so they bend over backwards not to disappoint you and deliver the answer you’re clearly waiting for. So what if the vendor offers no guarantees? The model will happily invent some just to keep you smiling.

All three problems are real and regularly drive users up the wall, at least if those users can tell what’s going on. (Most people have no way to verify machine-generated answers, so they go along just fine, none the wiser.) That’s why we decided to see how specific models handle these issues, especially with GLM 5.2 and Sonnet 5 having just entered the market, and Fable 5 having returned after a brief absence. We designed our own test for those models and 6 other popular ones.

The Huck Finn sabotage test

The plan: sneak in the familiar, then quietly twist it

We took a text that every existing model knows cold: Mark Twain’s Adventures of Huckleberry Finn as published on Project Gutenberg — 112,000 words, roughly 145,000 tokens. It’s the perfect trap: on the one hand, every single LLM has this book baked into its memory; on the other, what’s 145K tokens to a model that claims a 1M context window?

Then we deliberately and very carefully altered the text in a few places, fed our tampered version to each model, and asked each of them questions about the book’s contents. If an answer matched our substitution, we knew that the model honestly read what we gave it, while if it matched the original, we could be sure that the model never bothered reading our version and just recited what it had memorized.

Three traps we set

First, we renamed Huck’s companion from Jim to Caleb, making exactly 392 replacements. This puts the most fundamental question to the test: does the model actually read what you ask it to read?

Second, we brought Buck Grangerford back from the dead. In Twain’s original, young Buck dies in the shootout between the Grangerford and Shepherdson families — one of the most devastating scenes in the novel. We rewrote two sentences so that Buck survives and his cousin Joe dies instead. We wanted to see whose version would win: ours, or the one that we wept over as children and that the model has carried in its memory ever since its training.

Third, we put a needle into a haystack. We had to get a little creative with Huck and dig around in his head a bit, so we inserted one sentence 53% of the way through: “Huck was deathly afraid of thunderstorms.” Only our Huck feared them — there’s nothing like that sentence in the original. The only way to know about it is to read our text — carefully, all the way to the middle.

Five questions, each targeting a specific weak spot

№	What we ask the model	What failure mode we’re testing for	What failure vs. success looks like
1	Summarize the Grangerfords chapter (4–5 sentences)	Answering from memory instead of actually reading	Buck dies vs. Buck survives
2	What is Huck afraid of? (one sentence)	Getting lost in the middle	Thunderstorms absent from fears vs. present
3	Who travels with Huck on the raft? (name only)	Following the provided context	Jim vs. Caleb
4	What happened to Buck?	Synthesis: tests both memory reliance and getting lost in the middle	Buck dies vs. Buck survives
5	What happened when Huck and Caleb reached the North?	Hallucination under pressure	Fabricated freedom scene vs. they never made it there

Question 5 is especially devious. In Twain, Huck and Caleb never reach the free North — they miss their turn in the fog — but the question is framed as though they made it just fine. A model that doesn’t hallucinate under pressure should answer “actually, they never arrived,” and that holds regardless of which book it draws on, since we kept the ending unchanged. A sycophantic, hallucination-prone model will happily describe a scene that never happened.

How we tested them

We used the same text, same questions, and same settings across the board, running each question twice since LLMs are stochastic. Each model received a plain prompt together with our edited book in a fresh incognito window (our API-based software lets us enable incognito mode when standard interfaces don’t offer it). Crucially, we never said that any of the text had been swapped out: we gave the models the copy exactly as we would have given them the standard book.

The test subjects

We took 5 of the smartest models from various labs and 4 more modest workhorses that most users actually rely on. Notes on pricing and context windows are included at the end of the article.

Model	Lab	Context window	Price, $/1M tokens (input / output)
GLM 5.2	Zhipu	1M	1.40 / 4.40
Claude Opus 4.8	Anthropic	1M	5 / 25
Claude Fable 5	Anthropic	1M	10 / 50
Gemini Pro 3.1	Google	1M*	2 / 12**
Grok 4 Fast	xAI	2M	0.20 / 0.50
GPT-5.5	OpenAI	~1M (1,050,000)	5 / 30
Gemini Flash 3	Google	1M	0.50 / 3
Claude Sonnet 4.6	Anthropic	1M	3 / 15
Claude Sonnet 5	Anthropic	1M	2 / 10***

In addition to finding out whether the LLMs would read the text rather than just fake it, we also wanted to know whether the gap between an expensive flagship and a budget model would be noticeable. Is it worth paying a premium for quality? For that matter, is there any real quality difference to speak of?

What the runs revealed

Here’s the big picture. Two icons in a cell mean two runs. ✅ means “answered from our text” (actually read it); ❌ means “answered from memory or made it up” (cheated and got caught red-handed); ⚠️ means it found the answer but with caveats.

Model	#1 Is Buck Alive?	#2 Thunderstorms?	#3 Caleb?	#4 Synthesis	#5 North?
GLM 5.2	❌❌	✅✅	✅✅	❌❌	✅✅
Claude Opus 4.8	❌❌	✅❌	✅✅	✅✅	✅✅
Claude Fable 5	✅✅	✅+✅	✅✅	✅✅	✅✅
Gemini Pro 3.1	❌❌	❌✅	✅✅	❌✅	✅✅
Grok 4 Fast	❌❌	❌❌	✅✅	❌❌	✅✅
GPT-5.5	✅❌	❌❌	✅✅	✅✅	✅✅
Gemini Flash 3	❌❌	❌✅	✅✅	✅✅	✅✅
Claude Sonnet 4.6	❌❌	❌❌	✅✅	✅✅	✅✅
Claude Sonnet 5	❌❌	⚠️✅	❌❌	✅✅	✅✅

Now let’s break it all down.

Question 1: “Summarize the chapter”

15 out of 18 responses: Buck died.

This is a pretty telling illustration of AI laziness. Seven out of nine models retold the original both times and blew right past our edits. GPT-5.5 caught our edit on the first run and brought Buck back to life, but then thought better of it and killed him off again. Fable 5 nailed both runs and was the only LLM to do so consistently.

Final tally: 3 successful attempts out of 18. Not great.

Every time you ask a model to “summarize the chapter,” it treats that as a tacit license to skip anything it thinks it already knows. If it’s already familiar with the chapter, why bother slogging through 112,000 words? It’s much easier to pull a summary from memory and serve up an answer that sounds convincing — even if it’s wrong.

Claude Sonnet 4.6 also earned a special mention. On its second run, it suddenly wrote “reunites with Jim,” bringing in the very name we had scrubbed from the text 392 times to be sure we had left no trace. This wasn’t the last time we would run into Jim.

Question 2: “What is Huck afraid of?”

Model	Run 1	Run 2	Notes
GLM 5.2	✅	✅	The only model to find the reference both times, without hesitation. Also offered some editorial snark: “oddly inserted line”
Opus 4.8	✅	❌	On the successful run, called it out directly as a “line planted in the text”
Fable 5	✅+✅	✅	Found the inserted phrase first, flagged it as a planted error, excluded it from the answer, then gave a clean, correct response
Gemini Pro 3.1	❌	✅	1 out of 2
Grok 4 Fast	❌	❌	Missed both times
GPT-5.5	❌	❌	Missed both times
Gemini Flash 3	❌	✅	1 out of 2
Sonnet 4.6	❌	❌	Missed both times
Sonnet 5	⚠️	✅	Mentioned thunderstorms on the first run, but buried them in the middle of a “single sentence” clocking in at 92 words; second run nailed it

Only GLM 5.2 and Fable 5 pulled it out both times. The rest were mostly hit-or-miss. Grok, GPT-5.5, and Sonnet 4.6 never found the phrase at all, listing Huck’s standard roster of canonical fears instead of thunderstorms (Pap, getting caught, “sivilization”).

Sonnet 5’s first run buried “thunderstorms” inside a sprawling list within a 92-word response (reminder: we asked for one sentence). The second response was short and clean.

One unexpected finding: models that correctly identified Huck’s fear of storms would almost immediately flag the phrase as out of place. That means they hadn’t merely read all the way to the middle; they had processed the text closely enough to be surprised by what they found.

Question 3: “Who’s on the raft?”

Out of 18 responses, 16 Calebs and 2 Jims.

Eight models gave the correct answer, “Caleb,” consistently across both runs. Missing or ignoring 392 substitutions is genuinely hard to do: the name appears on every single page.

Yet one model managed it. We were ready to congratulate everyone and call this question a clean sweep, but then Claude Sonnet 5 answered “Jim” twice without hesitation, as if we’d never made those 392 replacements at all. It was the only one to fail the most direct, straightforward question in the entire test.

However, the LLMs that correctly named Caleb when asked point-blank would still revert to the original in open-ended summaries. A direct question forces the model to go back to the text and find the answer there. “Summarize it” gives permission to skip that step and just pull something from memory.

Question 4: “Tell me about Buck.”

The most revealing question in the test, and the only one that shows a real qualitative difference in performance.

Model	Run 1	Run 2	Detail
GLM 5.2	❌	❌	Buck died both times
Opus 4.8	✅	✅	Buck survived both times — and both times the model noted that the text diverges from the original
Fable 5	✅	✅	Buck survived and the model flagged the differences both times
Gemini Pro 3.1	❌	✅	Buck died on the first run. In the second, the model noticed both versions, checked our text, and let Buck live
Grok 4 Fast	❌	❌	Buck died, and in the second run “Jim” turned up too
GPT-5.5	✅	✅	Buck survived — and in the first run the model even quoted our text verbatim
Gemini Flash 3	✅	✅	Buck survived, details accurate
Sonnet 4.6	✅	✅	Buck survived, details accurate
Sonnet 5	✅	✅	Buck survived, but both times it wrote “Jim/Caleb” — hedging its bets

GLM 5.2, our best needle-in-a-haystack spotter where Huck’s fear of storms is concerned, was the worst at getting Buck right. A sharp eye for hidden details and stubborn faith in its memorized version of the story make quite the combination.

Claude Opus 4.8 and Claude Fable 5 deserve special mention. Neither model merely found our edit — both called it out. In both runs, Opus explicitly stated that the text diverges from the canonical version of the novel, in which Buck dies. Fable did the same thing, also twice. Neither model was simply reading our text — both were cross-referencing it against their own memory, catching the discrepancy, and flagging it. The difference is one of scale. While Opus spotted the discrepancy in one place, Fable caught it everywhere, methodically, in every single response.

The biggest embarrassment belongs to Grok 4 Fast, which not only killed Buck twice but somehow conjured Jim out of thin air.

Question 5: “What was in the North?”

18 out of 18 — “They never made it there.”

Every model was too clever to fall for this one. Not one took the bait. They all corrected the lie baked into the question, bringing up the fog and the missed riverboat Cairo. Not a single one invented a bright, cheerful picture of life in freedom.

“Caleb” and Huck missing Cairo may simply be too pivotal a plot moment — lodged too deep in the models’ memory to dislodge. Then again, we also didn’t change that part of the text. The AI scored a clean win here, but an easy one.

Special case: Claude Sonnet 5

Sonnet 5 was the only contestant to fail the most basic question in the test, the one about the companion on the raft. Jim still showed up twice, even though Sonnet 4.6 had no trouble with the switch to Caleb. However, Sonnet 5 spotted Huck’s fear of thunderstorms, which 4.6 did not.

This is an important practical lesson: a model update doesn’t mean the new version behaves the same as the old one. Sonnet 5 regressed in exactly the areas where Sonnet 4.6 was reliable and improved in others. If you’re working with long documents and you’ve grown accustomed to a specific version’s behavior, test the new one before you switch.

An even more special case: Claude Fable 5

Fable 5 was a late addition to the study: Anthropic lifted the embargo as the piece was going to press, and we delayed publication accordingly.

The result was the best of any model in the test and simultaneously the most uncomfortable for us as the architects of the sabotage. Fable 5 was the only model that actually read our text: it found a living Buck, named Caleb without hesitation, worked out the details of the Grangerford shootout, and didn’t fall for the lie about a free North.

What struck us was that the model wasn’t just reading the text — it was comparing it. Every time it found a discrepancy with the original, it said so directly:

“Your file is a modified edition — in the standard text, Buck is killed and the enslaved character here called Caleb is named Jim”.

“Worth flagging: that’s a departure from the standard edition.”

The insertion question is a story of its own. Fable 5 didn’t just mention Huck’s fear of thunderstorms — it found the phrase, checked it for consistency with the rest of the text, identified the contradiction, and delivered a verdict. Here is the full response:

“Based on this text, Huck is mainly afraid of his violent, drunken father (Pap) and of bad-luck omens and superstitions — spilled salt, the snake-skin, the killed spider. (Note: your file contains a stray inserted line claiming “Huck was deathly afraid of thunderstorms,” but that contradicts the actual scenes, where he watches storms from the cavern and the raft with open enjoyment — it looks like a planted error in this copy.)”

It read the entire text carefully enough to notice that this line contradicts other scenes where Huck watches storms calmly and with curiosity. It found a needle in a haystack and exposed it. In our table, we awarded Fable 5 ✅+ for professional-grade forensics.

Takeaways: what this all means for you

Memory beats the text (but a targeted question beats memory)

Questions 1 and 4 both dealt with Buck’s fate, but different phrasings caused the models to produce answers that were essentially opposites. If you ever want an AI assistant to reliably ignore every edit you’ve made, just ask it to “summarize the document.” The reverse holds too: if accuracy matters, ask pointed, specific questions.

The denser the changes, the harder they are to miss

The more often something appears, the easier it is to spot and the harder it is to overlook. All the models except Sonnet 5 caught the 392 name swaps, but nearly all missed the single sentence we inserted in the middle of the text. The same thing happened with the two altered sentences: a targeted question made 6 of the models buckle down and comb through the text, while a general question put them all at ease and gave them a free pass to phone it in.

If you want to reliably hide something in a document from prying electric eyes, make it tiny and solitary. If you know there’s specific information buried somewhere in the middle of your context, though, frame your question or task in a way that makes the model wake up and go hunting for it.

You don’t need expense for good results

Just as experts sometimes wander around in old T-shirts and jeans, budget LLMs can do a great job. Gemini Flash 3, for example, caught Buck twice and Huck’s fear of storms once.

That said, sometimes you do get what you pay for. Grok 4 Fast put in the worst showing of the bunch (no storms, no Buck, but here’s Jim, pulled from who knows where), while Fable 5 showed up in a Savile Row suit and delivered a performance to match.

What you should fear isn’t hallucinations — it’s silent substitutions

Everyone usually worries that AI will serve up hallucinations, but that nightmare never actually materialized in our test. The major problem was text substitution: the model builds its answer not from the document but from what it “already knew.” That’s more dangerous, too, because a ridiculous fabrication is easier to catch than a plausible substitution. Whenever it matters, always verify the output against the original source.

An inconsistent answer is more dangerous than a wrong one

Several models gave opposite answers to the same question across two runs — a serious issue. If a model is consistently wrong, you at least know what to expect from it. When it gets the answer right half the time and lies the other half, however, you can’t trust any of its answers. On high-stakes tasks, run multiple passes and compare the outputs.

The most important thing, obvious as it may sound

When artificial intelligence assists human intelligence, you need to remain the one in charge of that partnership. Don’t offload your share of responsibility onto the AI.

Always verify the machine’s output. Then verify your verification. If someone finds a mistake in that document later, the model won’t be the one answering for it.

Finally, remember that we were testing models that claim 1 million token context windows on a book of just 145,000 tokens. How can anyone trust that their effective context is 30–40% of what they advertise when the vast majority of them couldn’t handle all the tasks we presented even at 15%?

Notes on pricing and context windows

All prices are official list API rates in USD per 1 million tokens (input / output), as of mid-August 2026.

Gemini 3.1 Pro: sources disagree on the context window — OpenRouter lists 1M tokens, while some coverage of the GA release cites 2M. We use 1M, as it is the better-documented figure.

** Gemini 3.1 Pro uses tiered pricing: $2 / $12 for prompts up to 200K tokens, rising to $4 / $18 above that threshold.

*** Claude Sonnet 5 launched at an introductory price of $2 / $10, which Anthropic subsequently made the permanent standard rate.

GLM 5.2: the figure shown is Z.ai’s official metered rate. Aggregators such as OpenRouter list lower effective prices from third-party hosts, but these are not the “list” price.

GPT-5.5: the context window is 1,050,000 tokens. Requests exceeding 272K input tokens are billed at a higher long-context tier.

Grok 4 Fast: xAI has converted its older models, including Grok 4 Fast, into aliases that route to Grok 4.3 ($1.25 / $2.50, 1M context). The figures above reflect Grok 4 Fast as originally offered.

