Model Collapse Is Not New, and It Is Not Unique to AI

Summary

The article argues that the most important risk of AI-generated content may be intellectual homogenization rather than simple reuse of existing code. When an AI assistant proposes a product’s features, identity, pricing and launch plan, users can feel like architects while mainly choosing from options shaped by someone else. A study of 1,506 people found that participants who wrote with an opinionated language-model assistant shifted their own views toward the assistant’s opinions. The author connects this social effect to model collapse: a 2024 Nature study found that training models on recursively generated data causes irreversible defects in which the tails of the original distribution disappear. Those tails represent rare, unusual material that may be especially important for discovery and progress. The article compares this dynamic with historical monopolies over interpretation, arguing that the printing press restored disagreement and variation by putting texts in more hands. AI differs because people adopt its authority voluntarily for convenience rather than under institutional coercion, which can make a shared answer especially difficult to challenge. The author does not call for abandoning AI. Instead, the article recommends seeking human and dissenting opinions, consulting people outside one’s field, protecting the original goal before prompting, and treating AI advice as input rather than a substitute for personal taste and judgment.