Paper: Optimal AI Training for Human-AI Decision Making

Afinetheorem · x · 2026-07-10

Economists released a new paper exploring the core factors determining AI value. The research highlights that an AI's value depends on human usage—whether humans trust, verify, or ignore its predictions—forming a "composite experiment" in decision theory. The paper models AI capability as a trade-off boundary between "coverage" (how often it attempts to predict) and "accuracy." Due to varying human verification costs and error penalties, the optimal AI behavior is discontinuous: if verification is cheap, the AI should always attempt to predict; if errors are costly, it should only speak up when highly confident. Thus, the optimal AI training method entirely depends on the specific human user and their decision parameters.

Related event: Study: AI Value Hinges on Human Collaboration, Not Just Accuracy(8 posts)→

Original post →

More from AGI Musings

AGI Musings channel →