Human Response to Decision Support in Face Matching: The Influence of Task Difficulty and Machine Accuracy

TL;DR

This study investigates how human performance in face matching responds to AI decision support under varying task difficulty and machine accuracy. Through three online experiments, it shows that task difficulty strongly shapes both performance and the influence of machine suggestions, with high-accuracy systems improving accuracy, while low-accuracy or changing systems can degrade performance, especially on harder tasks. The authors introduce measures like the Influence Factor $IF$ and Confirmation Probability $P(C)$ to quantify these interactions and explore static versus variable accuracy and the effect of change notifications. The findings have practical implications for designing decision-support interfaces, highlighting the importance of task context, user perception, and transparent communication about AI reliability in high-stakes settings. The work also discusses ethical considerations and calls for extending evaluations to real-world domains and more nuanced interaction patterns.

Abstract

Decision support systems enhanced by Artificial Intelligence (AI) are increasingly being used in high-stakes scenarios where errors or biased outcomes can have significant consequences. In this work, we explore the conditions under which AI-based decision support systems affect the decision accuracy of humans involved in face matching tasks. Previous work suggests that this largely depends on various factors, such as the specific nature of the task and how users perceive the quality of the decision support, among others. Hence, we conduct extensive experiments to examine how both task difficulty and the precision of the system influence human outcomes. Our results show a strong influence of task difficulty, which not only makes humans less precise but also less capable of determining whether the decision support system is yielding accurate suggestions or not. This has implications for the design of decision support systems, and calls for a careful examination of the context in which they are deployed and on how they are perceived by users.

Figures (4)

• Figure 1: Survey screenshots.
• Figure 2: Participant initial and final average accuracy distributions for the Normal Set and the Hard Set, for 5%, 50%, and 95% Machines in Experiment 1.
• Figure 3: Participant accuracy for DEC machines (with and without notification) with the Normal Set (left) and the Hard Set (right).
• Figure 4: Results from the exit survey, from the participants who interacted with some static machine. They were asked whether the suggestions of the machine had been useful to make a decision 1. when they were not sure, 2. when some of the images were blurry or the quality was not good, 3. faster, to 4. confirm their decision when they were certain about it, 5. make more accurate decisions, and 6. make them feel more confident about the answer. We show the average of the responses across the six questions.