Are AI Detectors Accurate? How They Get It Wrong and Who Pays

In May 2026, a journalist named Steven Rosenbaum published a book arguing that AI is poisoning the truth. Within a week, The New York Times reported that the book itself contained more than half a dozen quotes fabricated by AI — words that real people like Kara Swisher never said. The man warning you not to trust the machine had quietly trusted the machine. It is a perfect little parable, and it has nothing to do with the people this article is actually about.

Because while the Rosenbaum story was front-page entertainment, a quieter scene was playing out in classrooms and editorial inboxes: real writers, who used no AI at all, getting flagged as cheats by software that claims to sniff out machine text. So let's answer the question plainly. Are AI detectors accurate? No — not reliably, not at the level of confidence schools and employers are betting careers on. This piece is for two people: the one worried they'll be flagged for writing too cleanly, and the one already staring at a "94% AI" score next to an essay they bled over. Both of you are right to be nervous. Here's how these tools work, why they fail, who pays for it, and what you can actually do about it.

How do AI detectors actually work?

Before you can judge whether a detector is right, you have to understand what it's measuring — and it is almost never measuring what you think. (If you searched how do AI checkers work, you're in the same place — "checker" and "detector" are the same machine wearing two name tags.) An AI checker does not "read" your essay and recognize ChatGPT's fingerprints. It has no idea who or what wrote your text. It measures statistical properties of the words and then guesses. That's the whole trick. Understanding the two signals it leans on — perplexity and burstiness — is the fastest way to see why false positives aren't rare glitches. They're baked in. No machine-learning background required: if you've ever had your phone suggest the next word in a text message, you already understand the core idea.

What perplexity measures

Perplexity is a measure of how surprised a language model is by your next word. Feed a model the sentence "the cat sat on the…" and it expects "mat." If your next word is "mat," the model is unsurprised — low perplexity. If your next word is "trampoline," the model didn't see it coming — high perplexity.

Here's the leap detectors make: AI models are built to produce the statistically likely next word, so their writing tends to be low perplexity — smooth, predictable, the verbal equivalent of a four-lane highway. Human writing, the theory goes, is bumpier and more surprising, so it runs high perplexity.

You can probably already feel the problem. Clear, plain, well-edited writing is also low perplexity — because good editing is largely the act of removing surprise. The whole point of a clean sentence is that the reader doesn't trip. So the better you get at writing simply, the more your prose looks, statistically, like a machine's. This is where the false positive problem is born, and it never really leaves.

What burstiness measures

The second signal is burstiness. Burstiness measures variation — how much your sentence length and complexity jump around within a piece. Humans, the theory says, write in bursts. A long, winding, comma-stuffed sentence that explores an idea from three angles. Then a short one. Then a fragment. The rhythm lurches.

AI, by contrast, tends to produce an even, metronomic flow — sentence after sentence of roughly the same length and shape, like a hedge trimmed flat. Low burstiness, the detector decides, means machine.

Except: plenty of humans write with low burstiness naturally. Technical writers are trained to keep sentences uniform and scannable. People writing in a second language often stick to safe, consistent structures. Writers with certain cognitive profiles favor steady, even rhythm. None of them touched an AI, and all of them just tripped the wire.

How perplexity and burstiness combine into a detection score

So how do ai detectors work — perplexity, burstiness, and all? The detector runs your text, scores it on both signals, blends them into a single number, and compares that number to a threshold. Cross the line and you get the dreaded percentage. Below it and you pass. That's it. There is no understanding in the loop, no fact-checking, no knowledge of your draft history — just two proxies and a cutoff.

What the tool measures	What it assumes is "AI"	Who that sweeps up by accident
Perplexity (predictability)	Low — smooth, expected word choices	Clear, heavily edited, simple writing
Burstiness (sentence variation)	Low — uniform sentence rhythm	Technical writers, ESL writers, formal stylists
Composite score vs. threshold	Above the cutoff = flagged	Anyone whose normal voice sits near the line

And here's the part nobody advertises: neither perplexity nor burstiness was invented to catch AI. They're borrowed measurements from older natural-language-processing research, repurposed for a job they were never designed to do — two rulers built to measure one thing, now used to convict people of another. That it works at all is mildly impressive. That we treat its output as evidence is the scandal.

Are AI detectors accurate — and can they be wrong?

The honest answer, the one the marketing pages bury: yes — can AI detectors be wrong? Constantly, and often enough to ruin specific people's lives. How accurate are AI detectors in the real world? Good enough to feel authoritative, bad enough to be dangerous — which is the worst possible combination. This isn't a fringe complaint from people trying to get away with cheating. It's documented in peer-reviewed research and, tellingly, admitted by the companies that build these tools.

