AI's Use of Knowledge in Society
The title is a direct nod to Hayek (1945).
Brynjolfsson and Hitzig argue that AI can shift the optimal locus of control in organisations through two channels: by codifying local knowledge that used to be tacit, and by expanding the processing capacity needed to aggregate and interpret dispersed information.
This erodes the advantage of having people deciding on the spot, making centralised coordination more viable.
But they recognise the limits: AI fails on rare scenarios (the "long tail"), on embodied knowledge, and at the same time it empowers the periphery by giving it access to knowledge that used to live only at the centre.
The tension between centralising and decentralising is not resolved — it intensifies.
Central argument
Brynjolfsson and Hitzig argue that AI reshapes the optimal locus of control within organizations through two distinct mechanisms: it codifies previously tacit local knowledge, and it expands the capacity to aggregate and process dispersed information. Both effects make centralized coordination more viable than Hayek's original analysis would permit — the classic argument that only the person on the spot can hold the relevant knowledge loses force when AI can capture and transmit that knowledge upward. However, the authors resist a clean centralizing conclusion: AI simultaneously empowers the periphery by democratizing access to knowledge that once resided only at the center, and it fails systematically at the long tail of rare scenarios and embodied knowledge, preserving irreducible advantages for local human judgment.
Critique
The framework treats organizational structure as primarily an information problem, which risks underweighting the political and incentive dynamics that shape how AI-generated knowledge is actually used. Even if AI can technically codify local knowledge and expand central processing capacity, those at the periphery have well-documented reasons to filter, delay, or distort what they surface — and centralized actors have their own motivated reasoning in how they interpret aggregated signals. The model of a neutral information conduit may overstate what AI changes in organizations where power and knowledge are already entangled.
Why it matters for product
For a CPO, the central tension in the essay maps directly onto the perennial question of how much product autonomy to grant to embedded teams versus how much to coordinate centrally through strategy, data, and platform. If AI genuinely erodes the informational advantage of the person closest to the user — by codifying discovery insights, surfacing usage patterns, or synthesizing qualitative feedback at scale — it weakens the traditional justification for highly decentralized squad models. At the same time, the authors' 'long tail' caveat suggests that edge cases, novel markets, and embodied user knowledge will remain the irreducible domain of local product judgment, which argues against collapsing that autonomy entirely.