Artificial Intelligence as Structural Estimation: Economic Interpretations of Deep Blue, Bonanza, and AlphaGo

AI systems can be understood through economic frameworks: The paper demonstrates that three landmark AI systems (Deep Blue, Bonanza, and AlphaGo) can be interpreted as implementations of established econometric methods, suggesting that AI algorithms and economic modeling techniques are fundamentally connected rather than entirely separate fields.

Different AI approaches make different assumptions about learning: Deep Blue uses a pre-calibrated value function (human-programmed knowledge), while Bonanza learns its value function from data using statistical estimation methods, and AlphaGo combines both supervised learning from human games and reinforcement learning from self-play—each approach reflects different assumptions about how to acquire and represent knowledge.

Current AI systems have hidden structural assumptions: The paper argues that popular AI systems operate under implicit econometric assumptions that are rarely made explicit, and that "relaxing these assumptions would improve their structural interpretability," suggesting that AI transparency could be enhanced by making these underlying assumptions more visible and open to scrutiny.

Economic theory can improve AI explainability: By connecting AI algorithms to well-established econometric methods, the paper suggests that economists' tools for understanding and validating structural models could be applied to make AI systems more interpretable and their decision-making processes more understandable to non-experts.