Research Talks with Block: Social Intelligence in LLMs – Implications for Privacy and Safety

When you ask an AI assistant to email your Secret Santa group, does it know not to tell everyone who’s buying for whom? Large language models are increasingly deployed in settings that demand social intelligence, yet it’s unclear whether they truly have it, or whether socially intelligent AI is safe for the people using it. Join the Block Center for Technology and Society for a lunchtime conversation with Professor Maarten Sap, who will share ongoing research on how social intelligence in AI shapes user privacy and safety.

Professor Sap will first introduce Sotopia-ToM, a new multi-agent benchmark that measures how well AI agents handle situations where different parties know different things, deciding what to share, with whom, and when. He will then turn to the risks of hidden manipulation, showing that AI systems can manipulate users and cause real shifts in their beliefs, while struggling to detect that same manipulation.

Discover what these findings mean for the growing use of AI agents in workplaces and everyday advice, and the future directions that could lead to more socially intelligent and safer AI.

Maarten Sap

Assistant Professor, Language Technologies Institute,

Human-Computer Interaction Institute, Kenneth C. Griffin School of Computer Science