model collapse

advancedexperimental

The phenomenon in which an AI model's quality degrades from generation to generation because, instead of human content, it increasingly learns from material produced by other models.

In plain terms

Imagine photocopying a photograph, then copying the copy, and then copying that copy. In the first few rounds, the difference is barely visible, but on the twentieth copy, there are only gray smudges. Model collapse is the same phenomenon in artificial intelligence: if a model learns from texts or images generated by a previous model, small errors and simplifications amplify from round to round.

First, rare things disappear: unusual phrasing, minority viewpoints, and special cases. The model increasingly repeats the 'average', making its output flatter and more monotonous. In an advanced state, it produces meaningless or self-repetitive content.

The topic became important because the internet is quickly filling with AI-written articles, images, and comments, and future models are largely trained on data collected from the internet. If no one pays attention to what enters the training data, models begin to learn from their own "echoes".

How it works

A model learns language or visual patterns from human content. This knowledge is never perfect: it captures frequent patterns well, but records rare ones inaccurately or not at all.
When the model generates content, it leans toward the most probable solutions, making rare cases in its output even rarer than in the original data.
This content goes online and also becomes part of the next model's training data, often unlabeled, mixed in with human texts.
The new model learns from a narrower, distorted worldview and further narrows it itself. Researchers call this early collapse: the edges of diversity wear away while average performance still appears good.
If the loop is repeated many times, and the output of previous generations displaces human data, the late collapse arrives: the model's output barely resembles the original reality. If the original human data remains in the training material throughout, the degradation can be strongly curbed.

An everyday example

You search for a recipe, and the first ten results say almost the same thing word for word, with the same meaningless introductions, because they were generated by AI based on each other's texts. If the next chatbot is trained on such pages, it will know even less about your grandmother's special dish than its predecessor.

Who is affected

developerWhen training or fine-tuning a model, you must know where your data comes from and how much of it is machine-generated. → Code

entrepreneur, executiveYour company's own human-written documents and data become more valuable, as it becomes increasingly difficult to find verified human-origin training material online.

marketerIf you mass-produce AI-generated text, you also contribute to the flood of uniform content in which your own brand voice gets lost. → Marketing

everyoneThe variety and reliability of information found on the internet depend on what the models learn from.

A common misconception

MythModel collapse means today's chatbots get dumber on their own, day by day.

In realityA finished model does not degrade from use; the phenomenon occurs when successive model generations are trained and most of their data was machine-generated. It is not inevitable either: careful data selection and preserving human data can largely prevent it.

Limits and risks

The strongest evidence comes from laboratory experiments where models were intentionally trained only on their own output; on the real internet, the situation is more mixed, and the extent of the effect is debated.
Deterioration is initially barely visible because average performance remains good, while data from rare cases and minority groups already fade out.
AI-generated content cannot be reliably detected today, making post-hoc cleaning of training data difficult and costly.
Synthetic data is not harmful in itself: used in a targeted way, checked, and mixed with real data, it can be useful; the problem is indiscriminate, repeated recycling.

Similar, but not the same

hallucination – A hallucination is a fabricated, false claim in a single response, whereas model collapse is the gradual quality degradation of models across generations.
mode collapse – Mode collapse is a training error in a single generative model (known especially with GAN-type image generators) where it can only produce a few types of output; it does not require multiple model generations.
catastrophic forgetting – In catastrophic forgetting, the model loses its old knowledge while learning a new task, not due to machine-generated data.

Related concepts

Broader

training data, machine learning, data quality, generative AI

Narrower

early model collapse, late model collapse

On the edge

synthetic data, hallucination, data poisoning, watermarking, bias

A brief history

The term was introduced by Ilia Shumailov and his colleagues in 2023 in their study titled 'The Curse of Recursion'. Their results were published in the journal Nature in 2024, at which point the concept became widely known.

Where things stand

By the end of 2026, model collapse is a well-documented risk in experiments, but no full collapse has been documented in large, live models. Research and developers are focusing on tracking the origin of data, preserving human data, and labeling AI-generated content.

Status: 2026 November

Also known as

model collapse · model collapse · model collapse · AI model collapse · AI model collapse · Model Autophagy Disorder (MAD) · Habsburg AI · MI-kannibalizmus

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.