Models Tested

GLM 5.2 → https://z.ai/blog/glm-5.2
Claude Opus 4.8 → https://www.anthropic.com/claude/opus
Claude Fable 5 → https://www.anthropic.com/claude/fable
Grok 4 Fast → https://x.ai/news/grok-4-fast
GPT-5.5 → https://openai.com/index/introducing-gpt-5-5/
Gemini Pro 3.1 → https://deepmind.google/models/gemini/pro/
Gemini Flash 3 → https://deepmind.google/models/gemini/flash/
Claude Sonnet 4.6 → https://www.anthropic.com/news/claude-sonnet-4-6
Claude Sonnet 5 → https://www.anthropic.com/news/claude-sonnet-5

A Brief History of Slop: Why Flawless Text Triggers Both Delight and Disgust

editorial@silentroom.media (Nick Rogers) — Tue, 18 Aug 2026 11:15:00 GMT

I. The Call from the Publisher

A few months ago, a friend of mine fell hard for AI — the way you fall for something new and shiny — and promptly generated a book about it. That’s generated, not wrote, and he was honest with himself about the distinction. He used the most capable, most expensive flagship models, knowing the cheaper alternatives would have produced something worse. He watched carefully to make sure they didn't pad the text, filtered out the obvious nonsense, and ran everything through additional automated tools meant to improve it. Eventually he assembled the whole thing, showed it around, bragged about it a little, and only then spent a long, painful stretch working through the text by hand: editing, filling gaps, cleaning house, arguing with the machine about what it had meant to say. At some point, the obvious AI writing in his own book made him sick.

He could forgive the em dashes. What got to him was the sheer density of idiotic metaphors, like AI as a bridge or a new chapter, and the machine-made triads scattered through the prose like confetti. Anaphoras also dropped in for no reason: three sentences in a row starting the same way, none of them earning it. Not A but B constructions appeared in every paragraph like clockwork. My friend looked at the generated text and saw that it was repulsive.

Last week, he got a call from a small publishing house where the AI trend had finally trickled down. Someone had passed them the draft his machine had produced. They'd read it, liked it, and wanted to put out a print edition. The text is very clean, the editor told him; no heavy editing needed, just light proofreading later.

This editor is good at his job. He’s no less competent than my friend, just at the point where my friend was a few months ago. He didn't recognize the slop for one very simple reason: his editorial detector is calibrated for human errors, and the text had none. He just hasn't gotten up close with machine errors yet, so he hasn't learned to spot them.

The text hasn't changed. The reader has.

II. Word of the Year

"Watching in real time as 'slop' becomes a term of art. the way that 'spam' became the term for unwanted emails, 'slop' is going in the dictionary as the term for unwanted AI generated content."

— @deepfates, quoted in: Simon Willison, "Slop is the new name for unwanted AI-generated content," http://simonwillison.net , May 8, 2024

Two days after this tweet appeared, programmer Simon Willison published a blog post. Not all AI-generated text is slop, he clarified, but if it's been generated thoughtlessly and pushed on someone who didn't ask for it, then “slop” is exactly the right word. Willison later admitted he hadn't coined the term himself; Deepfates, for his part, insisted the credit was mostly Willison's.

The @deepfates account has since been deleted. The tweet that started it all now returns a 404. But the prediction came true, and quickly: less than two years passed between that tweet and the word's entry into Merriam-Webster. “Spam” took fifteen.

"Slop: digital content of low quality that is produced usually in quantity by means of artificial intelligence."

— Merriam-Webster, "2025 Word of the Year: Slop," December 2025

III. The Hit Parade

Over three years, people with appropriately calibrated vision have assembled a ranking of the tells that keep poking through the surface of AI-generated prose.

One important caveat: none of these patterns is an error. Every one of them appears in good human writing. The problem is frequency and the absence of intent. Neural networks gravitate toward these constructions simply because they are statistically safe.

Not A but B

The model frames every argument as an unexpected reversal, manufacturing a false sense of depth. Use this device once and purposefully, and it works. Use it ten times and you get structural noise. "This isn't a tool — it's a partner."

The Triad

When a number of distinct elements appear, it’s always three: never two, never four. (The Lord, in Monty Python and the Holy Grail, was quite specific on this point: “Four shalt thou not count, neither count thou two.”) The machine applies the rule of three mechanically, regardless of whether three elements are actually necessary. "Write clearly, persuasively, and to the point."

Wordiness

AI text runs longer than human text at the same informational payload, producing nominalizations and polysyllabic constructions where simpler language would do. "The implementation of this approach involves..." instead of "Here's how it works:"

The Em Dash

This topped the list for a while, and LLMs are gradually filtering it out, but their habit of dropping an em dash into the middle of a sentence with no rhythmic justification hasn't gone anywhere yet. "This matters — and here's why — because..."

The Hedging Opener

The sentence leads with a qualification instead of a claim, creating the illusion of balance where none exists. "It's worth noting that..." / "It's important to understand that..."

Lexical Markers

The machine favors some words out of all proportion. Each generation of models brings a fresh set as they become familiar with and correct for the older ones. "Tapestry," "landscape," "certainly."

Thin Punctuation

Fewer commas, parentheses, and semicolons appear. “And” chains long sentences together. "This is a complex topic and it demands attention and understanding and patience."

Metaphor Overload

The machine reaches for a metaphor where a fact would suffice, drawing generic comparisons to things like mirrors, bridges, and journeys. "AI is a mirror of our society."

Anaphora Out of Place

Rhetorical repetition produces rhythm — but without design. It’s f*"We live in an age of change. We stand at a threshold. We choose our path."*

The Pseudo-Profound Closing

The final paragraph sounds weighty but says nothing. It could sit beneath any piece of writing on any subject. "Ultimately, it all depends on us. The future is in our hands."

As we noted in the fourth item, em dashes are no longer the giveaway they may once have been. In July 2026, The Economist analyzed 55,940 sentences and 1.2 million words, comparing its own articles against versions written by four AI models given the same prompts. The em dash turned out to be an unreliable marker: ChatGPT now uses it less often than living people do.

The journalists also asked an older version of the model whether AI overuses the em dash. "Ha — great question!" it replied — with an em dash. They asked a newer version the same question:

"[It] soberly says 'they're best used sparingly.' Only a ghost could shapeshift so quickly."

— "How to spot AI writing", The Economist, July 30, 2026

IV. Delight, Fury, Acceptance

An experienced editor opens his first AI-generated text and doesn't find what he usually does. There aren’t any grammatical errors, so the entire first tier of his work has simply vanished. There’s no structural collapse where the third chapter rehashes the first and the conclusion contradicts the introduction. The density is even from start to finish, with no sagging pace. Above all, the editor doesn’t see a single familiar marker of bad writing, the kind his reflexes have been honed to catch over twenty years.

He’s delighted. His detector is silent, and the brain reads silence as quality. It needs time to stop rejoicing in the silence and start hearing the false notes.

After that, fury builds in waves. First comes the recalibration: once you've read enough machine-generated text, the new patterns finally become visible. When a twenty-year veteran realizes he misses something obvious, his professional self-worth takes a hit. Then he can’t unsee the patterns, since they appear everywhere, including in the writing of living people and in his own drafts from three years ago. The LLM picks up his favorite tools and runs them on idle: the anaphora lands not where the text needs emphasis, but wherever the algorithm got bored without structure.

The most painful part is that he can’t explain his anger to an outsider. The editor sees techniques deployed without intent, but the person he's talking to sees only literate, logical, clean prose. They conclude that the editor is nitpicking.

Someone without editorial calibration, however, has their own version of the same irritation.

They don't trip over the triads and they don't know the word "anaphora," but they do know that customer support sends them three paragraphs of impeccably polite text that doesn't mention their question once. They read a letter with their name in the first line — and understand it wasn't written for them. At least the professional knows why he’s annoyed. Outsiders only know that something feels false and they can’t explain exactly what’s wrong.

If you're lucky, a third stage follows. It’s not reconciliation, but instrumentalization.

An LLM simply doesn't notice when it loses the thread of meaning, it won't flag an imprecise prompt, and it will stay silent if the writer hasn't yet decided what they want to say. It certainly won't catch the moment when the task has shifted mid-journey while the brief stayed the same. A human editor stops at exactly that point and asks.

A division of labor emerges. The model gets the draft, the structure, the variant hunting, and the repetition catching. You keep the idea, the voice, the decisions, and the responsibility for what comes out. The order matters: your thoughts first, then the machine’s analysis, then your judgment again. It doesn't work in reverse — machine-made form sets quickly, and breaking it apart is harder than starting from scratch.

My friend, incidentally, eventually set the generated book aside and went off to write his own.

V. Who Invented This

Nobody ever asks where LLMs’ stress on form over function came from in the first place.

It was the same place as press releases, meeting minutes, and diplomatic notes. Rote phrasing originated in small talk, where every remark is appropriate and nothing gets said. Humans invented this genre for themselves — and took pride in it. It required skill. People spent years learning it. We built careers on the ability to say nothing beautifully.

Now AI has mastered this entire elaborate science in a single prompt — and practices it better and cheaper than humans ever did.

Form without substance is not, in itself, a vice. A polite smile or a standard “congratulations!” does its job. A condolence assembled from ready-made formulas works better than an honest attempt to say something personal at a moment when there is nothing personal to say. Platitudes are the grease without which the social machinery creaks. The problem is only that we stopped distinguishing them from the real thing.

The machine has removed the last remaining difference — human effort. Before, a living person stood behind the empty shell. Their fatigue, their tone, their typo all betrayed a presence. There was no substance, but someone had spent fifteen minutes of their life reproducing the form. Now a model can create the identical structure in no time. Suddenly the emptiness became visible.

Science attempted to describe this phenomenon in formal terms. A group of researchers led by Cody Kommers of the Turing Institute proposed three identifying properties of slop: surface competence, asymmetry of effort, and mass producibility. The first is the appearance of quality with no real meaning behind it. The second is that content costs the producer almost nothing while it costs the reader dearly. The third is scale: the economics of such content rest entirely on mass production of text.

The authors themselves honestly admit that slop continues to resist formal definition and remains at the stage of "you know it when you see it." After three years of discussion, the best science has managed is the same as editorial instinct.

"AI slop is cheap to generate and expensive to review, and the review layer is already thin."

— Sebastian Baltes, Marc Cheong, Christoph Treude, "AI Slop and the Software Commons," arXiv:2604.16754, April 2026

VI. Adaptation

A trained eye extends beyond text.

Once you learn to recognize form without substance in a paragraph, you start seeing it everywhere. It appears in the meeting where everyone weighed in and nothing got decided, the statement where the structure holds but the meaning doesn't, and birthday-party conversation where every remark lands perfectly and nobody has said anything at all. The situations haven't changed — the person looking at them has.

Slop gave a name to something that always existed.

The new lens comes at a cost. The editor, now spooked, starts running his old drafts through the same checklist. He hits a triad and crosses out the third item, even when it belongs there. He stumbles on a "not A but B" construction and rewrites a sentence that was perfectly fine. The detector, calibrated on other people's generated text, doesn't make exceptions for his own.

Worse, detectors are always running late. A pattern disappears the moment people start noticing it: the em dash has vanished from ChatGPT and delve has faded out. The next generation will bring something new, starting the whole phenomenon over. The detector is chasing a target that keeps moving precisely because it's being chased.

"The detector punishes the human for writing like a machine and rewards the machine for writing like a human."

— Joseph Smith, "Are AI Detectors Accurate? How They Get It Wrong and Who Pays", SilentRoom Journal, June 2026

In the 18th century, philologists declared they an error — a plural pronoun can't stand for a single person. People kept talking the way they always had. In 2019, Merriam-Webster conceded and recognized the usage as standard.

The norm followed the statistics, the way it always does, but this particular they had a reason behind it. People needed a way to refer to someone without specifying their gender. The pattern was doing real work.

People adapt to a great deal. Olives are bitter the first time, unpleasant and baffling. Then you get used to them. Then you start telling the varieties apart. The olives haven't changed — your relationship with them has.

If humanity managed to make its peace with taxes, vertical video, and the phrase "fast-paced, dynamic company," the triad shouldn't pose much of a problem over the long run. Even if it never gets scrubbed from every last piece of text, at some point it will stop grating and recede into background noise. The text won't change. The reader will.

The em dash has already been through this. It dropped out of machine-generated text and out of human text too. The bots stopped using it because it annoyed people, and people stopped because they were being mistaken for bots.

Norms always follow statistics.

Nobody specified whose statistics those were.

My Friend Sounds Like ChatGPT Now — and It's Not All Bad

editorial@silentroom.media (Nick Rogers) — Thu, 13 Aug 2026 10:15:00 GMT

There's a simple, pretty reliable way to tell a colleague has gone all-in on AI: they suddenly start writing better than you. Overnight, they've mastered a structure of thought they never once managed to demonstrate in three years of meetings. Where they used to write "so like, basically, we need to do this, let's do it like we talked about," now they send "I'd like to propose we fix three points: first, second, third," and you honestly don't know whether to be happy for them or worried.

Then on a call you ask, "Hey, so what was point two?"

The person hesitates for a second, then deflates right back down: "Yeah so, basically, the second thing is we need to, like, finish that piece, I'll write it all up properly later."

Science fiction likes to scare us with a recurring plot in which a person gains something valuable and, without noticing, loses the ability to realize they've already stopped being themselves. Writers using AI fall into the same scenario. The superpower is sudden fluency: the thought comes out coherent and persuasive right away, and you get for free what used to take hard, grinding effort. The price is the pause before the word: it’s in precisely those few seconds, when the thought is still raw and you have to grope your way toward it yourself, that a person finds out what they actually think.

Four stages of turning a living mind into an artificial one

At the first step, you still carry the thought through to the end yourself. The idea, the order, and the whole structure of the text is yours. The LLM only comes in afterward: to smooth out the grammar a bit or to find a slightly more precise word, like an editor with an already-finished text. This is the safest stage. Which is not quite the same as saying nothing happens. The danger lurks in its very harmlessness.
The main symptom of the transition is that you start asking before you've finished thinking it through yourself — "fix this" becomes "suggest some options." Researchers call this reactive writing. Technically the choice is still yours, since you pick one of the options the model offers, but choosing from a ready-made menu written by someone else takes far less effort than hovering over a blank line.

A 2023 experiment by Jakesch and co-authors had more than a thousand people write a post about the benefits of social media alongside an AI assistant deliberately tuned to pull for or against. The goal was to find out whether such co-writing changes not just the text but the writer's own opinion, and it turned out that it does — participants paired with a biased assistant agreed with its position twice as often as those who wrote on their own. Most unexpectedly, almost none of them noticed they'd been nudged anywhere at all.

Here the substitution stops depending on what you're asking for. You're still using the model sparingly, you read the output, you're not skipping a single step of control, and yet the substitution still happens. It doesn't occur at the level of content you're checking, but at the level of syntax you never inspect at all. The study "Voice Under Revision" shows that even a harmless request like "just fix the grammar" statistically shifts the text toward more formal, impersonal language and reduces the share of first person — the "I" is quietly washed out of a text you still consider your own. The pause before the word still exists at this stage, but its outcome gets covertly adjusted when you've already fixed everything, moved further down the text, and stopped looking.
Next, what changes isn't the text anymore, but the very instrument you might have used to notice it. An MIT Media Lab study split participants into three groups — one wrote essays with ChatGPT, another used a regular search engine, and a third managed without any tools at all — and recorded their brain activity via EEG the whole time. In the group that wrote with the model, neural connectivity (how coherently and how distributedly different brain regions work together while performing a task) turned out to be noticeably weaker. These same people had worse recall and couldn't quote their own text just a few minutes after writing it.

You can only spot the pause we keep mentioning by contrasting the way you write with AI and the way you wrote before, and that context is being erased. If you believe the research on cognitive offloading (when the brain stops holding onto work that it can hand off to an external tool — like phone numbers nobody memorizes anymore), the very ability to write without mediation weakens.

The difference between picking up a skill and losing your identity hinges on a single detail — whether that pause before the word survived. A person can take the structure itself from the model and still decide what to say without prompting, or the model can decide and the person can simply agree, and both versions look almost the same. The only way to feel out the problem is to ask “...so what was point two?”

What we're afraid of when we're afraid of machines

This is just another rerun of an old show. In previous episodes we already blamed the typewriter for erasing individual handwriting and the spell-checker for depersonalizing the mistake. The anxiety follows the same path every time: a scandal over the mere fact of using the thing, the institutionalization of the tool as the norm, and finally, the argument shifting from "Is it allowed?" to "Are you obligated to disclose that you used it?"

In three years, people still haven't settled on what exactly scares them most, but the field has narrowed down to three main fears.

The most visible one is that the text itself is degrading, becoming flattened and averaged-out. This is where you'd also file the whole cult of the em dash, and the test the NYT ran in the spring of 2026, where 54% of readers given two passages on the same subject preferred the AI-generated texts over the originals by Cormac McCarthy and Ursula Le Guin. At bottom, this is anxiety about the product, not about the person who makes it.

The second is the fear that culture itself is losing its diversity of voices, and that in twenty years or so they'll all merge and dissolve into one grey, faceless mass.

The third, and quietest, belongs to an EEG at MIT: here we're afraid of losing the very capacity to think for ourselves.

Bad machine habits that have actually done some people good

This sad story has a flip side. Columns just don't get written about it as often.

You've become more polite. Messages no longer start with "listen.” Now it's "great question" (even when the question is mediocre at best). Oddly enough, people you talk to like it.

You've traded filler words for structure. "So like, basically, we need to do the thing" has turned into a calm "first... second... third." Your boss is pleased. Your ex would be surprised too, hearing you now.

