What Experts Learn from AI: Evidence from Professional Go
Abstract
Whether experts can learn from artificial intelligence is central to how AI will shape human capital. Using 4.58 million professional Go decisions graded by AI, I exploit staggered adoption of AI-based training, dated by each player’s first AI-popularized opening move. Adoption reduces the points players lose by 13 percent. Beyond the opening, players improve without further agreeing with AI. Gains do not differ measurably with prior ability and extend to positions whose best move lacked professional precedent. Performance gaps widened by half as older professionals fell behind. Learning from AI thus improves even experts’ judgment without narrowing skill gaps.