A theoretical model argues that training an AI solely for unconditional predictive accuracy can be suboptimal once human verification and downstream decisions are considered.
Research question
How should predictive models be trained when their outputs are combined with human analysis and other information before a decision?
Method & data
A theoretical framework treats AI prediction as one component of a broader composite experiment and studies heterogeneous users and monopoly model trainers.
Main finding
The abstract states that economically optimal training can differ discontinuously from accuracy-maximizing training as the surrounding decision environment changes.
Why it matters
Model evaluation should reflect how predictions are actually used, not only unconditional benchmark accuracy.
What to verify
The public abstract does not include an empirical application or quantify the size of the gap between training objectives.
Show public abstract used for this summary
AI predicts; humans use its predictions to make decisions. These predictions are combined with human verification and analysis, queries to other statistical models, and so on. The economic value of an AI, therefore, depends on how it interacts with the surrounding decision environment. We describe the value of AI as part of this "composite experiment" where AI makes a coarse prediction of the state of the world, show what this means for optimal model training via a geometric argument, explain why optimal training can be discontinuous in economic variables, and study how heterogeneous users or monopoly model trainers affect these results. In particular, maximizing the unconditional accuracy of AI predictions is generally suboptimal.