Understanding the influence of design-related factors on human–AI teaming in a face matching task

What the paper says

There has been a steep rise in the use of decision aids enabled by artificial intelligence (AI) for facial identity verification. The use of such systems and the impact of their implementation on human decision-making is still not well understood. The current study aimed to explore factors associated with the design of the paradigm and of the presentation of AI-enabled predictions. Across three pre-registered experiments, we examined the impact of (a) stated AI accuracy, (b) mismatch frequency (i.e. proportion of match and mismatch pairs), and (c) advice type (binary only vs. binary + similarity rating) on performance in a one-to-one face matching task. Participants' performance generally improved when aided by AI compared to a baseline without decision support. The largest improvement was observed when no information on the AI's overall accuracy was provided. Further, the frequency of mismatches did not influence performance, but biased responses. Finally, similarity ratings marginally improved overall performance and increased users' certainty in their decisions, but did not help to dismiss inaccurate predictions. Additionally, two findings were consistent across all experiments. First, participants often failed to dismiss inaccurate AI predictions, resulting in significantly lower performance compared to accurate predictions. Second, on a group level, the human-AI team did not outperform the AI alone, though examination of individual performance showed that some participants were able to exceed the AI's accuracy. These findings contribute towards determining appropriate design formats for AI prediction in a human-in-the-loop system, so that the performance of the human-AI team can be maximised.