AI and Humans Diverge in Aesthetic Evaluation of Beauty
A new study on arXiv (2605.18759) compares how humans and large language models (LLMs) evaluate beauty in visual scenes. Researchers used questionnaire items from prior human studies on beauty ratings, bodily sensations, and emotions, and presented them to LLMs for direct comparison. The analysis reveals divergences in aesthetic evaluation between humans and AI, highlighting differences in sensibility beyond cognitive tasks. The work is part of ongoing AI alignment efforts, probing convergence or divergence in aesthetic experiences.
Key facts
- Study compares human and LLM aesthetic evaluations
- Uses questionnaire items on beauty, bodily sensations, emotions
- LLMs are approaching or surpassing human performance in cognitive tasks
- Aesthetic evaluation depends on internal processes of the observer
- Research is part of AI alignment efforts
- arXiv paper ID: 2605.18759
- Published on arXiv
- Focus on visual scene beauty perception
Entities
Institutions
- arXiv