AI Scientists Need a Theory of Scientific Value

Abstract

An AI scientist must decide which questions deserve an answer and what would count as an advance. Current systems learn these judgments from scientific work, expert evaluations, and task-specific measures of success. As they take greater responsibility for choosing research, the grounds for those judgments become part of the design problem. We argue that AI for science should make the development of a theory of scientific value a foundational research task. Such a theory should explain why a contribution is worth recognizing and support judgments that can be checked against explicit principles and evidence. New relationships and stronger grounds for existing conclusions offer concrete starting points. They also reveal why value depends on what is already known, why some contributions resist a common scale, and why scientific advances can demand new standards. Developing this theory would make the aims of AI research open to systematic justification, formal investigation, and empirical challenge.

Author's revised preprint, September 8, 2026. The repository record preserves the original deposited version.