Multimodal Human Trust Perception
Benchmarking emotion and trust perception in advertising, with Capital One
Sep 2025 – Mar 2026 · In partnership with Capital One
How do people perceive trust and emotion in multimodal content, and how well do models capture those perceptions? This project collected human judgments on advertisements. It then benchmarked models against them and studied how reliable the human labels are.
Highlights
- Collected and benchmarked the EVALUATE dataset (920 ads), exposing gaps in how models capture emotion and trust.
- Showed via label-stability analysis that survey design systematically shifts human annotator judgments.