What is AI Cannibalism, and can it be fixed?

Dubai: AI models trained on AI generated content can degrade over time, a problem called model collapse. Here is what it means, why it matters, and a surprising new fix, in easy to understand terms.

Type a question into an AI search tool today, and there is a decent chance part of the answer was written by another AI, not a person. That is becoming a problem. As more of the internet fills up with AI generated text, images and analysis, AI companies increasingly end up training their next models on that same AI generated content. Researchers have a blunt name for it: AI cannibalism.

What AI cannibalism actually means

Think of it like making a photocopy of a photocopy. Each generation loses a little more detail, until what is left barely resembles the original. AI cannibalism works the same way. A model trained mostly on real, human made data learns the full range of how people write, think and disagree. But when a new model is trained largely on the output of an older model rather than on fresh human data, it inherits that older model's blind spots and mistakes, then adds a few of its own. It is an old computing idea with a new name, coined by Steven J. Vaughan-Nichols, GIGO : garbage in, garbage out.

Do this across several generations, and researchers call the result model collapse: models that grow steadily less accurate, less original and more repetitive, because they are increasingly working from a copy of a copy of a copy.

A widely cited 2023 study on the subject, "The Curse of Recursion," found that models trained this way begin to forget the rare, unusual edges of human knowledge and drift toward a bland, averaged out version of reality. One 2024 paper in Nature put it more bluntly still: the model becomes poisoned with its own projection of reality.

Why it is happening now

This is not a distant, theoretical risk. OpenAI chief executive Sam Altman said in 2024 that the company was generating around 100 billion words a day. A large share of that output ends up published somewhere online, which means the next generation of AI models, trained on fresh scrapes of the internet, are increasingly learning from text that earlier AI systems wrote.

Researchers point to a few distinct ways this plays out. Errors made by one model generation get inherited and amplified by the next. Rare, unusual information gets quietly erased, since AI generated text tends to gravitate toward the most common, predictable answer rather than the full range of what is actually true. And feedback loops set in, where a model's own biases get reinforced rather than corrected, generation after generation.

The practical effects are already showing up. Technology writers have described noticing AI search tools increasingly returning figures that only loosely resemble the real financial filings they are meant to summarise. A 2025 Bloomberg study of retrieval augmented generation, a technique meant to keep AI models grounded in real, current information, found that it can backfire too, in some cases making leading models more likely to leak private data or produce misleading analysis, not less.

Is there a fix to this?

Here is where the story gets more hopeful. Researchers from King's College London, the Norwegian University of Science and Technology and the Abdus Salam International Centre for Theoretical Physics recently tested how easily model collapse can be prevented, using a simpler, more mathematically transparent class of models than a full scale AI system.

What they found: adding just one real, human generated data point back into the training process was enough to prevent collapse entirely, even when synthetic, AI generated data vastly outnumbered it.

"Computer scientists will have the tools to prevent this potentially disastrous scenario," said Professor Yasser Roudi of King's College London, who worked on the study.

The finding, published in Physical Review Letters, does not mean the problem is solved. The researchers tested it on simpler statistical models, not the vast, complex systems behind tools like ChatGPT or Gemini, and they are still working out whether the same principle holds at that scale. But it does suggest the fix might not require an impossible amount of fresh human data, just a consistent, deliberate trickle of it.

What is actually being done about it

In practice, AI companies are already leaning on a mix of approaches. Retrieval augmented generation lets models pull in fresh, external information rather than relying purely on what they memorised during training, though as the Bloomberg findings show, it is not a complete fix on its own. Companies are also investing more in data curation, the unglamorous work of cleaning, labelling and tracking where training data actually came from, so AI generated content can be filtered out or balanced against real human material. Regulation is starting to catch up too, with the European Union's AI Act introducing transparency requirements that touch on data quality, and early, similar efforts now under discussion in the US.

None of this makes AI cannibalism a solved problem. But it does suggest the future is less about avoiding synthetic data altogether, and more about making sure a real, human voice never entirely disappears from the mix.

And in my opinion, AI may be able to cannibalise itself but it can never eat or erase human language.

Sources

• Kerner, S.M., "AI cannibalism explained: A model failure," TechTarget

• Vaughan-Nichols, S.J., "Some signs of AI model collapse begin to reveal themselves," The Register, May 2025

• Prada, L., "AI Models Are Cannibalizing Each Other, and It Might Destroy Them," June 2025

• King's College London, Norwegian University of Science and Technology and Abdus Salam ICTP, study on preventing model collapse, published in Physical Review Letters

• Shumailov, I. et al., "The Curse of Recursion: Training on Generated Data Makes Models Forget," 2023

• Citizen Code, "The AI paradox of cannibalism"