ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›AI MEANINGS›WHAT IS MODEL COLLAPSE?
AI MEANINGS · MEANINGBEGINNER

What is model collapse? AI getting worse by learning from AI

Model collapse means an AI model getting worse because it was trained on AI-made content, losing rare details first until output turns bland.

READ6 min
WORDS1,162
SECTIONS7
TYPEMEANING
CHECKED1 OCT 26

Model collapse means an AI model getting worse because it was trained on text or images made by other AI models instead of by people. Each round loses a little of the rare and unusual material, until the output becomes bland and repetitive.

◈ IN PLAIN TERMS

Think of photocopying a photocopy, again and again. The first copy looks fine. By the twentieth, the fine detail has gone and everything is smudged into the same grey. Model collapse is that, for AI: copies of copies, with the unusual details fading first.

In a sentence: “If future chatbots learn mostly from AI-written posts instead of human ones, they could suffer model collapse.”

What the term means

The plain point: training an AI carelessly on AI-made content can damage it, and the damage starts with the rare cases.

The best-known study is by Ilia Shumailov and colleagues, published in the science journal Nature in 2024. They found that feeding models lots of machine-made content without care causes lasting defects.1

The “tails” are the uncommon things: rare words, unusual facts, minority views, odd styles. Each round of copying tends to drop some of them. The next model learns from what is left, so the uncommon things show up less and less.

A simple example

The plain point: it happens over generations, when each new model learns from the last one’s output.

Imagine a model trained on real dog photos, which include plenty of common breeds and a few rare ones. It makes new dog pictures, mostly of the common breeds. A second model is trained only on those pictures. It sees even fewer rare breeds, so it makes fewer still. A few rounds later, the rare breeds are gone. This is an illustration of the idea, not an experiment from the study.

MODEL COLLAPSE · COPIES OF COPIES
An illustration of the idea, not an experiment from the study: the rare cases fade first.
First model learns from real photosPlenty of common dog breeds, and a few rare ones.
It makes new picturesMostly of the common breeds.
The next model learns only from thoseIt sees fewer rare breeds, so it makes fewer still.
A few rounds laterThe rare breeds are gone; the output turns bland.
Reasoning, 1 Oct 2026 — summarises this page’s dog-photo illustration and its section on what the term means.

How an AI is actually built goes through what the Nature paper measured, including how the overall score can hide the early damage.

TAKEAWAY

Model collapse is about losing variety first. A model can still look fine on average while the unusual cases quietly disappear.

Is it inevitable? The debate

The plain point: researchers disagree on how big the real-world risk is. It depends on whether AI-made data replaces human data or is added to it.

  • The warning. As more of the web is written by AI, future models may train on more of it. The Nature authors framed their study around that future.1
  • The counterpoint. A later study found that when AI-made data is added alongside the original human data, instead of replacing it, collapse was avoided in their tests.2

In other words, the danger is greatest when real human material gets pushed out. That makes the supply of human-made pages part of the story.

Why it matters to you

The plain point: it connects AI quality to the health of the open web you use every day.

  • The web is filling with machine text. The state of the automated internet tracks how much, with sources.
  • Human writing becomes more valuable. If people stop publishing because AI takes their readers, the human material shrinks. AI and the open web covers that side.
  • It is not the dead internet theory. That theory claims the web is now mostly bots, by design. Model collapse is a narrower, tested idea about training. Dead internet theory meaning separates what is supported from what is not.

Where the word comes from

The term was introduced by Ilia Shumailov and colleagues in a preprint posted on 27 May 2023, “The Curse of Recursion: Training on Generated Data Makes Models Forget”, which said: “We refer to this effect as Model Collapse”.3 The reviewed version of the study appeared in Nature on 24 July 2024, and a free copy sits on PubMed Central, the US National Library of Medicine’s open archive.4

Questions people ask

What causes model collapse?

Training a model, again and again, on material made by earlier models instead of by people. The Nature study put it in one line: “indiscriminate use of model-generated content in training causes irreversible defects in the resulting models”.1

Can model collapse be prevented?

The research points to one answer: keep real human data in the mix. The Nature authors write that “we need to make sure that access to the original data source is preserved”,5 and a later study avoided collapse in its tests by adding AI-made data alongside the original data instead of replacing it.2

Is model collapse the same as AI slop?

No. AI slop is low-quality AI-made content that people see online; model collapse is what can happen to a model trained on too much of that kind of material. That is why the Nature authors expect real human data to become “increasingly valuable”.5 See AI slop meaning.

Learn more on this site

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. What is model collapse?

  • A model getting worse because it was trained on content made by other AI models
  • A data centre losing power
  • A chatbot that stops replying in a long chat
Show the answer

Answer: A model getting worse because it was trained on content made by other AI models. Like photocopying a photocopy, each round loses a little rare material until the output turns bland and repetitive.

2. What tends to disappear first in model collapse?

  • The most common words and images
  • The model’s ability to use numbers
  • The rare and unusual material
Show the answer

Answer: The rare and unusual material. Each round tends to drop some of the uncommon things, so a model can look fine on average while the unusual cases fade.

3. What does the research suggest helps avoid model collapse?

  • Training only on AI-made content
  • Keeping real human data in the mix instead of replacing it
  • Making the model much bigger
Show the answer

Answer: Keeping real human data in the mix instead of replacing it. A later study avoided collapse in its tests by adding AI-made data alongside the original human data.

Related terms: Inference · Latent space · Mixture of Experts (MoE) · Open weights · Overfitting · Parameters (weights)

SOURCES
  1. 1Shumailov et al., AI models collapse when trained on recursively generated data, Nature, 24 Jul 2024, read at source 1 Oct 2026“We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear. We refer to this effect as ‘model collapse’”. ↑a“Here we consider what may happen to GPT-{n} once LLMs contribute much of the text found online.” ↑b“We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear.” ↑c
  2. 2Gerstgrasser et al., Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data, arXiv:2404.01413, revised 29 Apr 2024, read at source 1 Oct 2026replacing “real data by each generation’s synthetic data does indeed tend towards model collapse”, while “accumulating the successive generations of synthetic data alongside the original real data avoids model collapse”. ↑a“accumulating the successive generations of synthetic data alongside the original real data avoids model collapse” ↑b
  3. 3Shumailov et al., The Curse of Recursion: Training on Generated Data Makes Models Forget, arXiv:2305.17493, submitted 27 May 2023, read at source 12 Oct 2026“We refer to this effect as Model Collapse and show that it can occur in Variational Autoencoders” ↑
  4. 4Shumailov et al., AI models collapse when trained on recursively generated data, Nature 631(8022):755–759, 24 Jul 2024, open-access copy on PubMed Central (CC BY 4.0), read at source 12 Oct 2026“We refer to this effect as ‘model collapse’ and show that it can occur in LLMs as well as in variational autoencoders” ↑
  5. 5Shumailov et al., AI models collapse when trained on recursively generated data, Nature, 24 Jul 2024, PubMed Central copy, read at source 12 Oct 2026“we need to make sure that access to the original data source is preserved” ↑a“the value of data collected about genuine human interactions with systems will be increasingly valuable” ↑b

Shumailov et al., AI models collapse when trained on recursively generated data, Nature 631, 24 Jul 2024, read at source 1 Oct 2026.

Gerstgrasser et al., Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data, arXiv:2404.01413, revised 29 Apr 2024, read at source 1 Oct 2026.

The comparisons on this page (photocopying a photocopy, the dog photos) are this site’s own ways of explaining, not claims from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE