Third-party cyber evaluations involving OpenAI models
openai.com·13h ago
AI Research
arXiv CS.AI·May 21·5 min read
Interoceptive Divergence in Aesthetic Evaluation and Implications for Human-AI Alignment
TL;DR
A gap exists in understanding how humans and AI systems, particularly large language models (LLMs), evaluate aesthetic beauty. The study explores interoceptive divergence, which refers to differences in internal bodily perceptions, in aesthetic evaluations between humans and AI.
✦ Why It Matters
Engineers can enhance AI systems by incorporating human-like aesthetic sensibilities to improve alignment with user preferences.
Key Takeaways
- Explore how AI perceives beauty differently than humans.
- Incorporate emotional responses into AI aesthetic evaluations.
- Consider interoceptive divergence in AI design.
- Align AI systems more closely with human values.
Full Summary
As AI, especially large language models (LLMs), increasingly matches or exceeds human cognitive abilities, understanding the nuances of human experience, such as aesthetic evaluation, becomes crucial. This research investigates interoceptive divergence, which is the difference in internal bodily awareness, in how humans and AI assess beauty in visual scenes.
Using a series of experiments, the study measured participants' aesthetic judgments alongside AI-generated evaluations. Findings reveal that humans and AI often diverge significantly in their aesthetic preferences, with humans showing a broader range of emotional responses.
These results suggest that while AI can mimic human-like responses, it may lack the deeper sensibility that characterizes human aesthetic appreciation. This divergence poses challenges for developing AI systems that align closely with human values and experiences, particularly in creative domains.