Model Collapse

Model collapse is a phenomenon where AI models trained on data generated by other AI models progressively lose quality and diversity over successive generations. Each generation of the model produces output that is slightly less varied and nuanced than the previous one, until the output becomes repetitive, generic, and low-quality.

How it happens

The process unfolds in stages:

1. Generation 1: A model trained on human-created data produces high-quality, diverse output.
2. Generation 2: A new model is trained partly on Generation 1's output. It performs well but with slightly less diversity — the distribution of its output narrows.
3. Generation 3: Trained partly on Generation 2's output. The narrowing accelerates. Rare but valid perspectives, unusual phrasings, and minority viewpoints begin to disappear.
4. Generations 4+: Each subsequent generation loses more of the original richness. Output converges on the most common, most average patterns — the statistical centre of the training data.

The result is AI that sounds increasingly generic. Unusual ideas, creative phrasing, and diverse perspectives are gradually squeezed out, replaced by the most probable, most average output.

The research behind it

Researchers at Oxford and Cambridge published landmark findings demonstrating model collapse in controlled experiments. They trained successive generations of language models, each generation partly on the previous generation's output, and observed consistent quality degradation. The findings showed that model collapse is not a theoretical risk — it is a mathematical inevitability when AI output feeds back into training data without sufficient human-generated content to maintain diversity.

Other research teams have replicated and extended these findings, showing that the effect is robust across different model architectures and training approaches.

Why it matters for the internet

The internet is increasingly full of AI-generated content. Blog posts, articles, social media comments, product descriptions, and reviews are being produced by AI at enormous scale. If future AI models are trained on this AI-saturated web, model collapse becomes a real risk at the civilisation level.

This creates a paradox: the better AI gets at generating content, the more AI content appears online, and the harder it becomes to train the next generation of models without encountering its own output.

Practical implications for content creators

Model collapse matters for anyone creating content, even if you are not training AI models:

• AI-generated content becomes more generic over time as the models that generate it converge on average patterns. If you are relying heavily on AI to create all your content, you may notice it becoming blander and more formulaic.
• Original human writing becomes more valuable, not less. As AI-generated content floods the internet, genuinely original human perspectives, unusual insights, and authentic voices stand out more.
• The best approach is human-AI collaboration: Use AI for first drafts, research, and structure, but add your own insights, experiences, and perspectives. This keeps the human signal strong.

How to avoid contributing to model collapse in your work

• Do not publish raw AI output as finished content. Always edit, add your perspective, and inject original thinking.
• When building AI training datasets, prioritise human-created source material.
• Be sceptical of AI-generated "research" that may itself be derived from previous AI outputs.
• Value and invest in original writing, reporting, and creative work — it is the foundation on which useful AI depends.

The bigger picture

Model collapse highlights a fundamental dependency: AI models are only as good as their training data. The quality of AI depends on a continued supply of diverse, high-quality, human-generated content. This gives original human creators — writers, researchers, journalists, artists — a critical role in the AI ecosystem, even as AI becomes more capable.