Automation Bias in Action: Eye Tracking of Humans Reading Screening Mammograms with and without AI Prompts.
Authors
Affiliations (7)
Affiliations (7)
- Translational Medical Sciences, School of Medicine, University of Nottingham, Clinical Sciences Building, City Hospital Campus, Nottingham City Hospital, Hucknall Rd, Nottingham NG5 1PB, UK.
- Health and Medical Sciences Group, University of Cumbria, Lancaster, UK.
- Department of Colorectal Surgery, Nottingham University Hospitals NHS Trust, Nottingham, UK.
- School of Computer Science, University of Nottingham, Nottingham, UK.
- Department of Screening Programs, The Cancer Registry of Norway, Norwegian Institute of Public Health, Oslo, Norway.
- Department of Health and Care Sciences, Faculty of Health Sciences, UiT, The Arctic University of Norway, Tromsø, Norway.
- Nottingham Breast Institute, Nottingham University Hospitals NHS Trust, Nottingham, UK.
Abstract
Background Incorrect artificial intelligence (AI) suggestions can lead to automation bias; however, their impact on medical image interpretation is underresearched. Purpose To assess how incorrect AI suggestions influence the diagnostic accuracy, read times, and visual search behavior of readers interpreting screening mammograms. Materials and Methods In this retrospective multireader paired study conducted between September 2024 and February 2025, 10 National Health Service Breast Screening Programme mammography readers evaluated a test set of two-view mammography screening examinations. The test set included true-positive (TP), false-negative (FN), false-positive (FP), and true-negative (TN) AI suggestions, verified by 3 years of follow-up or histopathologic analysis. In round 1, readers interpreted cases without AI. In round 2, conducted 6 weeks later, a commercially available AI tool was used as decision support, displaying prompts with a region score of 10 or higher (scale, 0-100). Eye-tracking cameras recorded readers' fixations-maintained gaze-over specific image areas. Wilcoxon signed rank tests were used for paired comparisons between rounds, and Kruskal-Wallis tests compared cases with different AI outcomes. Results The test set (<i>n</i> = 60) included cases with 26 TP, 14 FN, 14 FP, and six TN AI suggestions. Median reader sensitivity was lower for cases with FN AI suggestions when reading cases with AI (39%) compared with unassisted reading (71%; <i>P</i> = .002). Reader specificity was higher for cases with FP AI suggestions (39% vs 21%; <i>P</i> = .004). A greater number of visible (TP and FP) AI prompts led to longer median read times, from 25 seconds (zero prompts) to 34 seconds (four or more prompts) (<i>P</i> = .001). Readers fixated less when reviewing cancer cases that AI failed to detect (FN suggestions) compared with unassisted reading (0.44 vs 0.47 fixations per second; <i>P</i> = .03). Shorter fixation durations were observed when readers interpreted cases with FP AI suggestions compared with unassisted reading (0.54 vs 0.56 second; <i>P</i> = .001). Conclusion Incorrect AI suggestions influenced both reader accuracy and visual search behaviors during mammography interpretation. The greatest negative impact was observed with FN AI suggestions; therefore, AI thresholds should be calibrated accordingly. © RSNA, 2026 <i>Supplemental material is available for this article.</i> See also the editorial by Clauser in this issue. See also the editorial by Abbasi and Giess in this issue.