Model collapse
noun
What happens when AI trains on AI output: each generation learns from the last one’s mistakes and flattens toward a confident average. The Flattening chapter calls it cannibalism, which is ruder and more accurate.
Also known as AI cannibalism
Model collapse is what can happen when generative AI is trained again and again on AI-generated material and begins to lose the rare, peculiar edges of the original human data. Common patterns get louder, unusual details fade, and eventually the great buffet of language starts serving fourteen trays of the same beige pudding. The book’s ruder name is AI cannibalism: the machine eating its own output and calling the smaller menu knowledge. It is not an unavoidable curse in every training design; keeping and accumulating real data changes the outcome, which is exactly why human-originated material matters. The strange sentence, the minority pattern, the awkward regional detail are not clutter. They are what averages erase first, and then the photocopy gets photocopied again.
See also AI slop Enshittification Workslop
References
- Human Still Wins (2026), Chapter 7 — The Flattening