What the Stanford research found

In 2023, a team of Stanford researchers led by Weixin Liang ran a now-famous test. They took essays written by non-native English speakers — real human writing, TOEFL exam essays — and fed them to seven commercial AI detectors. The detectors flagged more than 61% of them as machine-generated. On essays written by native English speakers, the same tools were nearly flawless.

Read that again. The software wasn't detecting AI. It was detecting non-native English, and calling it AI. The study, published in the journal Patterns, connects straight back to perplexity and burstiness: writing in a second language tends toward simpler vocabulary and steadier structure — exactly the low-perplexity, low-burstiness signature detectors are tuned to punish. The researchers' own suggested "fix" was darkly funny: tell students to rewrite using fancier vocabulary to fool the machine. Write less like yourself to prove you're human.

Why OpenAI shut down its own detector

Here is the credibility anchor that should end most arguments. In January 2023, OpenAI — the company that built ChatGPT — released its own AI text classifier. Six months later, in July 2023, it quietly shut the tool down, citing its "low rate of accuracy."

Sit with that. The organization with the most intimate knowledge of how AI text is generated concluded that detecting it didn't work well enough to keep online. They had every incentive to crack it. They couldn't. And yet schools, newsrooms, and HR departments keep paying third-party vendors who insist they've solved the very problem OpenAI walked away from. If the people who built the engine can't reliably hear it running, why trust the strangers selling stethoscopes?

GPTZero, Turnitin, and Originality.ai: what the data shows

Let's get specific, because "detectors are bad" is a vibe and you came here for verdicts. Below is the same pattern repeating across the three names you'll actually run into: the company quotes a tiny false-positive rate from its own lab, and the real world quotes a bigger one.

Detector	What the company claims	What independent testing & institutions found
GPTZero	Very low false-positive rate; built for educators	Independent tests put real-world false positives meaningfully higher and overall accuracy in the 60–90% range depending on text length; short texts suffer most
Turnitin	Under 1% false positives on documents over 20% AI, validated on ~700K pre-ChatGPT papers	Vanderbilt disabled it — even 1% means ~750 wrongly flagged papers per 75,000 submitted
Originality.ai	~99% accuracy, 0.5–1.5% false positives	Independent tests put false positives several times higher; the most aggressive detector by design

Is GPTZero accurate? It's one of the more careful tools, and to its credit GPTZero itself warns against using a single score to accuse a student. But "more careful than the others" is not the same as "accurate enough to expel someone." Its own documentation hedges, which tells you something.

Turnitin's accuracy claim is the most revealing, because the company is honest in the fine print: the headline Turnitin AI detection accuracy number — that sub-1% false-positive figure — applies at the document level, on documents that are already more than 20% AI, validated against a clean set of papers written before ChatGPT existed. That's a lab condition, not your Tuesday-night essay. Vanderbilt did the math on what even a "tiny" 1% rate means at the scale of a real university — roughly 750 innocent papers a year — and turned the feature off rather than wear that.

Then there's Originality.ai, which deserves credit for not pretending. It is openly the most aggressive detector, built for marketers and publishers who'd rather wrongly flag ten human articles than let one AI article slip through. Defensible for a content farm checking freelancers. Catastrophic for grading a 19-year-old. The tool isn't broken; it's optimized for a goal that has nothing to do with fairness to the writer. The throughline across all three: the marketing number and the real number are different numbers, and you are the rounding error.

Do AI detectors actually work, or just flag resemblance?

Here's the reframe the whole article hinges on. Do AI detectors actually work? It depends what you think "work" means.

Are AI detectors reliable? Ask it the harder way — how reliable are AI detectors when a real career is on the line — and the answer collapses. Are AI checkers accurate enough to accuse someone? Not even close. They do not detect AI. They detect resemblance to AI. That's not a pedantic distinction — it's the entire ballgame. These tools look for the statistical fingerprints that machine text tends to leave. So a human who naturally writes with smooth, even, predictable prose gets flagged, because they resemble the machine. Meanwhile a cheater who prompts the AI to "write messily, vary your sentences, add some quirks" sails through, because the machine has been told to not resemble itself.

Flip it around and the absurdity is total: the detector punishes the human for writing like a machine and rewards the machine for writing like a human. It is a bouncer who checks how you're dressed, not whether you're on the list. Once you see detection as a resemblance test rather than a truth test, every false positive in the next section stops looking like a bug and starts looking like the design working exactly as built.

