The End of the Link Era

Part 1. The Charges

In November 2024, Judge Colleen McMahon — who has seen enough aggrieved-publisher lawsuits to fill a monograph — dismissed the case brought by Raw Story and AlterNet against OpenAI. The outlets presented the classic grievance: their articles had been used to train ChatGPT without permission, without payment, without so much as a formal nod toward copyright. McMahon responded in a tone lawyers politely call withering and everyone else calls contemptuous. The substance of her ruling came down to a simple question: where, exactly, is the harm? Show me a concrete injury — lost traffic, a missed subscription, a reader who walked off to the robot. The plaintiffs couldn't. Case closed.

This was bad news for Raw Story, but surprisingly good news for people who want to understand what's actually happening with AI and journalism. A single court hearing exposed what the media industry keeps hiding from itself, and even the victims don't quite grasp where or how they were robbed.

This is the point where the big numbers come in. According to fresh measurements from Status Labs, the factual accuracy of SearchGPT — the product OpenAI is selling as a search engine replacement — sits at around 76%, compared to 98% for Google. Some 23% of its claims are unsupported by citations. A single SearchGPT response cites an average of 3.4 sources; Google's first page of results cites 8.2.

Almost every media critic will present these figures as a death sentence: look, AI search lies, AI search steals, AI search is strangling the primary source. But if you resist the hysteria and look at the numbers with a clear head, something else comes into focus.

Google and SearchGPT are different products solving different problems. Google is a library catalog: it tells you where things are shelved and leaves the rest to you. SearchGPT is the well-read neighbor who's consumed pretty much everything and holds forth about it at the kitchen table. That neighbor sometimes gets names wrong and forgets where they heard what, but the catalog occasionally sends you to a third-floor reading room that's been closed for three years. These are different genres of error, and conflating them is like faulting a pedestrian for being slower than a bicycle.

Then the situation gets genuinely entertaining, and I'll admit I've been waiting a long time to say this. The SEO industry — the very one that spent twenty years teaching newsrooms to write headlines for the algorithm, stuff copy with keywords, and churn out "10 Best Ways to Do Anything" — is now crying wolf the loudest. It's outraged that the new algorithm reads its content too well. A machine trained on texts written by humans to please another machine now summarizes them for a third machine so efficiently that the first machine is left out of the loop. This is not an epochal tragedy. It's the occupational trauma of a narrow professional class that has suddenly discovered its tool is obsolete. Blacksmiths in the early twentieth century felt much the same way. They, too, were convinced the end of the world was at hand.

No end of the world came, of course. There were just fewer horses.

Before we get into who's actually losing here and who's crying wolf the loudest, one thing needs to be established up front. My skepticism runs in both directions simultaneously. I don't buy the apocalyptic forecasts: journalism has never stopped burying itself ahead of schedule. But I don't buy the new-era evangelists either, the ones who show you SearchGPT and tell you that information will now flow to readers more cleanly, quickly, and fairly. It will simply flow differently, through different intermediaries. Who those intermediaries are, and how the new ones differ from the old — that's what we'll get into now.

Part 2. Who's Actually Losing

If you scrolled back through panel discussions on AI search from the past eighteen months, you'd get the impression that virtually the whole of written civilization, from village stringers to Pulitzer Prize laureates, is under threat.

SearchGPT does cut into organic traffic, and it does so aggressively: according to analytics data published in 2024, click-throughs from AI-powered search to news sites run consistently lower than from classic Google, and the share of zero-click answers — responses after which the user goes nowhere at all — is growing faster than anyone previously projected.

When you look more closely at whose traffic is actually dying, the tragedy evaporates pretty quickly. What's getting cut isn't journalism. It's just the lowest floor of it — the listicles, the endless "10 Best Running Shoes for People with Flat Feet," the SEO farms churned out in Chandigarh for $12 per thousand characters, and everything of that ilk. The paradox is almost elegant: AI-powered search is killing content that other AIs wrote to please a third AI. This isn't civilizational catastrophe. It's weeding the garden.

Journalism — in the narrow sense of reporting, investigation, or analysis written by someone who at least once picked up the phone and got their source on the line — is beyond an LLM’s reach. The machine has no primary source to paraphrase, and the moment a user asks "what's the latest in the such-and-such prosecutor case?" SearchGPT can only cite a journalist's latest article about the proceedings. Whether it comes back with a quote is a separate conversation.

The real loser isn't The New York Times: it'll sign a licensing deal, just as News Corp, Axel Springer, and Vox Media already have; it'll highlight "AI content licensing revenue" in red in an Excel column, and within three years content licensing will be part of the business model right alongside subscriptions. Stratechery and Bloomberg will get along fine without OpenAI, because they have direct relationships with readers who are willing to pay. The losers will be independent regional outlets, trade blogs, and others too big for Substack and too small for venture funding — the ones that spent ten years living off search traffic and never built anything beyond an SEO strategy. They don't have much of a voice, they don't have lawyers, and they have months, not years, to adapt. Raw Story, whose lawsuit Judge McMahon dismissed, comes from that tier.

Here a secondary but entertaining subplot enters the picture. In 2024, a group of researchers published a paper in Nature describing a phenomenon they called "model collapse": when an AI is trained on text other AIs generate, after several iterations it begins to degrade — losing rare vocabulary, impoverishing its distribution, and gradually dissolving into averaged-out noise [cite: 1, 2]. The effect has since been confirmed across several architectures, and while the industry has been trying to treat it with mixed datasets, the core problem hasn't budged.

