What Is Model Collapse, and Is It the Same Thing as the Web Writing Alike?

Model collapse is the degradation of models trained, generation after generation, on earlier models' output, rare content first. It is a training-loop result, not a reading of today's web.

By Jack Stovell · Published 20 October 2026

Short answer: model collapse appears when models are trained in a loop on earlier models’ output, and whether it appears depends on what happens to the original human data. This site tests a different question: whether the writing people actually publish and read has grown alike. Evidence for one is not evidence for the other.

Where the term comes from

The paper most people mean is “AI models collapse when trained on recursively generated data” by Ilia Shumailov, Zakhar Shumaylov, Yiren Zhao, Nicolas Papernot, Ross Anderson and Yarin Gal, published in Nature in July 2024. Its abstract states the finding plainly: “indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear.”

The formal definition calls model collapse “a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting the training set of the next generation.” The paper describes two stages. In early collapse, “the model begins losing information about the tails of the distribution”. In late collapse, the model “converges to a distribution that carries little resemblance to the original one, often with substantially reduced variance.”

In plain terms: train a model, sample from it, train the next model on those samples, and repeat. Rare things drop out first. The common middle survives, and over enough rounds even that drifts.

What the experiments did and did not do

The authors show the effect in Gaussian mixture models, variational autoencoders and a language model. For the language experiment they fine-tuned OPT-125m, a small model released by Meta, on the wikitext2 dataset. Each new generation was then fine-tuned on text produced by the one before. They describe the setting as “a proof of concept” and say they chose not to repeat it with larger models trained from scratch.

That is a careful result. It is also a result about a stated setup: a named model, a named dataset and a defined loop. The paper did not measure the public web, estimate how much of it is machine-written, or show that any commercial model now in use is degrading. Those are separate jobs, and the paper does not claim to have done them. Its wider warning, that “the use of LLMs at scale to publish content on the Internet will pollute the collection of data to train their successors”, is an argument about future training data that builds on the experiments. Today’s pages were never measured.

The slide usually happens in retelling. A lab result becomes a headline, and the headline becomes a claim about the world. By the third retelling it sounds as though someone watched the web eat itself. Nobody in this paper did.

Replace or accumulate: the condition that changes the result

The Nature paper’s own language experiment ran two settings. One had “no original data retained for subsequent runs”. The other had “10% of original training data preserved”, and there the authors report that “preservation of the original data allows for better model fine-tuning and leads to only minor degradation of performance.”

A 2024 preprint picks up exactly this point. “Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data” is by Matthias Gerstgrasser, Rylan Schaeffer and twelve co-authors (arXiv, April 2024). It argues that earlier studies “largely assumed that new data replace old data over time, where an arguably more realistic assumption is that data accumulate over time.” The abstract reports that “replacing the original real data by each generation’s synthetic data does indeed tend towards model collapse” and that “accumulating the successive generations of synthetic data alongside the original real data avoids model collapse”. The authors report similar results for diffusion models and variational autoencoders, plus a proof in a simplified linear setting.

So the two papers do not simply contradict each other. The Nature paper calls the process inevitable “even for cases with almost ideal conditions for long-term learning” within its setup. The preprint asks what happens when the setup changes. Throw the real data away and collapse appears. Keep it and add to it, and in their experiments it does not.

Commercial training pipelines are mostly not public, so it is hard to say which extreme they resemble, if either. Treat any flat “collapse is inevitable” or “collapse is a myth” with suspicion. The answer turns on the condition. Note the venues as well. Nature is peer reviewed. The arXiv page records a submission and a revision, which tells you nothing about peer review.

Model collapse versus a web that reads alike

This site tests a different claim: that since ChatGPT, published writing has drifted into one machine voice. The theory page sets out that claim and how the court tests it.

The two claims differ in what they study and how. Model collapse concerns a model’s outputs after it is trained on its predecessors’ outputs, and you show it by running the loop. Web convergence concerns text people read, and you can only show it by measuring that text. Two measurements matter: how much of the text is machine-made (prevalence) and whether it has grown more alike (convergence). Neither claim follows from the other.

They could connect. If much new web text is machine-written and it ends up in future training sets, the loop in the Nature paper looks less hypothetical. That is a hypothesis and should be labelled as one. The inference fails in the other direction too. A web that reads alike would not prove that any model has collapsed, and a model collapsing in a lab does not prove that the web reads alike.

What the court can and cannot say

The cases on this site measure published writing. None of them measures model training, so none can tell you whether a model is collapsing. What they can show is what has been counted about the text itself, and they divide along the line this post draws.

• Prevalence. Half the articles, and a plateau estimates what share of new English web articles are primarily AI-generated. It uses commercial detectors and carries their false-positive caveat.
• Convergence in the wild. The news that didn’t flatten asks whether English news articles became lexically more alike after large language models spread.
• Convergence in an experiment. Better stories, more alike tests whether writers given GPT-4 story ideas wrote stories that resembled each other more.
• Origins. Where the AI-isms came from looks at which kinds of human writing already used the habits now called AI-isms before ChatGPT.

Each case states its question, sample and caveats up front. Read the numbers there.

What to check before repeating the claim

1. Which loop? Look for a named model, a dataset and a number of generations. Without them it is a general assertion.
2. Was the original data replaced, partly kept or accumulated? The answer changes the result.
3. What is the claim about? A model’s outputs and the published web are different things.
4. Is there a measurement with a defined sample behind it? Or is it a story about what might happen?
5. What is the source? Is it a peer-reviewed paper or a preprint, and have you read the primary text or only a summary?

If someone cites the Nature paper to say the web has homogenised, they have jumped from a training loop to a corpus. If someone cites the accumulation preprint to say there is nothing to worry about, they have jumped too, because it also describes training conditions. Neither paper reads today’s pages. For that, go to the cases.

Sources

• Shumailov, Shumaylov, Zhao, Papernot, Anderson and Gal: AI models collapse when trained on recursively generated data (Nature, 2024)
• Gerstgrasser, Schaeffer et al.: Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data (arXiv preprint, 2024)

Questions people ask

What is model collapse?

Shumailov and colleagues define it in Nature (2024) as "a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting the training set of the next generation." The tails of the distribution go first.

Do AI models get worse over time?

Not as a general rule. Degradation has been shown in experiments where each model is trained on the previous model's output. Whether it appears depends on the setup, above all on whether the original human data is thrown away or kept.

Is model collapse happening to ChatGPT now?

Nothing cited here can tell you. The papers test controlled training loops, and commercial training pipelines are mostly not public. This site's cases measure published writing, not how any model is trained.

Is model collapse the same as the dead internet theory?

No. Model collapse is a result about training models on their predecessors' output. The dead internet theory and this site's narrower question are claims about the web itself. The theory page sets out what the court tests.

Can model collapse be avoided?

A 2024 preprint by Gerstgrasser and colleagues reports that "accumulating the successive generations of synthetic data alongside the original real data avoids model collapse" in their experiments. The Nature paper also found that keeping some original data led to only minor degradation.