# Model collapse: what happens when AI learns from AI

URL: https://ai-pulse.aglaganova.com/blog/model-collapse-what-happens-when-ai-learns-from-ai
Published: 2026-10-09
Author: Roman Aglaganov, Full-stack developer
Category: Research
Region: Europe

> Train a model on text written by a model, then do it again, and again. A study in Nature found that the rare stuff disappears first, and by generation nine you get jackrabbits.

![Diagram: a bell curve that gets narrower in each generation, from Gen 0 to Gen 9, ending in a line about jackrabbits](https://ai-pulse.aglaganova.com/api/media/file/collapse-1.png)

Language models learn from text written by people. But more and more of the text on the internet is now written by language models. So what happens when the next model learns from the last one’s homework?

## The photocopy of a photocopy

In July 2024, a team led by Ilia Shumailov published the answer in *Nature*. They trained a language model, used it to generate text, trained a new model on that text, and repeated the loop for several generations. They called the result *model collapse*: the models do not suddenly break, they slowly forget.

The first thing to go is the unusual. A model learns what is common and slightly under-represents what is rare; the next model learns from that slightly flattened picture and flattens it some more. Generation by generation, the tails of the distribution, the odd words, minority viewpoints and uncommon facts, disappear, and what is left becomes narrow, repetitive and confidently wrong.

## From church towers to jackrabbits

The paper’s most quoted example started with a passage about medieval church architecture. The first generations still wrote about cathedrals, if with growing confusion about which was where. By the ninth generation, a text about architecture had turned into a list of “black-tailed jackrabbits, white-tailed jackrabbits, blue-tailed jackrabbits, red-tailed jackrabbits”. Nobody asked about rabbits. The model just could not stop.

## Should we panic?

Not yet. The study’s experiments trained each generation entirely, or almost entirely, on the previous generation’s output, which is close to the worst case. Other researchers have shown that when synthetic data is *added* to human data rather than replacing it, collapse can be avoided, which is closer to how the real internet grows. AI labs also deliberately train on carefully filtered synthetic data with good results.

**The takeaway:** human-written data is becoming more valuable, not less. The archive of things people actually wrote, before chatbots joined in, is a resource worth protecting, and knowing where data came from matters as much as how much of it you have.

## Sources

- [Nature: AI models collapse when trained on recursively generated data (Shumailov et al., 2024)](https://www.nature.com/articles/s41586-024-07566-y)
- [Wikipedia: Model collapse](https://en.wikipedia.org/wiki/Model_collapse)