Are generative AI-facial composites of suspects too good? Accurate facial composites decrease memory strength for suspects
Ywomie A. Mota, Melissa A. BakerPurpose
This study aims to examine the artificial intelligence (AI)-witness effect, a recently observed phenomenon demonstrating impaired recognition performance for suspects following an AI-facial recall task, in which eyewitnesses describe a suspect’s face using a simulated AI interface and may then be shown an AI-generated facial composite of the suspect. In addition, this study examines whether the accuracy of the facial composite presented to eyewitnesses moderates the AI-witness effect and tests the interference and processing theoretical accounts for how AI-facial composites might affect recognition performance.
Design/methodology/approach
Mock witnesses viewed a suspect’s face and were randomly assigned to either complete an AI-facial recall task or a no AI-facial recall task control condition. Facial composite accuracy (accurate, inaccurate, no composite control) of the suspect was manipulated. Finally, participants identified the suspect from a showup. Logistic regression and signal detection analyses were used to assess memory strength and response criteria.
Findings
Results showed that when mock witnesses were presented with the accurate AI-facial composite of the suspect, engaging in the AI-facial recall task was associated with impaired recognition performance, worse memory strength and a less strict, more liberal response criterion, compared with the control condition.
Research limitations/implications
The findings are based on a controlled experimental paradigm and require replication across different samples and more ecologically valid settings. Future research should further disentangle the interference and processing accounts.
Practical implications
The results highlight the need for careful evaluation of AI-facial composite systems prior to their widespread use in criminal investigations. Findings are best interpreted as evidence of memory interference under controlled conditions, and further research is needed before conclusions can be drawn about real-world forensic practice.
Originality/value
This study provides one of the first tests of the AI-witness effect, extending established research on verbal overshadowing and facial composite construction to the context of generative AI-facial composite systems.