The serpent eating its tail: Model collapse in generative AI

Updated October 2026

Every generative model is an imperfect copy of its training data. When the next model learns from that copy instead of the original, the losses compound. Researchers call the result model collapse, and they often compare it to the ouroboros, a serpent eating its own tail. With most new web pages now containing some machine-written text and teams relying more on synthetic data, it has become a practical data-pipeline question rather than a thought experiment.

What is model collapse?

Model collapse is a degenerative process in which generative AI models trained on recursively generated data, the outputs of earlier models, progressively lose information about the true data distribution. Rare patterns vanish first. Eventually outputs become repetitive, homogeneous, and disconnected from reality. It is also called model decay, AI cannibalism, or model autophagy disorder (MAD).

The term was established by Ilia Shumailov and colleagues from Oxford, Cambridge, Toronto, and Imperial College in a 2024 paper in Nature. They fine-tuned the OPT-125m language model on Wikipedia text, then trained each new generation on text written by the previous one. A prompt about 14th-century architecture ended, nine generations later, in a list of jackrabbit types. The same effect appeared in variational autoencoders and Gaussian mixture models, which means it is a property of learning from generated data in general, not a quirk of one architecture.

Early and late collapse

Early collapse is quiet. The model loses low-probability events: specialist terms, minority dialects, rare medical conditions, unusual but valid reasoning paths. Suppose 1% of the original data uses specialized medical vocabulary. A model trained on it may reproduce that vocabulary in only 0.5% of its outputs. The next model, trained on those outputs, sees even less of it. Aggregate benchmarks barely move, so the loss often goes unnoticed.

Late collapse is visible. After enough generations, the model loses most of its variance. It confuses concepts, repeats itself, and converges on a narrow set of outputs that bear little resemblance to the original distribution. Mathematically, a rich distribution shrinks toward a single point.

Why model collapse happens

Shumailov and colleagues traced collapse to three sources of error. Each is small on its own. Recursion makes them compound.

•  Statistical approximation error. Any training set is a finite sample. Events with low probability may simply not be sampled, and once they are missing from one generation, the next generation cannot learn them back. This is the main driver.

•  Functional expressivity error. A model cannot represent the true distribution exactly, so it assigns probability in the wrong places, sometimes to outputs that never occurred in reality.

•  Functional approximation error. Training procedures such as stochastic gradient descent have their own biases, so even a perfectly expressive model would not learn the distribution perfectly.

A factor most explanations miss: synthetic data is rarely sampled at full diversity. Teams generate it with low temperature or nucleus (top-p) sampling to keep quality high, and those settings cut off the least likely tokens before any training happens. The tail is trimmed twice, once by decoding and once by learning. It works like photocopying a photocopy with the contrast turned up each time.

Is model collapse inevitable? What the research says

Early coverage treated collapse as unavoidable. Research since 2024 gives a more useful answer: it depends on how synthetic data enters the pipeline.

How to read these results together

The studies do not contradict each other. They measure different kinds of collapse.

Catastrophic collapse is avoidable. When real data is kept and synthetic data is added to it rather than swapped in, models do not degrade into gibberish. Gerstgrasser and Roudi both show this.

Subtle collapse is harder to avoid. Lost tails, homogenized outputs, and diminishing returns from scale appear even when synthetic data is a tiny fraction of the mix. This is the version most relevant to production teams, because it degrades exactly the rare, high-value cases that differentiate a model.

The limits matter. The single-data-point result was proven for simple statistical models and observed in restricted Boltzmann machines. Nobody has shown that it holds for frontier-scale language models. Treat it as evidence that real data is a powerful anchor, not as a dosage guide.

How much of the web is already AI-generated?

Two large 2025 studies put numbers on the contamination. Ahrefs analyzed 900,000 new English-language pages in April 2025 and found that 74.2% contained AI-generated content. Only 2.5% were purely AI-written. 71.7% mixed human and AI text, and 25.8% were purely human. A separate Graphite analysis reported by Axios sampled 65,000 Common Crawl URLs and found that AI-written articles rose from about 5% in 2020 to about 48% of new articles by May 2025, briefly passing human-written articles in November 2024 before levelling off.

Three implications follow for anyone building training data:

•  Detection cannot be the only filter. Most contaminated content is mixed, and detectors classify mixed text poorly. A binary AI-or-human filter will both miss AI text and discard useful human writing.

•  Data vintage and provenance gain value. Text collected before 2022, licensed archives, and data with known human authorship are more trustworthy inputs for pre-training datasets, which is why frontier labs invest heavily in curation.

•  Retrieval is affected too. A RAG system is only as reliable as the corpus it retrieves from. Grounding a model in a web index that is half synthetic does not guarantee grounding in reality.

Warning signs: how to detect model collapse early

Aggregate accuracy is the wrong place to look, because early collapse hides in the tail. These signals catch it sooner. Track them across every data refresh and model version, not once.

The last signal is the hardest to automate and the most reliable. Independent AI evaluation with domain experts catches degradation in exactly the cases where automated judges, often models themselves, share the same blind spots. For more on why this matters, see the human difference in high-stakes AI evaluation.

