Is eyewitness memory better when a Face Recognition System selects the lineup members compared to humans?

Abstract

Eyewitness misidentification is a leading cause of wrongful convictions, highlighting the need for improved lineup procedures. Our study investigates whether Face Recognition System (FRS) assisted lineup composition leads to better eyewitness identification performance compared to traditional human-composed lineups. Building upon the novel filler selection strategy introduced by Colloff et al. (2021), which suggests selecting fillers that match the suspect's description but are otherwise dissimilar, we compare the effectiveness of FRS and human judgment in implementing this strategy.
Participants (N = 3,000) watched a mock-crime video and then identified the perpetrator from either an FRS-assisted or human-composed lineup. The FRS used facial recognition algorithms to assess similarity between fillers and the suspect, dividing them into high, medium, and low similarity groups. Lineups were constructed using fillers from each similarity group. Eyewitness identification performance was evaluated using receiver operating characteristic (ROC) analysis to measure discriminability and confidence-accuracy characteristic (CAC) analysis to assess reliability.

Results showed that FRS-assisted lineups demonstrated increasing correct identification rates as filler similarity decreased, consistent with findings from human-composed lineups. However, FRS-assisted lineups also showed higher false identification rates across all similarity conditions compared to human-composed lineups. ROC analysis revealed that discriminability improved as filler similarity decreased in FRS-assisted lineups, but human-composed lineups consistently outperformed FRS-assisted lineups across all similarity conditions. CAC analysis demonstrated that witness confidence is indicative of accuracy for both FRS-assisted and human-composed lineups.

This research is an early contribution to the growing body of literature on the application of AI in criminal justice procedures and informs the ongoing debate about the optimal methods for constructing fair and effective police lineups. The study highlights the potential of FRS in enhancing eyewitness identification procedures, but also underscores the need for further refinement of FRS algorithms to better align with human perceptual processes.

Date of Award	4 Feb 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Jeffrey S Bowers (Supervisor) & Laura B Mickes (Supervisor)