AI is eating itself, and the disease already has a name

In November 1987 the British Ministry of Agriculture admitted it had a new disease on its hands: bovine spongiform encephalopathy. Its human variant went on to kill more than 230 people around the world, and left behind a crisis of confidence, falling sales and more than four million cattle slaughtered. All of it even led the European Union to create EFSA (the European Food Safety Authority), laying the failures of control completely bare. In English it was never known by its technical name, but as mad cow disease, and that is how we remember it.

How does a biological disease come to affect the world of computing? Well, this is quite a can of worms (a very large one). It comes from a parallel drawn by eight researchers at Rice University, in Houston. Josue Casco-Rodriguez and Sina Alemohammad lead the research, with Richard G. Baraniuk as senior author. As you already know, the cause of mad cow disease was that cattle were being fed with the remains of cattle. Well, the origin of this condition in computing lies in feeding AI models with material generated by AI. MAD, Model Autophagy Disorder, is the name they gave it, stating it explicitly in the paper: by analogy with mad cow disease. Carrying with it the idea of a system that goes mad from eating itself. An absurdity Kafka would have signed, and Camus would have taken to the stage. The researchers coined the term in 2023, and in 2024 the renowned British journal Nature, cited worldwide, published the general version of the phenomenon. It showed that the first thing the engines stop generating is personality. Whatever was rare or stood out above the average. What survives in the generation is the norm.

Nature distinguishes two moments. In early collapse the model starts by losing the tails of the distribution, that is, the minority, the exceptional. In late collapse it is no longer that the rare disappears: everything starts to look like everything else. And it is not a process for twenty years from now: Nature itself headlined its popular piece “AI models fed AI-generated data quickly spew nonsense”. (No joke.)

To picture it: the image is that of photocopying a photocopy. Nobody needs the fifth copy explained to them. Everything is still legible, but the edges eat away the fine detail. A typeface as perfect as Helvetica could end up looking like Comic Sans.

There is, however, other research trying to find out how much generative content (content made by AI) an engine needs before reaching that collapse. And for now there is no real consensus. What is clear is that the result depends on the diet. The AI you use, what does it feed on?

First of all, it is worth getting ahead of something I can already hear: that this is fixed by writing better prompts. No. A prompt trains nothing. It arrives when the model is already made and can only ask for what is already inside it. During training the minority is lost, precisely what makes an image exceptional, and no prompt can bring it back.

Since 2 August 2025, the European AI regulation requires (article 53) the publication of a sufficiently detailed summary of the content each model was trained on. And since 2 August this year that obligation has stopped being merely an obligation. It can now be sanctioned, and the European AI Office has already started looking at who has published and who has not. A laboratory at Trinity College Dublin, the AI Accountability Lab (AIAL), led by Dr Abeba Birhane, known for auditing bias in training datasets, has been collecting that documentation since then and grading it from A+ to F: whether the information is given clearly and whether it is of any use to an author. To date they have located 44 summaries; of these, 39 are already graded, and they have drawn up another list of nineteen models that ought to have that report and do not.

Among the image generators, Bria with its FIBO model takes a B+ for clarity and an A+ for usefulness, and its 3.2 sits just behind. Then Adobe Firefly, with a C+ and a B+. Flux.3, from Black Forest Labs, drops to a B and a C. Microsoft's and Meta's stay at a D+ for usefulness. For Sora 2, OpenAI's video model, there is no document at all, nor for the earlier versions of Flux, which are the ones half the sector has installed. Midjourney is nowhere to be found.

I want to share some numbers I would not call anecdotal. Those documents, in which each manufacturer must state what it trained its engine with, are scored by section: public sources, private sources, web crawling, user data, synthetic data and processing. Well, in the public sources section Gemini 3 Pro scores 6 out of 100 for clarity and 0 for usefulness. Grok 4.5, 7 and 0. Phi-4, from Microsoft, stays at 3 and 0. That zero is not a low mark: it means that, with that document in front of them, an author can do nothing. Neither find out whether their work is inside, nor ask for it to be taken out.

Well, giving them the benefit of the doubt, that lack of information could be explained by the fact that models already on the market before 2 August 2025 have until August 2027 to comply.

Distilling this information and bringing it down to earth: if not even the European Commission has this data, a studio is hardly going to have it.

What we do answer for is our own. Offering quality material sustains the visualisation work and, in turn, the client's project. The famous seals of guarantee, the very ones meat ended up needing.

From cattle we learned too late what was in the feed. With this we should still be in time.

Web references

New Scientist, Brain disease drives cows wild, 5 November 1987 (the piece contemporary with the British Ministry of Agriculture's admission). New Scientist

CDC, Clinical Overview of Variant Creutzfeldt-Jakob Disease: 233 vCJD deaths worldwide between 1996 and 2023. CDC

BBC News, 2 December 2024: 4.4 million cattle slaughtered in the United Kingdom. BBC

Regulation (EC) No 178/2002, establishing EFSA. EUR-Lex

Alemohammad, S., Casco-Rodriguez, J., Baraniuk, R. G. et al. (Rice University), Self-Consuming Generative Models Go MAD, arXiv:2307.01850 (2023), ICLR 2024. arXiv

Shumailov, I., Shumaylov, Z., Zhao, Y., Papernot, N., Anderson, R., Gal, Y., AI models collapse when trained on recursively generated data, Nature 631, 755-759 (24 July 2024). Nature

Gibney, E., AI models fed AI-generated data quickly spew nonsense, Nature News, 24 July 2024. Nature

Gerstgrasser, M. et al., Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data, arXiv:2404.01413. arXiv

Dohmatob, E. et al., Strong Model Collapse, arXiv:2410.04840. arXiv

Regulation (EU) 2024/1689, articles 53(1)(d) and 111(3). AI Act Service Desk

European Commission, template and explanatory notice for the public summary, 24 July 2025 (library). digital-strategy.ec.europa.eu

Blankvoort, D. A. H., Pandit, H. J. and Gahntz, M. (2026), Quality Assessment of Public Summary of Training Content for GPAI models required by AI Act Article 53(1)(d), FAccT 2026, DOI 10.1145/3805689.3806755. arXiv

AI Accountability Lab, Trinity College Dublin, GPAI Training Transparency (accessed 2 September 2026). aial.ie