A study published in Imaging Neuroscience in September 2025 reported that GPT-4V assessed social features across 468 images and 234 videos, and showed that the model's results correlated with a human consistency rate of r≈0.79 compared to a pool of ratings from 2,254 people; the researchers also correlated these ratings with fMRI imaging of 97 participants, indicating activation of similar social cognitive regions (LOTC, pSTS, aSTS, TPJ). The results suggest that the model not only superficially categorizes emotions, but also mimics some brain processing patterns - opening possibilities for expanding measurement tools in experimental psychology and neuroscience, although the extent of generalizability across broader populations still needs further complementary studies.
Imaging Neuroscience study - GPT-4V mimics the brain's processing of social cues
