Using confidence, decision time, and confidence entropy to predict accuracy for online and real eyewitnesses.

It is important to identify factors that police can use to distinguish accurate from inaccurate eyewitnesses (i.e., reflector variables). In this study, we examined three such variables known to reflect eyewitness accuracy: confidence, decision time, and confidence entropy. We examined data from an online study (N = 1,072) and a field study involving real eyewitnesses (N = 75). Confidence was measured in three ways: human-coded verbal statements, AI-classified verbal statements, and numeric ratings. We compared the correspondence between these confidence metrics, how multiple reflector variables relate to accuracy, and whether these patterns differ across two datasets. All confidence metrics predicted identification accuracy to a similar extent, but numeric confidence derived from online participants was the strongest predictor of identification accuracy. In general, including multiple reflector variables better predicted decisions than any reflector variable alone, but there were diminishing returns associated with each additional variable. The results support indirect theoretical models of retrospective confidence and have practical implications for lineup procedures and policy, emphasizing the importance of documenting confidence and other reflector variables to assess the reliability of eyewitness identifications in the field.