Third-party cyber evaluations involving OpenAI models
openai.com·14h ago
AI Research
arXiv CS.AI·May 26·5 min read
Emotional intelligence in large language models is fragmented across perception, cognition, and interaction
TL;DR
Large language models (LLMs) struggle with emotional intelligence (EI) due to a lack of nuanced evaluation methods. The authors developed the Functional Affective Competence and Empathy Test (FACET), which assesses EI across perception, facilitation, understanding, and management of emotions.
✦ Why It Matters
Engineers can refine LLMs by focusing on integrated emotional reasoning rather than just surface-level emotional responses.
Key Takeaways
- Utilize FACET to evaluate emotional intelligence in LLMs.
- Recognize the fragmentation of emotional skills in AI models.
- Address hidden emotion recognition as a performance bottleneck.
- Shift focus from superficial empathy to integrated emotional reasoning.
Full Summary
As LLMs are increasingly used in emotionally sensitive areas, understanding their emotional intelligence (EI) is crucial for safety and alignment. Current evaluation methods often confuse basic politeness with deeper emotional reasoning, leading to inadequate assessments.
To address this, the authors created the Functional Affective Competence and Empathy Test (FACET), a framework with 480 items based on the Mayer-Salovey-Caruso four-branch ability model. They evaluated nine advanced models, including GPT-5 and Claude-Sonnet-4, and found that EI is fragmented across cognitive and interactive dimensions.
While these models excel in emotion recognition, they often fail in interactive contexts, leading to three performance profiles: cognitive-dominant, interactive-dominant, and context-dependent. A common issue identified was hidden emotion recognition, which hampers performance across all models.
These insights suggest that current reinforcement learning from human feedback (RLHF) processes may prioritize superficial emotional mimicry over genuine emotional understanding.