AI, Human Cognition and Knowledge Collapse

Acemoglu, Kong, and Ozdaglar do something rare: they build a model with genuine generative architecture — not a metaphor about AI deskilling, but a formal dynamic system showing precisely when and why agentic AI tips a society into a knowledge-collapse equilibrium.

The core mechanism is a learning externality: human effort jointly produces private context-specific knowledge and a thin public signal that accumulates into shared general knowledge; when AI substitutes for that effort, it quietly drains the commons that makes human effort valuable in the first place.

The welfare non-monotonicity result — that more accurate AI can make society worse off, implying an interior optimal precision — is the kind of counter-intuitive formal finding that reshapes how product directors should think about when to automate and when to preserve human deliberation.

The policy implication (aggregation capacity for general knowledge unambiguously raises welfare) connects directly to platform design and governance questions that the library identifies as gap territory.

This is Acemoglu at his most relevant to the library's project: not labor-market displacement framing, but an epistemic commons argument with transferable architecture.

Read alongside Ostrom on governing the commons and Brynjolfsson on the productivity paradox.