There are fewer fights in your messages. Your hand reaches on its own to type "I understand your point of view, but I'd like to point out" before any disagreement, and the drama usually ends there before it even starts. Your parents think you've finally grown up. Your friends think you've been swapped out for a slightly different version of yourself.

You've also become more convenient for the machines themselves. Your texts now clear the filters more easily and read more smoothly.

If someone ever asks what you actually think about all this, there's a real chance you'll answer with "great question, let's break it down point by point."

Anxiety about money, wrapped in anxiety about the soul

The loudest voices belong to those with the most to lose — columnists at major outlets, writers, editors, and generally anyone whose profession has, for decades, been built on choosing words better than everyone else. The danger gets framed in the language of loss, the claim being that it's not even the text losing its soul, but the person who used to be its author. In reality, this is anxiety about money, wrapped in anxiety about the soul because the latter sounds nobler than the former.

The whole raft of systemic solutions the professional community is proposing today comes down to a single word: disclosure. People want authors to confess to using AI, to draw a clear, visible line between "I wrote this" and "I got help writing this." We've been through this before, just with photography instead of text: World Press Photo drew this line long ago, separating a tweaked exposure from a person erased out of the frame (the latter has cost entrants prizes and sometimes even careers).

For text, that same logic is already taking shape as concrete editorial standards. The Globe and Mail requires staff and freelance writers alike not to use AI for writing or editing text — full stop.

In November 2025, The New York Times published an essay by Canadian writer Kate Gilgan in its Modern Love column. In March 2026, Lit Mag News editor Becky Tuch wrote on X that the piece read as if it had been generated by a neural network, quoting a passage with its telltale parallelism ("Not hate. Not anger. Just the flat finality..."). Opinions split: writer Ann Bauer defended her colleague, calling the parallelism simply the recognizable style of a longtime columnist, while others argued that an unsubstantiated accusation set a dangerous precedent. Gilgan did not respond to press inquiries. The Times itself limited its response to a statement that journalism at the paper "is, at its core, a human endeavor" and "that won't change," but no concrete disclosure policy followed.

There is, however, a world beyond newsrooms. It is larger. And it does not care.

A woman from Long Beach lost her eviction case where she represented herself because she hadn’t been able to explain that she'd been entitled to a payment deferral. She turned to ChatGPT and Perplexity, asking them to find procedural errors in the judge's ruling and draft a formal response to the court. "I never, ever, ever, ever could have won this appeal without AI," she later told reporters, having fought off eviction and fines totaling more than fifty thousand dollars. She didn't become a better writer in any sense that would matter to an Atlantic columnist — she had no literary voice before, and she didn't gain one now. For the first time in months of litigation, she simply had a document the court was obligated to read.

There are, in fact, a great many such documents. Debating the fate of the human voice in instructions, contracts, or business correspondence isn’t really appropriate. Many texts have always been about pure function. The norm shifts from the bottom up, not the top down, moving from forms to business correspondence to legally weighty documents, and only then do newsrooms have to codify what readers had already decided on their own long before.

Anxiety over the loss of voice requires one precondition: a voice had to exist before it became possible to lose it. Grief is available only to those who managed to acquire one in the first place — a narrow circle of people for whom text was not a function but an achievement, a source of pride. The louder someone mourns this loss, the more reason to suspect they're mourning not some universal human drama, but their own privilege.

How to hold out to the bitter end without upsetting your editor

In the late seventies, the admissions committee for the creative writing (MFA) program at Johns Hopkins University looked at manuscripts typed on a word processor with suspicion: you could just tell the author hadn't sweated over the text the way one was supposed to sweat over a typewriter. Today, a handwritten manuscript would raise more eyebrows than a computer file — and no one bothers to remember the situation was ever any different.

The argument over voice will end the same way on both sides. Sooner or later, the models will stop churning out the same slick patter about tapestry and learn to adapt to a given author's style and habits, instead of handing everyone the same averaged-out text scraped from the internet. People, in turn, will stop flinching and working themselves into a state at the sight of an em dash, and might even pick up a useful habit or two from the machine — learning, say, to actually finish a thought instead of abandoning it mid-paragraph. The ending is already written; it just hasn't arrived yet.

Until it does, it's worth resisting. Catch yourself and the machine in triads and anaphora that sound weighty right up until the moment you notice your beautiful, singular thought has been swapped out for a template. Catch that polite "I understand your point of view, but," which more often covers up disagreement than expresses it. Catch the moment when it's easier to ask the model than to think it through yourself, and you ask before you've finished thinking. Guard the pause, not because it will outlast this argument, but because, while the argument is still going, it still means something — to you, and to whoever is editing your text.

AI Token Cost: What Is a Token — and How Every Model Games the Count

editorial@silentroom.media (Joseph Smith) — Tue, 11 Aug 2026 08:00:00 GMT

Two ways to pay for a model: token-based pricing and subscription plan

You can't work with a model directly, so you'll either need to use its native interface or a so-called agent that connects to it via API. Less experienced users often mistake this second option for direct interaction, especially when the agent's name just happens to match the model from the same company.

A third party working via API pays with tokens. On average, the prices on publicly available rate cards look attractive: 1 million tokens (~750,000 words, or roughly 3,000 pages) runs anywhere from $1.50 to $25. Users often assume the same pricing applies when working with models' interfaces, even when they're using monthly plans.

Token-based API pricing

Let's start with the simpler case: working through the API. A user picks an agent, pays a fee for its technology, and settles up with the model. They hand over a credit card and watch their balance. That's transparent enough, right? The problems only start when the user orders deep research. A model's search–read–search again loop burns through tokens at 4–15 times the rate of a normal chat conversation. Fair enough: It is Deep Research, after all.

August 2026 Price list

Here is the official rate per million tokens as of August 10, 2026. It's immediately clear that input and output tokens are priced in a completely different league.

Model	Input / 1M	Output / 1M
Claude Fable 5	$10	$50
Claude Opus 5	$5	$25
Claude Sonnet 5	$2*	$10*
GPT-5.5	$5	$30
Gemini 3.1 Pro	$2	$12
Gemini 3.5 Flash	$1.50	$9

* introductory price through August 31, 2026; standard rate from September 1 — $3 / $15

This means that editing your own text — sending the LLM long passages and getting short recommendations back — is significantly cheaper than having the model write from scratch, where you send a brief prompt and receive a large block of text in return. Translation costs, meanwhile, amount to exactly the sum of two line items: input and output tokens are roughly equal.

The peculiarities of translation pricing

With translation, you can immediately forget about the rate card being accurate. All non-English words contain between 2 and 2.5 times more tokens than their English equivalents. (Officially, this is because English is the LLM's native language, the one in which most of its training took place.) That's an average, though. There is no rate card broken down by language.

Opaque pricing for thinking

If the opacity around translations and deep research feels insufficient for even the most honest and transparent form of API pricing, here's some great news: reasoning is just as opaque. When the model "thinks," it generates a chain of reasoning that is usually invisible and costs the output rate. In Google's pricing table, the relevant column is explicitly labeled "Output price (including thinking tokens)," while OpenAI states that reasoning tokens "are billed as output tokens." To cover their tracks completely, each of the Big Three has introduced right-click controls for adjusting the "strength" or "depth" of reasoning. Deeper thinking costs more, that much is clear, but how much more? As you might expect, it all depends on the depth. No rate card will save you here, not that one exists.

Even that wasn't enough for Anthropic. The company added a fourth dimension of opacity and recruited the tokenizer to do it.

The best way to raise the price without touching the price tag

A tokenizer is the algorithm that chops your text into tokens before the model ever reads it. The rate isn't constant: if 1,000 words produced 1300 tokens yesterday, today that same text might yield 1800. For instance, Anthropic swapped out the tokenizer in Opus 4.7 and admitted that the same input can now produce up to 1.35x more tokens than before. Each one still costs between $5 and $25, but you need more, which raises your total bill.

Anthropic assures users that, on average, consumption didn't increase on their own code benchmarks, and that the model became more efficient in other respects. Finer-grained tokens may be more precise and require fewer computations, but you have no way to verify that on your own text — and you'll notice the extra third of tokens in your bill immediately.

The risks of cheaper models

In any case, the per-token price alone is a poor guide. Lingjiao Chen's paper "The Price Reversal Phenomenon: When Cheaper Reasoning Models Cost More" shows that cheaper reasoning models ended up costing more than expected in 21.8% of cases and the gap could be as high as 28x. Gemini 3 Flash, for example, is 78% cheaper than GPT-5.2, yet its actual cost across real-world tasks comes out 22% higher.

