Recent advancements in Artificial Intelligence have produced general-purpose Large Language Models (as used in AI chatbots such as chatGPT) that millions of people use daily. These models not only produce text that is startlingly like human writing and spoken language, but have shown remarkable ability to understand people’s emotions and even offer empathy. But we are unable to objectively gauge their true understanding of emotions, because we do not have rigorous definitions and tests that measure social and emotional understanding in AI. Without objective measures, we cannot assess progress in developing these models, identify blind spots or potential harms, or understand how to improve the emotional understanding of AI. This project addresses this knowledge gap, by defining what such social and emotional understanding entails, and devising new ways to measure the emotional capacity of AI language models. In parallel to the technical work, this project will also produce policy guidelines for designing and implementing AI that relies on understanding of human emotions, and educational materials to provide teenagers and adults with better AI literacy, and enable them to navigate the AI landscape with an awareness of its dangers. This project will ultimately enable practitioners to build more capable and safer conversational AI conversational that can better understand their human users. The project combines theoretical frameworks from psychology with rigorous methods from