Abstract:
We propose a computational benchmarking approach to evaluate human decisions in contexts where AI often outperforms humans. By comparing choices to a “strong” AI (near-optimal) and a “restricted” AI (human-like computational limits without biases), we separate cognitive constraints from behavioral factors. Using professional chess data, we show that time pressure, fatigue, and complexity shape decisions, and that faster choices often deviate from the benchmark yet can still improve quality.
