Large language models are proficient in solving and creating emotional intelligence tests.

Area of Science:

• Artificial Intelligence
• Psychology
• Cognitive Science

Background:

• Large Language Models (LLMs) exhibit broad expertise but their emotional intelligence capabilities are largely unexplored.
• Assessing emotional intelligence in AI is crucial for understanding human-AI interaction and AI's role in sensitive applications.

Purpose of the Study:

• To investigate if LLMs can accurately solve performance-based emotional intelligence tests.
• To evaluate the capability of LLMs to generate novel emotional intelligence test items.

Main Methods:

• Five LLMs (ChatGPT-4, ChatGPT-o1, Gemini 1.5 flash, Copilot 365, Claude 3.5 Haiku, DeepSeek V3) were tested on standard emotional intelligence assessments.
• ChatGPT-4 generated new items for these tests, which were then administered to 467 human participants across five studies.

Main Results:

• LLMs achieved an average accuracy of 81% on emotional intelligence tests, significantly outperforming the human average of 56%.
• LLM-generated test items showed statistically equivalent difficulty to original tests.
• While some psychometric properties differed slightly between original and generated tests, effect sizes were small, indicating strong correlations.

Conclusions:

• LLMs demonstrate a robust capacity for understanding and responding to emotional intelligence assessments.
• LLMs can generate emotionally relevant test content that is comparable in difficulty to human-created tests.
• These findings suggest LLMs possess accurate knowledge of human emotions and their regulation.