AI Models Can Now 'Breed' — And That Reveals Why They Go Bad

Something odd is happening in the AI industry. Instead of always training new AI from scratch on human writing, companies now feed models each other's output, retrain them on their own past work, and glue several models together into one. That looks less like engineering and more like a family tree — parents, children, inherited traits. So a researcher at Imperial College London, Giorgio Gilestro, asked a simple question: what if we just used the math biologists already built for animal populations?

The answer, it turns out, is yes — almost exactly. He found that a model retrained only on its own output follows the same equation biologists use to track shrinking populations. That matters because it explains a problem AI companies quietly worry about: models fed too much of their own writing get dumber, repeating themselves and making more mistakes. The fix is like adding fresh genes. Real human-written text acts like new arrivals joining a village, and the study found something surprising — what counts is the raw amount of real data, not what percentage of the mix it is. A little fresh data diluted in a huge pile of AI text still helps, as long as there's enough of it.

The most useful discovery is about merging models. If you combine two AI models by averaging them — splitting the difference — you lose the advantage of having two parents at all. That's the same old argument biologists used against Darwin in the 1800s. But if you merge them so each keeps its strongest talent, the child outperforms both parents. He repeated this across random runs, and it held every time.

There's a darker finding too. Models that learn clashing rules — different ways of formatting things, say — become unable to merge at all. It's not about growing apart over time; it's specifically about learning contradictory habits. As AI systems start having generations of their own, knowing which ones can still mix matters for anyone using them.