Part of a series “old econ papers”
The Use of Knowledge in Society is usually summarized as “central planning doesn’t work because markets are better”. A lot of economically useful knowledge is local, temporary and difficult to write down:
- this machine started making a strange noise this morning;
- this customer will pay more if we deliver two days earlier;
- this supplier is unreliable despite looking good in the database;
- everybody at the lab knows this protocol needs ten more seconds than the documentation says. Prices are useful because they compress some of that distributed information into a number that lets people coordinate without understanding the entire system.
AI makes the center much smarter
Hayek wrote in a world where moving local knowledge into a central system was expensive, and processing it was even more expensive.
Erik Brynjolfsson and Zoë Hitzig’s “AI’s Use of Knowledge in Society”: if AI can codify more local knowledge and process much more of it, some decisions that previously had to be decentralized can move upward.
But AI also makes the edge much smarter
Give the same model to a junior employee and they suddenly have access to knowledge that previously required asking a senior specialist. In a field study of customer-support agents, an AI assistant increased average productivity, with larger improvements for less experienced and lower-performing workers.
I think AI attacks Hayek’s computation problem more than his discovery problem
Suppose an AI really can ingest every relevant fact about my company — it still doesn’t know what would happen if we tried a completely different product, organization or scientific hypothesis. That information does not exist yet; it has to be created by experimentation.
Company A tries one product, Company B tries another, and Company C does something stupid that unexpectedly works. Customers choose, most experiments fail, and the resulting information is generated by the competitive process itself: markets don’t only aggregate knowledge, they also run lots of partially independent experiments and kill the bad ones.
AI can simulate more possibilities, but the world still has to answer many questions. This is especially obvious in science — I can ask an AI scientist to generate 100,000 plausible hypotheses, but eventually some subset needs a real experiment. From my Two ICML 2026 AI For Science Papers: physical execution and verification are still expensive.
This changes how I think about AI organizations
A common prediction is that AI will flatten companies because every worker can do more. Another is that it will create giant centralized organizations because one management layer can see and control everything. Both are plausible, and the right question is probably:
For which decisions does AI reduce the value of local knowledge more than it increases local capability?
Customer support may become more centralized, while small product experiments become more decentralized. Research may split: centralized models and compute, but thousands of autonomous experiments at the edge. Governments could centralize administrative processing while still needing local experimentation — the shape of organizations should change differently depending on which kind of knowledge is scarce.