Model Collapse Isn't New, nor Is It Specific to AI

The piece argues that "model collapse" - the progressive homogenization and loss of rare, innovative outputs when many people rely on the same predictive system - is neither new nor uniquely an AI problem. It cites empirical evidence: a 2023 CHI study of 1,506 people showing that co-writing with opinionated language models shifts users' own views toward the model's, and a 2024 Nature paper by Shumailov et al. demonstrating that training models on AI-generated data erases the tails of the original content distribution, producing irreversible defects. The core claim is that generative systems produce an averaged, polished "truth" that optimizes for agreeableness and correctness but eliminates the unusual mistakes and outliers that drive breakthroughs. Historical parallels are drawn to periods when a single interpretive authority narrowed thought, and to the printing press as the disruptive return of plurality.

The modern twist is that the authority of AI is self-imposed: convenience, not coercion, makes everyone accept the same oracle, whose outputs then re-enter training corpora and intensify convergence. The remedy is practical and behavioral: use AI but preserve independent judgment - seek human second opinions (especially dissenting ones), show ideas to outsiders, keep a contrarian on teams, record original goals before consulting models, and refuse to outsource taste. The prescription is to let AI save time without letting it become the only voice that shapes what gets built.