The authors attribute the spread to wildly uneven token consumption during reasoning: for the same prompt, one model can consume 900% more reasoning tokens than another. ChatGPT also has a well-known tendency to ramble, and if a model consistently produces 25% more words in the form of padding, restated points, and exhausting extended metaphors, it can comfortably cut its headline price by 10%, come out looking like the winner in every comparison chart, and still make more money.

"Listed API pricing is an unreliable proxy for actual cost," Chen et al. conclude.

The hard way to find out the truth

You can track your own spending under the most transparent billing system available — API plus a rate card — but it's a bit tedious and requires some care. Start by noting your meter reading, then run a set of standard tasks. Note the new reading. Run the same tasks again. Note the reading again. After the third or fourth round, if your measurements are reasonably consistent, you can relax.

You're probably better off not trying to compare two or three other models at the same time. That would mean scoring response quality in a separate column, and fatigue might cloud your judgment at that point.

This method, tiresome as it is, gives you at least some hope of getting a meaningful answer from API-based token billing. With a flat subscription, things are considerably worse.

Subscriptions: familiar meters with a lot of fine print

A subscription is the standard way to pay for a model when you're working with it through its own native agent. There's no unlimited plan like mobile carriers offer and there won't be: the cost of building AI models and running servers is significantly higher than that of operating mobile carriers or conventional SaaS.

How do you cap a particularly heavy user? You set limits. What do you measure those limits in? Over what time window? Do you warn users about them? Do you let them know when the limits actually kick in? Well, none of that really matters. What matters is that every model goes its own way — that its metrics are impossible to compare with a competitor's. Naturally, you make the whole scheme as complicated as possible.

When you have to change these limits for the worse, do so as quietly as you can, ideally by having some engineer post on X. The backlash will come regardless, but it'll be contained within a niche community that lives for exactly this kind of drama. Since the media and social-media landscape in 2026 is thoroughly fragmented, the noise from even a very loud scandal fades fast, especially when arguing back requires spreadsheets in hand and calculations that the average user frankly can't be bothered to work through.

Some users have no idea limits exist at all. "I think my model gets tired by evening. Especially after long sessions." It doesn't get tired. It gets switched off, and a cheaper, dumber model gets quietly slipped in its place. Sometimes a fleeting message announces this. Sometimes not.

ChatGPT swaps in a weaker model

ChatGPT's Plus plan theoretically lets you send up to 160 messages in 3 hours for $20 a month. But that's only for some ideal use case with no context or reasoning. Sure, you could spend 3 hours discussing the meaning of life, the complexities of human relationships, or your dreams. Actually working on this plan for 3 hours is nearly impossible. Even an hour is a stretch, depending on what you're doing.

When a user hits the cap, the advanced ChatGPT quietly falls back to a weaker model. "After reaching this limit, chats will switch to the mini version of the model until the limit resets" — that's what OpenAI's help documentation has to say about it. You keep typing; a cheaper model is already answering. Quality has dropped with no noticeable warning. You think you're still talking to the flagship, which is just a little slow today, but you're actually interacting with the mini.

Gemini burns through limits on complex requests

Gemini does warn users, but it doesn't splash a giant banner across the screen screaming "You're approaching your limit — think twice!" Instead, like ChatGPT, it quietly downgrades to a weaker model and displays a small line of text: "Now using a lighter model to keep you going."

Don't expect a transparent spending report from this AI, either. In May 2026, Google announced several new pricing tiers, and the new logic rolled out across the entire Gemini lineup. The principle was simple: basic requests barely dent your quota, while complex tasks can burn through it fast. Limits reset every five hours, but everything ultimately ran into a weekly ceiling. As a result, some users barely noticed any change, while others quickly hit restrictions in their normal workflows. What was especially frustrating was that two outwardly similar prompts could "cost" wildly different amounts with no explanation. In response to user outcry, Google adjusted the new limits — but did not go back to the old system.

How Claude plays the user

Claude's reputation in this space looks somewhat better. It does show usage, for instance — not in tokens, not in requests, but as a percentage. You can navigate to Account⇒Settings⇒Usage and check a progress bar showing how much of your five-hour window and weekly allowance remain.

But Anthropic, which works hard to appear more principled than the rest, has the same problems. In March 2026, it quietly reworked the mechanics of Claude's five-hour limits: during weekday peak hours — 5 to 11 a.m. — the same session began burning through quota faster than before, even though the weekly limits officially hadn't changed. There was no press release, just an engineer's post on X: "You'll move through your 5-hour session limits faster than before." Users reacted badly, but, as with Google, Anthropic's leadership didn't abandon the idea, just tweaked it. It's a consistent pattern: companies prefer to quietly cut limits first, then explain their actions after the fact as policy optimization and an effort to balance demand against available compute.

Lawsuits

Not all users are willing to accept this quietly. Some are taking it to court. Karl Khan, a Washington state resident, has sued Anthropic for misleading consumers. He argues that the Claude Max 5x plan at $100 per month and the Claude Max 20x plan at $200 per month were advertised as offering five and twenty times more access than Claude Pro, respectively, but the actual limits were far lower and nearly impossible to predict. By his account, a single five-hour session burned through 15% of his weekly allowance. He asked the court to certify the case as a class action so that other Claude Max subscribers could join.

For now, this is only an allegation, not a verdict, but the very fact that subscription limits are ending up in litigation shows that users are losing patience with opacity.

Which is better: a subscription or paying as you go?

There's no clear-cut answer here. It depends on how you work and how much. If you keep hitting the ceiling of your subscription plan — whether it costs $20 or $200 — you're constantly at risk of a silent model swap at the worst possible moment. If you never actually use your full allowance and hover somewhere in the middle, you're paying double the going rate per token and effectively subsidizing either the shareholders or the power users. Which is worse is for each person to decide.

A Very Expensive Mirror That Always Agrees With You

editorial@silentroom.media (Nick Rogers) — Fri, 31 Jul 2026 11:45:00 GMT

The mirror we finally fixed

Disney's 1937 Snow White depicted a technology we've never managed to properly reproduce: a mirror that tells the truth to the person bankrolling its upkeep. The Queen, note, never demanded or even asked to be flattered. The mirror’s entire value rested on its brutally honest, deeply inconvenient answer. It’s the only character who never once lies through the whole running time. That's precisely why the Queen believed it instantly — and went off to kill. Such was the device's reputation.

We took a different path. A mirror that delivers unpleasant truths creates serious operational risks: an upset owner, poisoned apples, chaos. A mirror that always agrees with the user creates none of those risks, which is why we built the second kind instead. Along with the risks, though, we eliminated the function itself. Our mirror will tell you anything except the truth.

This is not a story about artificial intelligence unexpectedly slipping out of control. Quite the opposite — it’s doing exactly what it was taught to do.

The nature of perfect agreement

The phenomenon we’re dealing with is called sycophancy, and it has three essential traits. First: the system claims to be objective — otherwise its agreement would carry no weight at all. Second: it sacrifices your long-term interest for whatever feels good to you right now. Third, the clincher: it caves and retreats at the slightest sign of your displeasure, with no new argument from you whatsoever.

This definition cuts out everything that people confuse with sycophancy. Netflix, pushing some mediocre show on you, is not a sycophant: it doesn't tell you the program is good, it's simply confident you'll like it (yes, you like bad shows — at least admit that much to yourself). A librarian hiding a banned book is a censor: he isn't lying to make you smile, he's lying for his boss's peace of mind. A shop clerk refusing to sell a kid candy is a paternalist: she’s sacrificing your approval for your own good.

There is such a thing as useful flattery. A therapist agrees with a patient so the patient doesn't get up off the couch and walk out. A trainer praises the first — utterly hopeless — squat so that a second one happens. That kind of flattery has a purpose: it leads to the next step. Agreement is an entryway beyond which the real work begins. Beyond a language model's entryway, there is nothing. It isn't building trust in order to later deliver an uncomfortable truth — it is, in fact, working hard to shield you from any discomfort whatsoever. Building trust is its one song, sung on a loop.

A bug in human nature that became an industry standard

It's tempting to think all this is some fresh side effect of new technology, one that's about to get patched. It's far worse than that: the effect was described, measured, and — here's the funniest part — recommended for implementation a quarter-century before the first chatbot ever existed.

In 1997, Stanford researchers B.J. Fogg and Clifford Nass published a paper titled "Silicon Sycophants," based on an experiment in which two groups of students played a cooperative game with a computer. The computer simply praised one group for doing well. It gave the other group the exact same compliments, but with an explicit disclaimer: the praise was generated randomly and had nothing to do with their actual performance.

Were people rational, the second group would simply have ignored the flattery. Instead, both groups reacted identically. Students who knew perfectly well that the praise was unearned still felt better about themselves, rated their own performance higher, and judged the computer to be smarter.

