AI agents risk 'knowledge collapse' — new paper warns of cognitive commons tragedy

In a recent paper, Kallel and El Louadi (2026) scrutinize the dynamic model by Acemoglu, Kong and Ozdaglar (2026a) that warns of 'knowledge collapse' — a self-reinforcing deterioration of humanity's common knowledge base driven by agentic AI. The original model posits a complementarity between general human knowledge and context-specific local knowledge, with a learning externality where everyone contributes to a thin public signal. If AI substitutes for private human effort but cannot regenerate the public signal, and if human effort is sufficiently elastic, the system can tip into a low-knowledge equilibrium. The authors provide a measured appraisal, noting partial empirical evidence like a 25% decline in public knowledge sharing on Stack Overflow, and connect the model to literature on knowledge commons and model collapse.

Kallel and El Louadi then present five structural criticisms: the model's fixed knowledge taxonomy, the assumption that AI cannot generate general knowledge, the unmeasured effort elasticity parameter, and its political futility. They argue the model's true value is not in predicting certain catastrophe but in identifying a credible negative externality — the 'tragedy of the cognitive commons.' The remedy lies in better aggregation of human knowledge. The paper concludes with a research agenda focused on measuring effort elasticity and designing institutional mechanisms to preserve the collective intelligence commons in an age of pervasive AI.