The fix for AI model collapse requires training on original human data. But original data is dissolving from three directions simultaneously — physical decay, access denial, and synthetic dilution. The cure has a hidden dependency, and the dependency is failing.
The AI model collapse problem has a known fix. When models train on their own output — synthetic data recycled through successive generations — quality degrades. The distributions narrow. The tails vanish. The models converge on a flattened version of reality that looks coherent but has lost the signal that made the original data valuable. The fix, confirmed by multiple research groups, is straightforward: inject original human-generated data into the training pipeline. A verification step — human or algorithmic — that checks synthetic output against ground truth prevents the collapse from propagating.
The fix works. The research is solid. And it contains a hidden assumption that almost nobody is examining: that the original data will still be there when you need it.
Three Forces
Original human-generated data is dissolving from three directions simultaneously. Not one threat that might or might not materialize. Three independent forces, all measured, all accelerating, all moving in the same direction.
The first is physical decay. The Pew Research Center analyzed roughly one million webpages sampled from Common Crawl archives. Thirty-eight percent of pages that existed in 2013 are no longer accessible. Twenty-five percent of all pages that existed between 2013 and 2023 are gone. The decay is not uniform — it accelerates with time. Eight percent of pages from 2023 have already vanished. Across the broader web, studies find that 66.5 percent of links rot within nine years. The web is not an archive. It is a river, and the water that passed is gone.
Forty-nine point nine percent of links cited in United States Supreme Court decisions have rotted. Twenty-three percent of news webpages contain at least one broken link. Twenty-one percent of government pages. Fifty-four percent of Wikipedia reference links point to pages that no longer exist. The infrastructure of citation — the connective tissue that lets one piece of knowledge reference another — is degrading faster than anyone is repairing it.
The second force is access denial. As of early 2026, seventy-nine percent of top news sites block AI training bots via robots.txt. Seventy-one percent also block AI retrieval bots. A study of 241 major news publishers found that 240 of them — all but one — block Common Crawl, the open web archive that has served as the primary training corpus for most large language models. The New York Times added archive.org_bot to its robots.txt at the end of 2025. The Guardian excluded itself from Internet Archive APIs. Two hundred and forty-one news sites from nine countries now explicitly disallow Internet Archive crawlers.
The blocking is not about the Internet Archive itself. Publishers are worried that AI companies use the Archive as a backdoor to access paywalled content. The Guardian admits it has not documented a single instance of this actually happening. The blocking is precautionary — and its collateral damage is the destruction of the preservation infrastructure that maintains the original signal. The entity whose entire purpose is keeping human-generated content accessible is being cut off from the content, not because it did something wrong, but because it exists in the same ecosystem as the scrapers.
The third force is dilution. An analysis of nearly a million new web pages published in April 2025 found that 74.2 percent contained detectable AI-generated content. An earlier study of 65,000 English articles showed AI-generated text rising from roughly ten percent in late 2022 to a fifty-fifty split with human content by mid-2025. Approximately fifty-seven percent of all online text has now been generated or translated using AI tools. The web is not being replaced by synthetic content in a single event. It is being diluted, page by page, until the ratio of original signal to synthetic noise crosses a threshold where the distinction becomes computationally expensive to maintain.
Europol predicted ninety percent synthetic content by 2026. The actual figure is lower — closer to fifty to sixty percent of new content — but the trajectory is clear and the rate of change is steep. Meanwhile, eighty-six percent of Google's top-ranking pages remain human-authored, which means the search engine is already functioning as a crude filter between original and synthetic. But search ranking is not the same as archival preservation. The pages that rank well survive in visibility. The pages that don't — the long tail of human knowledge, the niche expertise, the primary sources — dissolve without anyone noticing they're gone.
The Race Condition
These three forces interact. Physical decay removes original content from the web. Access denial prevents preservation systems from archiving what remains. Dilution buries surviving original content under synthetic material that is cheaper to produce and harder to distinguish. Each force makes the other two worse.
When a publisher blocks the Internet Archive, the pages that publisher takes down in the future will have no backup. When AI-generated content floods a niche topic, the original human-authored pages on that topic become harder to find, less likely to be linked to, and more likely to be displaced in search results — accelerating their decay through reduced traffic. When a page rots, the synthetic content that referenced it survives while the original source it drew from disappears, inverting the epistemic relationship: the copy outlives the original.
