Larger facial ID galleries increase witness misidentification risk, study finds

A new paper from researchers including Genesis Argueta and Kevin W. Bowyer examines how scaling up facial recognition galleries impacts the accuracy of witness lineups. The study simulates a common forensic workflow: a probe image from surveillance video is matched against a gallery of driver’s licenses or booking photos. The algorithm’s top-ranked (rank-one) image becomes the “suspect” in a lineup shown to a witness. By testing galleries of 500, 5,000, and 24,000 images, the team found that larger galleries significantly increase the probability that the rank-one image is a non-mated (incorrect) match—and that witnesses are more confident in those erroneous selections.

The implications are stark. The authors note that this facial identification process has already contributed to at least 9 known wrongful arrests in the U.S. Their data suggests that as police and forensic agencies adopt ever-larger databases (e.g., statewide driver’s license photos), the risk of false positives grows non-linearly. The study questions whether an image derived from such a process should ever be placed in a photo lineup, and whether a witness lineup alone should constitute probable cause for arrest. The findings are a clear warning to law enforcement and courts about over-reliance on automated face matching for suspect identification.