When this happens to an AI, the model's output drifts toward a bland, confident average. Push it far enough, and researchers have shown models eventually degrade into repetitive nonsense. The phenomenon was formally identified by a team from the universities of Oxford and Cambridge, whose landmark study was published in the journal Nature in 2024. Their warning was blunt: train AI indiscriminately on AI-made content, and you risk a slow-motion breakdown in its ability to produce diverse, high-quality results.
AI is destroying itself with 'data cannibalism' — but there's a simple fixTo show you the most relevant results, we’ve omitted some entries very similar to those already shown. Repeat the search with the omitted results included.