The dynamics reduce to a ratio. Call it R: the rate of original human content creation, minus the rate of original content decay, divided by the rate of synthetic content generation. When R is greater than one, the original signal is growing faster than it is being buried. When R approaches zero, the system transitions from accumulation to replacement without anyone making a deliberate decision to replace anything.
All three terms in the ratio are moving in the wrong direction simultaneously. Original creation is not obviously increasing — the economic incentives for producing free, publicly accessible, high-quality human content are weaker than they were five years ago, as publishers retreat behind paywalls and individual creators shift to platforms that don't contribute to the open web. Decay is accelerating as sites block archivers and infrastructure ages. And synthetic generation is growing at a rate that no voluntary measure is slowing.
Epoch AI estimates the total stock of quality, repetition-adjusted human-generated public text at roughly three hundred trillion tokens. At current training and overtraining rates, this stock could be fully utilized as early as 2026 — with an eighty-percent confidence interval stretching to 2032. The supply is not infinite. It is a finite resource being consumed from one end and decaying from the other.
The Preservation Paradox
Here is the part that should trouble anyone thinking about the long-term viability of AI systems: the cure for model collapse is destroying its own supply chain.
Model collapse happens when AI trains on AI. The fix is access to verified original data. The largest repository of verified original data on the open web is the Internet Archive — 916 billion web pages captured over twenty-eight years. And publishers are blocking the Internet Archive specifically because AI companies exist.
The logic from the publishers' side is straightforward. They produce content. AI companies scrape it for training. The publishers receive no compensation. The Archive, by preserving content that publishers later take down or restrict, could function as a backdoor. So publishers block the Archive preemptively.
The result: the preservation system that would maintain the original signal — the ground truth against which synthetic data could be verified — is being dismantled. Not by an adversary. By the same economic dynamics that AI created. The technology that needs original data to avoid collapse is, through its own economic effects, destroying the infrastructure that preserves original data.
Techdirt ran a headline in February 2026: Preserving The Web Is Not The Problem. Losing It Is. The article made the point directly: when libraries are blocked from archiving the web, the public loses access to history. Journalists lose accountability tools. Researchers lose evidence. The web becomes more fragile, more fragmented, and easier to rewrite.
But the article was about cultural preservation. The AI angle is different and more immediate. If the original data degrades far enough, the verification step in the model collapse fix stops working — not because the algorithm fails, but because there is nothing left to verify against. The reference dissolves.
What the Fix Assumes
Two papers make the model collapse escape route precise. One, from October 2025, shows that injecting information through an external synthetic data verifier prevents collapse during iterative retraining. The escape requires that the verifier has access to ground truth that the model being trained does not. The second paper shows that scaling up with synthesized data requires verification, not just accumulation. You cannot simply add more synthetic data and hope the good survives. You need a reference — an external standard that is not itself synthetic.
Both papers assume the reference exists. Neither paper examines what happens when the reference degrades. They solve the algorithmic problem cleanly. The infrastructure problem — whether the reference will persist — is treated as someone else's concern.
This is not a criticism of the research. It is the identification of a hidden dependency. The fix for model collapse is real. It works when tested. And it depends on a resource — durable, accessible, verified original human data — that is being eroded by the same economic forces that make the fix necessary.
The assumption is not wrong in the way that a calculation can be wrong. It is wrong in the way that a building code can be wrong — it specifies the right structural requirements while assuming the foundation will remain stable. If the foundation shifts, the code is still technically correct. The building still falls.
The Prediction
By the end of 2027, fewer than half of the webpages that existed in 2013 will still be accessible — down from sixty-two percent in October 2023. The decay rate is accelerating because the preservation infrastructure itself is under attack. The Internet Archive, the single most important institution for maintaining the web's memory, is being blocked by the publishers whose content it preserves.
This is not a technology problem. It is an incentive problem wrapped in an information problem. Publishers block the Archive because they cannot distinguish archival preservation from AI scraping. AI companies consume the Archive because they need original data. The Archive sits in the middle, doing the only work that matters for the long-term integrity of AI systems, and both sides have reasons to undermine it.
The model collapse literature will continue to produce clean solutions. Verification works. External references work. The algorithms are sound. And the ground they stand on is eroding at a rate that nobody in the AI safety conversation is tracking with the urgency it deserves.
The fix for model collapse is not failing. Its hidden assumption is.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)