Takuto Sato, Rafał Rzepka · International Conference on Human-Agent Interaction (HAI) 2026 · 2026
DOI: 10.1145/3841580.3845731
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
When people delegate tasks to AI agents, oversight requires an evaluation axis from outside the AI, and brain measurement is a leading candidate. But brain data must first be mapped into the AI’s own representation space, so the oversight signal passes through the representation of the system it should oversee. Using four subjects from the Natural Scenes Dataset and MindBridge, we ask how much of a brain representation decoded through CLIP is already predictable from CLIP itself under adversarial perturbation. CLIP alone accounted for roughly three quarters of the across-image variance in the brain-routed quantity; a reproducible component remained in all four subjects, but added almost no incremental explanatory value beyond the CLIP-only score. Because the brain embedding is fixed from clean-viewing fMRI, our design quantifies redundancy in CLIP-mediated representational geometry rather than establishing statistical independence from CLIP. For HAI, the degree of redundancy must be measured for each mediating representation.
No comments yet — start the discussion below.