2025

Multimodal AI Safety Evaluation

SALT Lab · with Prof. Yun Huang

Systematic review of 176 multimodal AI systems using an LLM-assisted research pipeline (Gemini Pro, κ=0.717 IRR). Found 93% strip social context via modality-to-text conversion, and 45% lack ethical discussion — with evaluation over-relying on static benchmarks over human-centered assessment.

Papers

The Social Gaze of LLMs: A Literature Review of Multimodal Approaches to Human Behavior Understanding

Under review · Preprint