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

Using SocialReasoning Bench, we observed a stable pattern across models—agents execute competently, but fail to consistently improve the use