NBER paper says AI should be trained for human decision-making, not raw accuracy
Afinetheorem · x · 2026-07-24
A thread from NBER highlights a paper arguing that AI should be trained and benchmarked in the context of how humans actually use its predictions. The paper treats AI as part of a “composite experiment” in which machine forecasts are combined with human verification, judgment, and other models, and it argues that maximizing raw prediction accuracy is often suboptimal economically.
Key points from the paper preview:
- AI’s value depends on the surrounding decision environment, not just model accuracy.
- Optimal training can be discontinuous in economic variables.
- Heterogeneous users and monopoly trainers change what the best objective looks like.
- The paper reframes training around human decision use rather than isolated prediction quality.
More from AGI Musings
- A post predicts the rise of a future “Chief of Chat” job — joshwhiton · 2026-07-24
- Emad Mostaque says open-weights debates are really US-vs-China debates — emollick · 2026-07-24
- If LLMs solve existence but fail universal claims, math academia may barely change — JFPuget · 2026-07-24
- If an AI knows which memories you regret, can it use them as leverage? — PierceLilholt · 2026-07-24
- Allie Miller: AI is Moving Towards Multiplayer, Proactive, and Autonomous Agents — alliekmiller · 2026-07-24
- Gary Marcus’s 2022 AI predictions already look outdated by 2026 — dhadfieldmenell · 2026-07-24