AI Alignment Breaks at the Edge

In three linesarXiv paper showing AI alignment fails on edge cases: value conflicts, multi-stakeholder disagreement, epistemic ambiguity. Scalar rewards and average-case evaluation hide these failures. Authors propose 'Edge alignment': detection, evaluation, and governance to surface critical cases. Tested on 91 edge cases across 4 contemporary models.

Summary generated by Claude — human-verified