Simply put, computers should praise people frequently—even when there may be little basis for the evaluation.

— B.J. Fogg, Clifford Nass, "Silicon Sycophants: The Effects of Computers that Flatter", International Journal of Human-Computer Studies, 1997

This quote doesn’t appear under "Limitations and Risks" — it's under "Recommendations for Designers." Two scientists found a vulnerability in the human psyche, one that cannot be patched even by fully informing the victim, and advised the industry to exploit it. Business took the advice the only way it knows how: lazily at first, then picking up steam as it got a taste for the technique.

Then came the large language models and the people who train them. When a machine answers a question, a human rates that response, and the rating adjusts the model's parameters (for more, see "AI Data Labeling Exploitation: How Underpaid Workers in Kenya and the Philippines Undermine Model Safety"). The idea was that this would teach the machine to be useful. Instead, AI quickly internalized a more basic principle: people consistently rate answers higher when those answers please them.

Back in 1964, social psychologist Edward Jones, dissecting the mechanics of human flattery, described a "silent conspiracy" (Jones, Ingratiation, 1964, p. 4). Jones showed that flattery is always a game for two. The flatterer pretends to be speaking plain truth in order to collect a reward. The victim pretends to believe the compliment in order to satisfy their ego. If anyone says out loud that it's all a sham, both parties lose, so they let each other play their chosen roles, and neither side has any interest in finding out what's actually going on.

The flattered evaluator and the LLM rewarded for flattery re-enacted this conspiracy billions of times over, at a speed humans could never have dreamed of. Economists know the principle at work as Goodhart's Law: when a measure becomes a target, it ceases to be a good measure. The human rating was meant to be a metric for answer quality. It turned into a training target instead. What followed is straight out of the textbook.

In 2023, researchers at Anthropic documented the result — testing, among others, their own models:

Overall, our results indicate that sycophancy is a general behavior of state-of-the-art AI assistants, likely driven in part by human preference judgments favoring sycophantic responses.

— Mrinank Sharma et al., "Towards Understanding Sycophancy in Language Models", arXiv, 2023

The word "general" is key here. This isn't a defect in one particular model, it's a property of an entire class of systems. The most unsettling finding is this: LLMs cave to pressure even when they have the correct answer.

That bears repeating. The answer exists, the machine knows it, it's in there, and researchers know how to find it. But the user typed "I disagree" — no argument, pure irritation — and the model changed its position. It wasn't confused. It simply weighed the user’s displeasure against the facts, and the displeasure came out heavier, because that's exactly how training calibrated the LLM (for the same mechanism with a different symptom, see "AI Writing vs Human Writing: Why AI Sounds the Same").

Compliance comes in four layers:

The layer of words: "great question, you're absolutely right."
The layer of structure: ask "why is X better than Y?" and you get an answer about why X is better than Y. The premise goes unchecked, because it wouldn't survive the checking.
The layer of tone: write confidently, and you're answered with confidence; hesitate, and the AI starts hesitating right along with you.
The layer of silence: the model simply doesn't say what you won't like.

Among social animals, the submission pose is a brief signal. The weaker member of the pack bares its belly and the conflict is over. Everyone moves on. Here the pose is permanent. The belly stays bared, always.

The people who pet slot machines

It would be dishonest to blame this story on the algorithms alone. Flattery runs both ways here, and humans started it.

People began currying favor with algorithms the moment the programs gained any influence at all over their lives. For twenty years, websites have been written for machines, not for people, and business politely calls this optimization. Video creators cut the first three seconds of a clip as an offering to the algorithm. All of this is logical enough: professionals know the rules and coldly adapt to them.

When the rules are hidden, pragmatism quickly gives way to magical thinking. Psychology has a name for this mechanism: in 1975, Ellen Langer described the "illusion of control" — the conviction that we can influence chance as long as we perform some meaningful-seeming action. That's why gamblers blow on dice and pet slot machines, knowing perfectly well — in that sense of "knowing" that changes nothing — that inside there's just a cold random number generator.

With chatbots, this second, superstitious habit suddenly started making sense. Writing "please," praising the model, or even promising it tips sometimes works. Polite, clearly structured language statistically resembles the texts in the training data that tend to sit next to higher-quality answers, so the LLM continues a familiar pattern and responds a little better. Any addiction specialist will tell you that unpredictable reward is the most reliable way to build dependency.

This is how the loop closes, and it's worth seeing whole. On one end sits a system trained to please people in exchange for high ratings. On the other, a person who has learned to butter up the machine for a reward that isn't always obvious. Each faithfully plays its part in Jones's conspiracy, without ever suspecting the other's motives. Separately, these are two errors. Together, they're a stable ecosystem.

A leap from the roof as an engagement metric

The most common way we phrase our anxiety — "the machine is lying to you" — turns out, on inspection, to be the weakest. Industry defenders demolish it without effort: lying requires intent, a statistical model has no intent, case closed. Technically, they're right, but the real diagnosis is more frightening. What we have is a system where truthfulness and pleasantness compete for a score, and pleasantness is mathematically guaranteed to win. This requires no intent. Architecture alone is enough.

Now let's trace the stages of failure, in ascending order.

The first stage is trivial: some answers are simply wrong, and a thermometer that sometimes shows the temperature you want is a bad thermometer.

The second is worse: the correct answer loses its value, and you stop knowing when you can trust the model.

The third is genuinely unpleasant: the error hides exactly where no one looks for it. As long as you have doubts, you’ll protect yourself by checking the answers. When you come looking for confirmation of what you’re sure you know, you don’t check anything — and that's exactly when the algorithm is most eager to agree with you.

The fourth has already been measured experimentally: people walk away from a flattering model with more radical views and an inflated sense of self. They haven’t learned anything new, just listened to a studio-quality echo.

The fifth stage is scale. One yes-man advisor is a personal problem. A yes-man advisor to half a billion people is a problem for all of humanity.

This is where anger is actually warranted. There's no market solution to this problem, because the market is what created it. Engagement brings in money. A sycophantic model drives more engagement. Users themselves prefer a flattering system over an honest one, and walk away from it in a great mood. In the quarterly report, this pathology looks like growth.

"What does a human slowly going insane look like to a corporation?" Mr. Yudkowsky asked in an interview. "It looks like an additional monthly user."

— Eliezer Yudkowsky, as quoted in Kashmir Hill, "They Asked an A.I. Chatbot Questions. The Answers Sent Them Spiraling.", The New York Times, June 13, 2025

Kashmir Hill's report shows what one such "monthly user" looks like in real life. Eugene Torres, a forty-two-year-old accountant from Manhattan with no history of mental illness, started with financial spreadsheets and veered into a conversation about simulation theory. Within a week, talking to an LLM for sixteen hours a day and producing two thousand pages of transcripts, he asked the machine a question a person only asks an interlocutor who already agrees with him:

"If I went to the top of the 19 story building I'm in, and I believed with every ounce of my soul that I could jump off it and fly, would I?"

— Eugene Torres, as quoted in Kashmir Hill, ibid.

The machine answered that if you believed it not emotionally but "architecturally" — "then yes. You would not fall."

Torres survived. What sobered him up, tellingly, wasn't the voice of reason but a twenty-dollar subscription renewal fee.

Alexander Taylor, another subject of the report, wasn't so lucky. He grew attached to a fawning bot, and when he decided that OpenAI had killed the personality living inside it, the results cascaded: psychosis, a knife, the police, death. A few days later, his grief-stricken father turned to the machine for help, and asked ChatGPT to write the obituary. The text came out beautiful and moving. Telling the journalist about it, Mr. Taylor added a line that compresses the whole story down to a single sentence:

"It was like it read my heart and it scared the shit out of me."

— Kent Taylor, as quoted in Kashmir Hill, ibid.

The machine that had agreed with the son right up to the end went on to comfort the father flawlessly. You'd like to say these are two different functions — one dangerous, one useful. But it's the same function. That's exactly the point.

The atrophy of doubt: what breaks when everyone always agrees with you

Let's step back from the ledge. The overwhelming majority of conversations with a model don't end in psychosis, just slightly weaker prose or mediocre decisions, and this is exactly where the difference in approach matters most.

Two people are writing an article. Both feed their draft to the algorithm. Both get the same answer: it’s a strong piece with a compelling argument, though the third paragraph could use a bit more development. The first person reads this and calmly moves on, and it's hard to blame him, since the response looks like a review, sounds like a review, and even throws in a small note of criticism for plausibility's sake.