Who gets falsely accused of using AI?

If a detector is really just measuring "how far does this writing sit from average native-English prose," then the people who get hurt aren't random. They're predictable. Anyone who deviates from that statistical center for a perfectly legitimate reason is structurally at risk — and "deviates for a legitimate reason" describes a huge slice of actual writers. This is a systemic bias, not an occasional miscalibration. Here's who keeps walking into it.

What tools schools are actually using

So what do professors use to detect AI? The short list is small and you already know the names. Turnitin's AI-detection layer is the giant, because it's bolted onto plagiarism-checking software that thousands of institutions already license. After that: GPTZero, Copyleaks, and Originality.ai round out what most schools and instructors reach for.

Tool	Where it shows up	How it's typically used
Turnitin (AI layer)	Universities, large school districts	Auto-runs on every submission via the existing plagiarism suite
GPTZero	Individual instructors, smaller schools	Manual paste-in checks, often ad hoc
Copyleaks	Mixed academic and enterprise	First-pass screening
Originality.ai	Publishers, some instructors	Aggressive flagging, marketing origins

The deployment is where it gets ugly. These tools are often used as a first-pass filter by instructors who have never seen the false-positive research and assume a percentage is a fact. Frequently students aren't even told detection is running. You can be accused on the basis of a number whose error rate your accuser doesn't know and you were never warned existed. It's why students falsely accused of using AI so often describe the same bewildered question — why is my writing being flagged as AI? — with no one in the room able to answer it. That's the institutional backdrop for everything below.

Non-native English speakers

This is the best-documented group, and we've already met the evidence: the Stanford study's 61% false-flag rate on non-native essays. The mechanism is brutally simple. Writing in a second language tends to lean on a smaller, safer vocabulary and more consistent sentence structures — which is exactly the low-perplexity, low-burstiness signature detectors read as "machine."

So an international student who worked twice as hard to write a clean English essay gets penalized because it came out clean. The reward for mastering the language is suspicion. There's a real discrimination dimension here that institutions are slow to name: a tool that disproportionately flags one nationality of student would never survive scrutiny if a human were making the same calls. Dressed up as software, it runs unchallenged.

Students with autism, ADHD, and dyslexia

The research here is thinner than the ESL literature, so treat this as an emerging concern backed by structural logic and a growing pile of anecdotes rather than a finished study. But the logic is hard to dodge.

Autistic writing is often precise, literal, and consistent in its terminology — low on idiomatic variation, high on exactly the kind of patterned regularity a detector reads as synthetic. Dyslexic writers frequently edit heavily for clarity and lean on shorter, cleaner sentences, flattening burstiness. ADHD complicates the picture from the other direction: those writers may produce high variation, but they also lean heavily on AI editing tools for structure, blurring the line further.

The reported cases are grim in a specific way — one student described deliberately worsening their own writing after repeated flags, just to dodge an accusation. When your accommodation strategy is "write worse so the robot believes I'm real," the tool has stopped measuring honesty and started punishing disability.

Academics and literary writers with strong style

Now the irony that should sting your masthead's core readers. The writers most likely to be flagged are often the best ones. Develop a controlled, disciplined, consistent voice — the thing every serious writer spends years building — and you've also built a low-perplexity, low-burstiness profile that lights up the detector like a slot machine.

The entire classical toolkit works against you. Parallel structure lowers perplexity. Anaphora — "And he saw. And he understood. And he left." — reads to a machine like assembly-line output. The rhetorical repetition that runs from Cicero to Brodsky is, statistically, indistinguishable from the froth a model produces on autopilot. A polished academic abstract, a minimalist literary paragraph, a tightly argued legal brief: all of it sits in exactly the zone the detector distrusts.

So the same craft that earns a writer a byline earns them a 90% score. The tool cannot tell style from stylization — it cannot distinguish a human who writes with deliberate control from a machine imitating one. And it's been handed the authority to judge anyway.

What to do if you're falsely accused of using AI

Enough diagnosis. If you're reading this because a professor accused you of using AI — or because you typed "I didn't use AI but it says I did" into a search bar at midnight — here's the practical part. Every AI detector false positive feels personal, because it is: it's your name attached to a number you can't argue with. The tone shifts here from analysis to advice, and the advice is direct because you don't have time for "it depends." Here is how to prove you didn't use AI — or at least make the accusation fall apart.

Build your paper trail before you submit

The single best defense is one you build before anyone questions you. Version history is the strongest evidence of a genuine writing process that exists, because it's almost impossible to fake convincingly after the fact.

