kapynResearch

SocialReasoning-Bench: Measuring whether AI agents act in users’ best interests

SocialReasoning-Bench measures AI agent alignment with user interests. Agents perform tasks competently but often fail to optimize for user benefit, even when explicitly instructed. This highlights a critical gap in AI safety and alignment research.

Microsoft Research·May 11, 2026

Opening Kapyn…