Clean, human-generated text turns out to be a critical strategic resource. Not a "cultural heritage" (that's too solemn) — a resource, like rare-earth metals. For writers, this discovery is equal parts humiliating and encouraging. On the one hand, you've just been officially appraised by the ton. On the other, whoever controls the raw material has at least a theoretical point of leverage.

It's here, at this fork between "we've been robbed" and "we have leverage," that the essential task presents itself: ruthlessly dissecting our own comfortable illusions about the past.

The narrative of "The End of the Link Era" is, in large part, an exercise in revisionist history. You'd almost think that before OpenAI arrived, publishers were living in some kind of media paradise, where Google graciously walked readers by the hand to the original source and that source received a fair slice of the pie in return. That, of course, is fiction — and the editors writing these manifestos know it better than anyone. Google was never a partner to the press. Google was a middleman with its own algorithm, its own economics, and its own right to zero out your traffic because something shifted in the rankings. For twenty years, media companies adapted to that algorithm through gritted teeth, hiring SEO consultants and rewriting headlines a third time because the first and second versions didn't perform in the rankings.

The era isn't changing. The vendor is. Before now, you paid SEO agencies to keep the Google robot happy. Now you'll pay few-shot prompting engineers to keep the OpenAI robot happy. The budget moves from one line item to another. That's genuinely unpleasant for anyone who just finished paying off a mortgage on SEO revenue, but it's hard to call it the end of an age.

A new era is when the rules change. In this case, we’ve only switched cashiers.

Part 3. What Remains When the Links Are Gone

While the lawyers for OpenAI and the NYT trade lawsuits and bloggers publish manifestos about the death of original content, something far less dramatic and far more telling is happening at the University of Chicago. Ben Zhao's lab released two tools with names straight out of a teenage goth phase: Glaze and Nightshade. The first is an invisible "glaze" applied over an image, distorting it for machine vision while leaving it unchanged for the human eye. The second is a system of “poisoning” images in training datasets, which can shift entire categories within a model until a dog starts looking like a cat to the AI [cite: 3, 4]. Nightshade was downloaded more than a quarter of a million times in just the first five days after its release [cite: 5].

I wouldn't call these "tools of resistance" — that framing sits awkwardly with the engineering nature of the project. They’re a symptom of the greater issue. Creators recognized something fairly straightforward: legal protection doesn't work, lobbying only works if you have a lobby, and collective bargaining in creative industries functions roughly as effectively as an anarchist trade union. When the legal system refuses to see harm (looking at you, Judge McMahon), people start protecting themselves directly, even embedding that protection into the work itself.

The music industry needed roughly twenty years to go from the first recognition that digital copying had made its business model obsolete to the moment Spotify started paying rights holders something meaningful. In the text and visual ecosystem, the same cycle took eighteen months, not because we got smarter but because we had already seen the pattern before.

The current operational model for AI corporations demonstrates why people use Nightshade. In May 2024, OpenAI unveiled the GPT-4o voice assistant. Journalists immediately clocked the “Sky” voice option as strikingly similar to Scarlett Johansson, whose character in Her (2013) was effectively the conceptual blueprint for the entire project. Johansson responded that OpenAI had approached her about voicing the assistant and she had declined. Two days before the launch, Sam Altman reached out to her agent with another offer and got another no. Sky appeared in the product anyway, then vanished after the backlash [cite: 6, 7]. OpenAI insists Sky was recorded with a different actress, and that is most likely true, but the effect they were going for was still obvious.

The most interesting thing about this episode is not the violation. Strictly speaking, there may not be one. What's striking is the sequence: asked — refused — made it sound similar anyway — backed off under threat of a lawsuit. This is not malice, nor is it corporate ethics running off the rails. It is the company's operational model: ask permission where you have no choice, and skip it where you can apologize later. OpenAI behaves exactly the same way with text, but articles have no agent and no face. If Raw Story had Scarlett Johansson's cheekbones, Judge McMahon would have given the lawsuit a much longer look.

"The End of the Link Era" is a striking formulation, but a false one. Links regularly show up in academic papers, on Wikipedia, in The New York Times, and in this very column. What is ending is a short, historically contingent period in which exactly one intelligible intermediary stood between a text and its reader, operating on exactly one intelligible algorithm that gave rise to an entire profession. That intermediary is leaving. The replacement has not yet announced its rules, and in that pause — that technical, legally murky, economically anxious pause — two kinds of players win: those who have direct relationships with an audience willing to pay (Stratechery, niche Substacks with a thousand devoted readers), and those with enough lawyers for the long haul (NYT, News Corp, Axel Springer).

Independent newsrooms without paid subscriptions, mid-sized trade blogs, regional outlets, and tech sites that live on referral traffic find themselves in the position of medieval scribes at the debut of Gutenberg's printing press. Scribes didn't disappear overnight; they faded over roughly eighty years, gradually retraining as typesetters, proofreaders, and illustrators. Their work didn't vanish — it transformed. The contemporary mid-tier media landscape seems headed for exactly that kind of slow, not particularly graceful reinvention in some new role. Nobody has defined that role yet; we’ll figure it out mostly by trial and error.

That, if we're being completely honest, is the real problem: not OpenAI, not SearchGPT, not dying traffic, but the gap between intermediaries where nobody owes anybody anything. Legally, courts demand proof of harm, and that proof has so far proved impossible to find in the wild. Economically, licensing deals go to those who can afford to litigate. Moral categories have never mapped cleanly onto market processes.

One last thing. This column you just read — SearchGPT will summarize it in three sentences on the very first query, get Judge McMahon's name wrong, attribute the Status Labs accuracy figures to Google, and cite none of the original sources. That may, in fact, be the only truly compelling argument in defense of links, but I'm afraid the machine won't cite it either.