Fast, Accurate, But Sometimes Too-Compelling Support: The Impact of Imperfectly Automated Cues in an Augmented-Reality Head-Mounted Display on Visual Search Performance

The benefits and costs of explainable artificial intelligence in visual quality control: Evidence from fault detection performance and eye movements

“…Specifically, for correct detections, we expected lower reaction times (H1-a) and fewer errors (H1-b) with XAI than BBAI, in line with most previous research on cueing effects (e.g., Chavaillaz et al, 2018;Maltz & Shinar, 2003;Warden et al, 2023;Yeh & Wickens, 2001). Conversely, we did not have a single straightforward hypothesis for correct rejections but could imagine outcomes in either of two directions.…”

“…When pooled across all trials, XAI enabled faster and more accurate performance. In principle, it is not surprising to find XAI benefits, as the same has been shown for direct visual cues in various studies over the years (e.g., Chavaillaz et al, 2018;Maltz & Shinar, 2003;Warden et al, 2023). What is more surprising is that these performance benefits occurred despite our AI system's relatively low reliability of 70%.…”

“…But first, let us take a broader look at the cueing effects of our XAI highlights. A cueing or XAI system that has clear benefits will typically go along with a high risk of overreliance (Maltz & Shinar, 2003;Warden et al, 2023;Yeh & Wickens, 2001). Therefore, in Section 4.2 we will address the question whether the present study provided any evidence for overreliance on the (X)AI.…”

“…The effects of direct target cueing on human visual inspection performance have been investigated for at least three decades (e.g., Alberdi et al, 2004;Chavaillaz et al, 2018;Goh et al, 2005;Hättenschwiler et al, 2018;Huegli et al, 2020;Krupinski et al, 1993a;Maltz & Shinar, 2003;Povyakalo et al, 2004;Warden et al, 2023;Yeh & Wickens, 2001). In line with the integrated model of complacency and automation bias (Parasuraman & Manzey, 2010), a common finding is that visual cues generate large benefits when they are accurate, but large costs when they are not.…”

“…In principle, direct visual cues can take various forms, such as arrows, frames, object outlines, or heatmaps. Again, the most effective cues carry the highest risk of overreliance (Warden et al, 2023). For instance, when tumor areas in radiological images were cued by closed, complete shapes (i.e., full circles, as opposed to dashed-line circles or two vertical bars), people were better at detecting targets inside the cued area but worse at detecting targets outside (Krupinski et al, 1993b).…”