Key points
- Model collapse is what happens when AI output gets fed back in as training material, over and over, and the results grow less varied each time.
- The first thing to go is whatever was already rare. Common patterns survive; the edges get shaved off.
- It does not happen in one step. When a model is retrained again and again on only the previous generation's output, things narrow a little each generation. Whether that also happens in real settings, where human-made material keeps piling up alongside, is something researchers see differently.
- It is a different problem from the similarly named mode collapse, which happens inside a single model's training run, not across generations.
- Managing the material is the most commonly cited way to prevent it, and keeping human-made data coming in, controlling how much synthetic data gets mixed in, and keeping originals archived separately are named as the basics.
The analogy
Record a sound in a room, then play that recording back through a speaker and record it again. Once, almost nothing changes. Do it two or three more times, and it starts to shift. The very low and very high sounds drop out first, leaving only the middle range.
After about ten rounds, the many sounds that were once in the room are gone, replaced by a single dull hum. Whatever the speaker cannot reproduce well never makes it into the next recording, and once a sound drops out, it never comes back.
The tricky part is that you cannot tell what is missing just by listening to the last recording. Only someone who remembers the original room can tell what dropped out. That is exactly what happens when AI is trained on AI output, generation after generation.
In detail
AI output is creeping into the training material
Large models learn by scraping up the text and images piled onto the internet. But the share of that material actually made by AI is climbing fast. When the next generation of models is trained, some of what was made by the previous generation ends up mixed into the pile.
The trouble is that it is hard to sort out. Tools built to spot AI-made content only give a probability, and anything a person has touched up gets even harder to call. Whoever is collecting the data usually has no idea which pieces are which.
That makes this risk less a single company's mistake than a matter of the whole shared pool slowly getting mixed in. It is also why material from the era before AI output was common is becoming more valuable. How much it affects real models is still a matter of differing views.
The rare cases disappear first
A model imitates the distribution it learned from, but it imitates the common parts well and the rare parts poorly. So its output already contains less of the rare stuff than the original had. Train the next model on that output, and the rare share shrinks again.
This shaving repeats every generation. Regional dialects, unusual writing styles, uncommon compositions and color choices — the things sitting at the edges — go first. The middle holds up for a long time while the edges erode.
Surface quality stays fine for a while. Sentences read smoothly, images look clean. But underneath, everything is quietly drifting toward looking the same no matter what you ask for.
It only takes a few generations
Small-scale experiments show the pattern clearly. Train a model only on the previous generation's output, repeat that a few times, and the spread of results narrows noticeably within a handful of generations.
Errors stack up here. There is the gap from having only a limited sample, the gap from a model never fully capturing the true distribution, and the gap from training never converging perfectly. Each generation adds its own version of these, and the previous generation's approximation gets treated as ground truth by the next.
Real-world conditions are messier than the experiments, though. Human-made material keeps arriving, and there is a filtering effect too, since only the generated output people actually chose to keep survives. So collapse is less a foregone conclusion than a risk that shows up when the material goes unmanaged.
It is a different problem from mode collapse
The similar names make this easy to mix up. Mode collapse happens during a single model's training run — the model settles for a handful of outputs that work and gives up on the rest. It is a problem that starts and ends inside that one training process.
Model collapse is a problem between generations. One model's output becomes the next model's textbook, and that model's output becomes the textbook after that, narrowing the material itself across several generations. Training a single model well does nothing to stop it.
Preventing it is mostly about managing the material
First, keep bringing in human-made material. New human records keep refilling the edges, and archiving old original data carefully belongs here too.
Second, control the mix ratio. Rather than cutting out generated data entirely, many people consider it more realistic to keep the original human-made data and add generated data on top of it. There are results showing that accumulating alongside the original data, rather than replacing it, substantially slows the narrowing. Other studies find that effects can linger even at a small share, so views differ on which holds.
Third, mark and filter. Embedding an invisible marker at generation time and attaching a history to files lets later data collection screen some of it out. Using only material a person has reviewed and chosen helps too.
More precisely
Model collapse is described as the tail of a distribution getting shaved off generation by generation, narrowing the spread of outcomes. The cause is three kinds of error stacking up: the gap from having a finite sample, so rare events sometimes never get drawn at all; the gap from a model's limited capacity to represent the full distribution; and the gap from training never reaching a true optimum. As generations pass, one generation's approximation sits in the place of the next generation's ground truth.
The analogy breaks down in one place. Repeating a recording adds noise and makes the sound murkier, but in the early stages of model collapse the output often looks smooth and clean while rare cases disappear. What disappears is not noise but rare cases, which is why surface quality alone will not tip you off. And a recording only has one lineage, while real data pipelines keep having fresh human-made material mixed back in.
None of this means synthetic data is inherently bad. Data made for a purpose and verified by a person can genuinely help fill gaps. The problem is the loop where unfiltered output gets fed straight back in as the next round of material.
Try it yourself
Common misconceptions
It's easy to think using any AI-made data causes collapse, but actually there are results showing that keeping the original human-made data and adding to it, rather than replacing it, slows the narrowing a great deal. Other studies find effects can linger even at a small share, so views differ.
It's easy to think this is the same thing as mode collapse, but actually one is a skew that happens during a single model's training, and the other is material narrowing across generations — different problems.
It's easy to think worse output quality gives it away, but actually in the early stages sentences and images stay smooth while the rare cases quietly drain out, so it is hard to notice, and with long repetition quality itself can fall apart.
One-line summary
Model collapse is the risk that retraining AI again and again on only AI output makes the rare cases disappear first. How much it happens in practice is debated, but keeping human-made material and managing how it is mixed in is treated as the baseline precaution.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-18