Where the risk is highest

Not every use of generated data carries the same risk. The deciding factors are whether real data stays in the loop and whether something outside the model checks the output.

How to prevent model collapse

The research points to six practices. The first two cost little. The rest are where most of the protection comes from.

1. Accumulate, never replace

Keep the original human-generated data in every training mix and add synthetic data on top of it. This one decision separates the collapse results from the no-collapse results in the literature.

2. Track provenance and data vintage

Tag every source with origin, collection date, and generation method. Watermarks and AI detectors help, but treat them as partial filters, not guarantees. You cannot control the ratio of real to synthetic data if you cannot tell them apart.

3. Refresh the long tail with expert human data

Fresh human data matters most where the distribution is thinnest: specialist reasoning, rare edge cases, minority languages, professional judgment. General web text will not cover these, and synthetic data will thin them further. Targeted post-training data written or reviewed by domain experts puts back what recursion removes.

4. Verify synthetic data before training on it

Synthetic data is safe when something outside the generating model checks it. That can be a programmatic verifier, a rubric-based review, LLM-based quality assurance backed by human spot checks, or full expert review for high-stakes domains. A practical walkthrough is in our guide to building a synthetic data pipeline for SFT. When choosing between AI and human feedback for alignment, see RLAIF vs RLHF.

5. Ground learning in objective feedback

Imitation learning copies text, so it copies the errors in that text. Reinforcement learning in an environment learns from outcomes: the code passes its tests, the agent completes the booking, the proof checks out. Because the reward comes from the environment rather than from another model's writing, the signal does not degrade with each generation. This is why RL environments and verifiable rewards have become central to post-training. For a technical introduction, see inside the RL gym.

The caveat: environments shift the risk rather than remove it. Agents exploit loopholes in reward functions, and an environment only teaches what its designers anticipated. Human-designed tasks, realistic edge cases, and regular audits of agent trajectories keep the reward signal honest.

6. Evaluate independently, with humans in the loop

Use evaluation sets that no model in your pipeline has generated or seen, and keep human raters in the loop for the slices that matter. Reliability across repeated runs reveals degradation that single-run benchmarks miss. Toloka Arena applies this principle to agents, ranking them on how consistently they complete real-world tasks across repeated attempts.

Why verified human data is gaining value

Shumailov's team noted a "first-mover advantage": organizations that collected large volumes of human-generated data before the web filled with model output hold an asset that is getting harder to recreate. Newcomers face a choice between scraping an increasingly synthetic web and sourcing verified human data directly. The second option is now practical at any scale, from commissioning expert-written data to requesting ready-made datasets with documented provenance.

Frequently asked questions

What is model collapse in AI?

Model collapse is the gradual loss of quality and diversity in generative AI models that are trained on data produced by other models, including earlier versions of themselves. Rare information disappears first. Over many generations, outputs become repetitive and drift away from reality. The term was popularized by a 2024 Nature paper by Shumailov and colleagues.

Is model collapse inevitable?

No, not in its severe form. Research from 2024 to 2026 shows that collapse into nonsense is avoided when synthetic data is added to real data rather than replacing it. A milder form, where rare knowledge thins out and scaling gains shrink, can still appear with small synthetic fractions, so data provenance and fresh human data still matter.

Is model collapse happening to today's AI models?

There is no public evidence that a major frontier model has collapsed. Leading labs filter training data, retain large human-written corpora, and verify synthetic data. Subtler effects, such as more homogeneous outputs, are harder to measure and remain an active research question. The risk is higher for smaller teams that fine-tune on unverified generated data.

What is the difference between model collapse and mode collapse?

Mode collapse is a failure within a single training run, most familiar from generative adversarial networks, where the generator produces only a few kinds of output. Model collapse happens across generations of models, when each one trains on data generated by the last. Both reduce diversity, but they have different causes and fixes.

How much real data is needed to prevent model collapse?

There is no universal ratio. A 2026 Physical Review Letters study found that one real data point prevents collapse in simple statistical models, and keeping all original data prevents severe collapse in larger experiments. For production models, the practical rule is to keep every original human example and add fresh human data where coverage is thinnest.

Can synthetic data be used safely for training?

Yes, when it is verified before training, mixed with real data rather than substituted for it, used to fill specific gaps, and tracked with clear provenance. Synthetic data becomes risky when a model trains on its own unfiltered output over multiple rounds.

Conclusion

Model collapse is real, measurable, and largely preventable. The serpent only eats its own tail when real data leaves the loop. Teams that keep original data, track where every example comes from, verify synthetic data, and anchor training in human judgment and objective outcomes can use generated data without paying the recursion tax.

As the share of machine-written text on the web keeps growing, the bottleneck in AI development moves from collecting more data to collecting trustworthy data. That makes the long tail, the rare and expert knowledge that recursion erodes first, the most valuable part of any training set.

Related reading

•  How post-training teaches LLMs to reason

•  The backbone of large language models: understanding training datasets

•  Synthetic data is now available on Toloka Platform

•  LLM evaluation framework: principles, practices, and tools