How AI Aggregation Affects Knowledge

A team of prominent economists, including MIT's Daron Acemoglu, has published a groundbreaking paper titled "How AI Aggregation Affects Knowledge" on arXiv. The research introduces a formal economic model to analyze a critical feedback loop in modern AI: when AI systems aggregate human beliefs and outputs, then use that synthesized data to train future models, potentially degrading the quality of collective knowledge over time. The authors extend the classic DeGroot model of social learning by introducing an "AI aggregator" that trains on population beliefs and feeds signals back to agents, defining a "learning gap" to measure deviation from an efficient benchmark.

The study's central finding identifies a dangerous threshold related to update speed. When an AI aggregator updates its training on population data too quickly, the researchers prove that no set of training weights can robustly improve learning across environments. In contrast, slower-updating systems can find weights that help. This leads to a crucial architectural comparison: the paper demonstrates that local aggregators—trained on proximate or topic-specific data—robustly improve learning in all environments. Consequently, the research concludes that replacing specialized local AI models with a single, global aggregator (a common industry trend) will worsen learning outcomes in at least one dimension of the state space, creating a trade-off between breadth and depth of knowledge.