Fusing ChatGPT and human decisions in unfamiliar face matching

Unfamiliar face matching involves deciding whether two face images depict the same person or two different people. Individual performance can be error-prone but is improved by aggregating (fusing) the responses of participant pairs. With advances in automated facial recognition systems (AFR), fusing human and algorithm responses also leads to performance improvements over individuals working alone. In the current work, I investigated whether ChatGPT could serve as the algorithm in this fusion. Using a common face matching test, I found that the fusion of individual responses with those provided by ChatGPT increased performance in comparison with both individuals working alone and simulated participant pairs. This pattern of results was evident when participants responded either using a rating scale (Experiment 1) or with a binary decision and associated confidence (Experiment 2). Taken together, these findings demonstrate the potential utility of ChatGPT in daily identification contexts where state-of-the-art AFR may not be available.