Objective:

To evaluate the ability of three leading LLMs to recognize human facial expressions using the NimStim dataset, highlighting the significance of their socioemotional competence.

Key Findings:

• GPT-4o and Gemini 2.0 Experimental matched or exceeded human performance, particularly for calm/neutral and surprise expressions, with GPT-4o achieving the highest overall accuracy.
• Overall accuracy was 86% for GPT-4o, 84% for Gemini 2.0, and 74% for Claude 3.5, indicating a clear performance hierarchy.
• Fear was frequently misclassified as surprise across models, highlighting a common area of error.

Interpretation:

The findings indicate that LLMs are developing socioemotional competence, with potential applications in healthcare for recognizing mental health conditions such as depression and anxiety.

Limitations:

• All stimuli were static images, limiting generalizability to dynamic expressions.
• Actors were predominantly aged 21-30 and European American, which may affect results and applicability to diverse populations.
• The study relied on a single dataset, which may limit broader applicability and necessitates further validation across varied datasets.

Conclusion:

While LLMs show promise in facial expression recognition, further research is needed to enhance generalizability and explore multimodal emotion classification, particularly in diverse contexts.