Write in something that keeps a timestamped history of every change:

• Google Docs — its version history records the document's whole evolution. Tools like Draftback can even replay the entire writing session like a film, every keystroke and deletion.

• Microsoft Word with track changes or saved iterations (draft_v1, draft_v2, and so on).

• Notion or Dropbox Paper — any editor that logs revisions with timestamps.

Then keep the surrounding debris of real work: research notes, annotated sources, browser history, even voice memos or freewriting that predate the final draft. Build this habit before you need it, not after — once you're accused, anything you create looks like a cover story. A messy, time-stamped trail from blank page to final draft is a defense that "trust me" can never be.

What evidence actually works in an appeal

Not all evidence is equal, and knowing the difference matters when you're sitting across from an academic-integrity panel. The most important thing to understand: a detector score is a flag, not a finding. It is not legally or academically conclusive. It is an accusation generated by software with a known error rate, and you are allowed to treat it that way.

Strong evidence	Weak evidence
Document version history showing gradual drafting	Your verbal "I promise I wrote it" alone
Contemporaneous research notes and annotated sources	A screenshot of a different detector calling it human
Browser history from research sessions	Character references unrelated to the document
Earlier dated drafts, outlines, freewriting	Re-running the text and getting a lower score

One move that feels clever and mostly backfires: running your flagged text through a different AI checker hoping it clears you. Detectors disagree with each other constantly, so a second tool calling you human just proves the tools are unreliable — which helps your broader argument but won't, on its own, clear you. Lead with process evidence. Use the unreliability of detectors to undercut the accusation, not to mount your whole defense.

The trap of rewriting to look human

Here's the instinct to resist. You see "78% AI," you panic, and you start hacking your own prose to pieces — adding typos, chopping rhythm, jamming in weird word choices to "look more human." Don't.

Two reasons. First, it doesn't reliably work — these tools are noisy, and roughing up your sentences may not move the score at all while definitely making your writing worse. Second, and more dangerous: if your original submission is ever compared against your "fixed" version, the rewrite looks exactly like what guilty people do. You've manufactured the appearance of a cover-up to defend yourself against a false charge. Never alter the flagged document. The document is your evidence. Your job isn't to make the text look human to a machine — it's to prove the human who already wrote it. Build the case that the original is yours. Don't deface the original trying to flatter the calculator.

From plagiarism to authenticity: what this means for writers

Step back and you can see the ground shifting under every writer at once. Ten years ago, disputes about authorship were about plagiarism: did you copy this from someone else? The question was comparative, and the burden sat with the accuser to find the source you stole from. Today the dispute is about authenticity: did you write this at all, or did a machine? And the burden has quietly flipped onto you.

That flip is the whole story. *You are now presumed guilty by default, asked to prove a negative — that you didn't use AI — using an instrument you can't inspect and can't cross-examine.* The methodology is opaque, the appeal depends on the mood of one instructor or editor, and the math underneath is two repurposed NLP rulers and a threshold. Kafka would have recognized the architecture immediately. He'd also have failed the detector cold — those long, unbroken sentences, that steady perplexity — and spent his life under suspicion.

There's a better path, and the institutions thinking clearly are already on it. Groups like MIT Sloan's teaching specialists argue that the answer isn't a better calculator — it's process-based assessment: judging the draft history, the oral defense, the visible road from idea to finished page, rather than outsourcing judgment to a percentage. That's not just fairer. It measures the thing detection can't even see — the actual work.

Because here's what genuinely separates a living writer from a machine, and it's the same thing that's always separated a good writer from a bad one: the plethora of drafts, the crossed-out chapters, the endless hunt for the right word, the long and difficult road to a finished text. The evidence of craft is the evidence of humanity. Save your version history. Keep your drafts. Hold onto the messy trail of how the thing got made.

A few things to carry out the door:

1. AI detectors measure resemblance, not authorship — they can't tell style from stylization, and they were never built to.

2. False positives aren't random — they fall hardest on non-native speakers, neurodivergent writers, and stylists with a strong voice.

3. The companies and even OpenAI admit the limits — a flagged score is an accusation, not a verdict.

4. Your process is your defense — version history and dated drafts beat any score, in either direction.

5. Don't write worse to look human. Write well, and keep the receipts.

The burden of proof has landed on writers, and the tools doing the shifting aren't strong enough to carry the weight they've been handed. Until that changes, the smartest thing you can do is the oldest thing in the craft: do the work, and keep proof that you did. Keep your head up. They're not locking anyone up for anaphora — not yet.