AI models can excel in many fields, but their ability to understand emotions is still unclear.
Researchers from the University of Geneva and the University of Bern tested six AIs, including ChatGPT, using human emotional intelligence (EI) assessments. Surprisingly, the AIs performed better than the average human and even quickly created new tests, suggesting exciting AI potential in education, coaching, and conflict resolution.
Researchers tested six AI models on emotional intelligence using five well-known assessments. These tests, often used in research and business, included emotionally charged scenarios to measure how well AI understands, regulates, and manages emotions.
For example, what would be Michael’s most effective reaction if one of Michael’s colleagues stole his idea and was unfairly congratulated?
a) Argue with the colleague involved
b) Talk to his superior about the situation
c) Silently resents his colleague
d) Steal an idea back
This example illustrates how emotional intelligence assessments evaluate decision-making in challenging situations. Michael’s most effective reaction is option b)—talking to his superior—because it addresses the problem professionally and fairly, rather than reacting impulsively or seeking revenge.
Researchers gave the same emotional intelligence tests to humans and AI models. The AIs scored much higher—82% correct compared to 56% for humans—suggesting they understand emotions and respond intelligently.
Later, scientists had ChatGPT-4 create new tests, which were then taken by over 400 people, further exploring AI’s ability to assess emotional intelligence.
The AI-generated emotional intelligence tests were just as reliable, clear, and realistic as the original ones, which had taken years to develop. This suggests that LLMs can choose the best response in emotional scenarios and create new, context-specific tests.
These findings reinforce that LLMs, like ChatGPT, possess emotional knowledge and can effectively reason about emotions—a promising step toward AI-assisted education, coaching, and decision-making.
Journal Reference:
- Schlegel, K., Sommer, N.R. & Mortillaro, M. Large language models are proficient in solving and creating emotional intelligence tests. Commun Psychol 3, 80 (2025). DOI: 10.1038/s44271-025-00258-x
