SocialReasoning-Bench: Measuring whether AI agents act in users’ best interests
Microsoft Research introduced SocialReasoning-Bench, a benchmark for testing whether AI agents act in users' best interests. It measures both outcomes and process, adding a concrete evaluation signal for agentic AI systems as they move into higher-stakes workflows.