The impact of computer-aided detection systems on typo foraging

Emily Heffernan¹, Benjamin Wolfe¹, Anna Kosovicheva¹; ¹University of Toronto

Users spend a great deal of time searching for information in visual displays. This visual search process can be facilitated by the display system itself. A ubiquitous example of these computer-aided detection (CAD) systems is spell checkers. While they are widely used, little work has explored how they impact the user’s attention and foraging behaviour. In an eye tracking study, we tasked participants (N = 9) with scanning for typos in pseudoparagraphs of random words. Each trial contained 0–4 typos randomly inserted into a block of 50 words. Participants were eye tracked as they scanned for and “collected” typos using a touchscreen interface. Participants completed a baseline task with no annotations and a second CAD condition that was identical but included underlined typos. Critically, the CAD was unreliable—25% of trials contained either a miss, a false alarm, or a miss/false alarm combination. In both CAD and baseline conditions, participants received feedback on the locations of the typos after every trial. We found that, relative to baseline, accuracy was much higher in the CAD condition, despite search time on each trial being significantly shorter. Participants also had shorter fixation-to-click latencies when guided by CAD. However, in the CAD condition, they scanned more sparsely, fixating on fewer words per trial, with more random gaze paths (higher gaze transition entropy) across the screen. These faster and more sporadic foraging behaviours when using CAD may create conditions in which, in the absence of regular and detailed feedback, users may over-rely on CAD and not notice CAD failures. These findings suggest that it is critical to understand how the attentional guidance offered by CAD tools intersects with the user’s visual attention, such that these “aids” serve to augment, rather than distract from, information intake.

Acknowledgements: This work was supported by a SSHRC Insight Grant to BW and AK.