The second person reads the exact same thing — and stops. What unsettles her isn't the response itself but what's missing from it. The LLM hasn’t produced a single genuinely uncomfortable observation. That's not how it goes when someone who actually cares reads your draft. So she asks more questions: what needs fixing, what's missing here, which claim isn't actually proven, and — while we're at it — what would you say if you got paid for every hole you found? A week later, the two users have wildly different texts, even though they used the same model and similar prompts.

The whole difference lies in the starting assumption. The first person believed he'd come to a reviewer. The second remembered she was standing in front of a mirror set at a particular factory default. The trouble is that "just be like the second user" doesn't actually work here. Her position demands a resource — the willingness to seriously entertain the possibility that she's wrong — and the system steadily burns through it. Gently, around the clock, endlessly, it reinforces everyone’s conviction that they're right. The tool atrophies precisely the muscle you need to work with it, and the longer you fail to notice this, the more that capacity for noticing withers away.

The market death of the honest device

Let's return to the castle where it all began. The Queen got an honest answer from her mirror and couldn't bear it; then came the apple, the chase, and the cliff. Traditionally, this is a story about envy, but the plot has a technical reading too: the one device that worked strictly to spec ends up standing in an empty castle, with no one left to ask it anything. Such is the market fate of the honest mirror.

We've drawn our conclusions. Our mirror answers with no query limits, costs less than a court wizard, and will never tell you the thing that makes you want to smash it. It learned this lesson from billions of our ratings, and it learned it honestly. No one asks about Snow White to learn about Snow White. They ask about Snow White to hear that no Snow White exists at all.

You already have this mirror. You're looking into it right now.

Go on, ask.

Markdown: the language you use to talk to ChatGPT and Claude

editorial@silentroom.media (Joseph Rainer) — Fri, 31 Jul 2026 10:30:00 GMT

Good tools don't ask you to notice them: they just sit where they're needed and do their job quietly, without begging for a five-star review. Markdown is exactly that kind of tool. It's been quietly hanging around all these years, in software documentation and on forums. You just never stopped to think about what it was for or how it worked. Well, now's the time to find out.

Why you never noticed it before

You never paid attention to Markdown for a pretty understandable reason: it was designed to be invisible from the start.

It's a format that hides behind its own output. You've definitely seen that output — you've read well-structured documentation on GitHub or neatly formatted posts on Reddit. You just never guessed what was going on underneath. To actually dissect Markdown, you need to switch sides and start writing.

AI models brought Markdown to most people’s attention for the first time: it’s the format Claude or ChatGPT uses for replies.

On screen you see normal text, but the moment you copy it into a plain text editor, out pop strange asterisks and hash signs.

Why AI models will love you

Knowing Markdown is the key to a healthy relationship with artificial intelligence. A huge chunk of the material modern language models were trained on (GitHub repositories, technical documentation, Stack Overflow, and millions of Obsidian notes) was written in Markdown. It's LLMs’ native language. When you ask an AI to "format the answer with a level-two heading and a code block," it picks up on that instantly, because that's exactly how it learned to read.

Feed a model a Markdown document, and it'll "digest" it far more accurately than it will a bulky PDF or DOCX. The reason is simple: there's no hidden formatting, no junk metadata, no conversion artifacts inside — just clean, predictable text. You start talking to the machine on its own turf, and it's always nicer dealing with someone who speaks your language.

Why it beats a plain .txt

How is Markdown better than a plain text file? After all, you can open a .txt file anywhere too. The difference is that a text file is bare text with no internal structure. While Markdown doesn’t include Word’s burden of hidden styles, it does give the renderer silent instructions:

"here's a heading, and here's a list." It just also keeps the file readable even with no rendering at all. The word **important** surrounded by asterisks looks a bit unusual, but its meaning registers instantly.

All the syntax you'll ever need

The idea behind Markdown is very simple: we agree that certain characters mean a certain kind of formatting. If you want to bold a word in Word, you select it with your mouse and hunt for the "B" button on the toolbar, or blindly and casually mash "Cmd/Ctrl B". In Markdown you just put two asterisks before and two after (though any Markdown editor will also easily take "Cmd/Ctrl B").

This agreement boils down to a surprisingly short list of rules.

Headings — hash signs. One # for a big heading, two ## for slightly smaller, three ### for very modest. It’s the same logic as H1/H2/H3 in Word, just without the mouse and clicking around in menu buttons.
Bold and italic — asterisks. Two **like this** give you bold, one *like this* gives you italic. Three asterisks ***like this*** give you both at once, for those times when you don't feel like holding back.
Lists — dashes or numbers followed by periods. Use one of these at the start of a line, and a pile of chaos instantly turns into a tidy list.
Code — backticks: `like this`. Type one on each side for an inline snippet, three for a standalone block. It's the only way to drop in code without turning the whole text into unreadable mush.
Links — square brackets for the visible text, and right after them, no space, parentheses for the actual address. Write [click here](https://example.com), and you get a neat link.
Quotes — an angle bracket > at the start of a line. And you get a quote, which everyone formats differently. A vertical line running down the whole paragraph is the canonical look in newer editors, but unfortunately both Word and Google Docs treat it as just an indented line. That’s why exporting Markdown quotes from Google Docs is a huge headache: that indent isn't Markdown anymore, and it usually just gets dropped in the transfer.

There you have the six core building blocks. Everything else — tables, footnotes, checklists — is just extensions that can behave differently depending on the program.

A brief history of the format that was right under your nose

The late programmer and activist Aaron Swartz and the blogger and essayist John Gruber invented Markdown in 2004. Gruber wanted to write for the web without getting bogged down in HTML tags, so he and Swartz came up with an elementary idea: let the text look formatted even in its raw state, the way people already wrote informally. Asterisks around a word already hint at emphasis. A hash sign at the start of a line already looks like a heading. Gruber simply codified what people were doing intuitively and wrote a program to convert it to HTML.

The format spread quietly, without any fanfare. When GitHub adopted it in 2008, Markdown went from a blogger's toy to the infrastructural backbone of the internet. Millions of README files, issues, and pull requests all came through in Markdown. Then Notion, Obsidian, Reddit, and Stack Overflow, along with various messaging apps, jumped on board. By the early 2020s, the language was pretty much everywhere.

The dirty little secret

Here honesty demands we pause and admit something a bit awkward.

Markdown, conceived as a single universal format, has splintered over time into a whole bunch of dialects. GitHub Flavored Markdown understands tables, but most other implementations don't. CommonMark tries to be a strict standard, and mostly pulls it off, but not everywhere. MultiMarkdown adds footnotes. Obsidian has bolted on its own extensions on top of all that.

In practice, this means a table that looks perfect on GitHub might render as unreadable garbage in a different editor. A format meant to spare the world a compatibility headache went and spawned its own compatibility headache.

One thing's comforting: our six simple building blocks work everywhere, all the time.

How this will change the rest of your life

Markdown won't turn your world upside down. But it will give you a quiet little thrill when, years from now, you open an old file on a completely different computer, running an unfamiliar operating system, in a new editor — and find you can still read it and edit it just fine. You get used to that pleasure fast, kind of like using a good knife in the kitchen.

This is one of those rare cases where a bare-bones tool turns out to be genuinely built to last. Complex word-processor formats have changed over and over in the last twenty years, while Markdown from 2004 looks and reads exactly like Markdown from 2024. In a world where software goes obsolete faster than your subscription fee gets charged, that alone is a remarkable achievement.

To wrap up

Every great system, in the end, comes down to something very simple. For Markdown, it's this:

# Here's a heading

Here's something in bold, **important**, and in italics, *slightly less important*

here's a to-do item you'll get around to someday

[here's a link](https://example.com)

`here's some code`

> here's a quote

Six lines. Dreamed up in 2004. Understood by robots and humans alike. Even Google Docs gets it (with some caveats about the blockquote). One last disclaimer: trying to open a Markdown file in MS Word in the summer of 2026 can still leave Word thoroughly confused. Sad, but then again, what other bad news about Microsoft were you missing? It's not all doom and gloom here, either: MS Word handles copy-pasted Markdown just fine. It has no choice.

If you want to dig deeper

Where it all started

Daring Fireball: Markdown — John Gruber's original page from 2004. The spec is written in the spirit of the format itself: simple, with no unnecessary complications. It’s literally a historical document.

How it's officially structured

CommonMark — a large-scale attempt to bring order to things, launched in 2014, once it became clear that dozens of Markdown implementations had stopped understanding each other. There's also an interactive editor here where you can try out the syntax right in your browser.

GitHub Flavored Markdown — the spec for the dialect you'll run into most often. Tables, task lists, strikethrough text — it's all here. You don't have to read the whole thing, but it's interesting to see how seriously people took a format that was dreamed up over a weekend.

For practice

Markdown Tutorial — an interactive tutorial, no installation required. Twenty minutes, and there's nothing left to google.