How AI Aggregation Affects Knowledge

Summary

Extends the DeGroot social-learning model with an AI aggregator that trains on population beliefs and feeds synthesized signals back to agents, defining a learning gap as the deviation of long-run beliefs from the efficient benchmark. The main result is a threshold in updating speed - if the aggregator updates too quickly, no positive-measure set of training weights robustly improves learning, whereas such weights exist when updating is slow enough - and the architectural comparison shows local aggregators trained on proximate or topic-specific data robustly improve learning in all environments, so replacing specialized local aggregators with one global aggregator worsens learning in at least one dimension of the state. For a paper claiming that aggregation architecture determines epistemic quality, this is the closest formal economics result: global versus local aggregation is a hierarchy-versus-polyarchy claim in disguise, from an authoritative team, and it is missing from the knowledge base.

Contribution

Queued — not yet read.

Method

Queued — not yet read.

Relevance to RISE

Surfaced by the 2026-09-07 literature sweep as a high-priority candidate for the Architecture as Epistemology (ISR) paper (angle: organization theory applied to AI agents; epistemic quality of LLM outputs). Verified against: Verified by opening https://www.nber.org/papers/w35036 on 2026-09-07: full abstract read verbatim, 4 authors, NBER Working Paper 35036, issue date April 2026. Pre-window; flagged as centrally missing.

Critique / open questions

Queued — not yet assessed.

Key quotes

—