Every large language model in production was trained on the open web — the archived pages, the forum answers, the Wikipedia edits, the decades of writing that people put online mostly for free. That corpus is the reason any of this works. And a growing argument says the same technology is now quietly dismantling the thing it was built from.

The sharpest version of that argument this week comes from Vass Bednar in The Walrus, whose essay "Google Search Is Dying. What Comes Next Is Worse" drew a large, sustained discussion on Hacker News — usually a sign a piece has named something people were already feeling. The essay is worth engaging with directly, because the mechanism it describes is not a slogan — it's a feedback loop with an economic engine.

The loop

Start with what AI search actually does to the page underneath it. When a model answers your question in an overview box, you often never click through to the source. Bednar cites the small, telling failure mode — an AI summary confidently returning the wrong sunset time — but the fabrication is not really the point. The point is interposition: the model sits between you and the original, and when it is wrong, the correct page might as well not exist. When it is right, you still don't visit.

Now follow the money that used to flow through that click. A site earns its keep — ad revenue, subscriptions, donations, simple motivation to keep publishing — from people arriving. Strip the arrivals and you strip the reason the page exists. Wikipedia is the cleanest example of the paradox: AI systems increasingly read it directly to answer questions, rather than sending readers to it, which means the encyclopedia trains the models that reduce its own traffic, donations, and volunteer pipeline. It is, in Bednar's framing, building the infrastructure of its own decline.

That is the loop. The web produced the training data; the trained models reduce the traffic that funds the web; a thinner web produces less of the reliable, human-made data the next models need. Each turn is individually rational for the party taking it and collectively corrosive.

The disappearances are already happening

The essay's stronger move is to point out that this isn't a forecast — the erosion is visible now, and it's invisible by design. Link rot deletes pages continuously and silently; Bednar cites even the Library of Congress briefly losing sections of the U.S. Constitution from its site to a coding error. Whole archives vanish on a corporate decision: she points to the near-total deletion of the data-journalism outlet FiveThirtyEight's archive after Disney's layoffs. The Internet Archive — the closest thing the web has to a memory — is simultaneously fighting lawsuits and cyberattacks, while some publishers block its crawlers precisely because they fear the archive becoming AI training material.

Put those together and you get a specific kind of loss. Not censorship, which announces itself and can be resisted, but attrition — pages that simply stop resolving, archives quietly pulled, a collective record that thins without anyone deciding to thin it.

Where the argument is strongest, and where it's contestable

It is worth being honest about the limits of the thesis. "Google search is dying" has been declared for a decade, and the web has proven durable and adaptive before. Some of what AI summarisation replaces was low-value SEO sludge that few will mourn. And a model that answers a factual question directly is, for the user in that moment, plainly useful — the convenience is real, which is exactly why the traffic drains.

But the core claim survives the caveats, because it is structural rather than moral. You do not need anyone to be acting in bad faith. You need only an intermediary that is convenient enough to keep people from clicking, sitting on top of an ecosystem that was funded by clicks. The convenience and the erosion are the same event seen from two ends — which is why it is hard to argue against and harder to stop.

The part worth borrowing

Bednar's proposed direction is the least discussed and most useful part: treat core information infrastructure as a public good rather than assuming the market will maintain it. She points to Canadian precedents — CANARIE, SchoolNet, the Public Knowledge Project — and to France's state backing of privacy-preserving search alternatives, as evidence that a society can choose to fund the commons instead of waiting for it to be profitable.

That reframing matters because it changes the question. The debate about AI and the web usually gets stuck on attribution and copyright — who owes whom for the training data. Those fights are real, but they're rear-guard. The forward question is whether the archive, the encyclopedia, and the index — the load-bearing infrastructure everyone including the AI labs depends on — get treated as things worth deliberately maintaining, or left to decay because the entity best placed to profit from them has the least incentive to keep them alive. The models ate the web to learn. Whether they get to keep learning depends on whether anyone decides the web is worth feeding.