Not yet human-reviewed
This source record was written automatically by the research pipeline and has not yet been reviewed by the maintainer or integrated into a topic page. Treat it as a provisional summary and verify claims against the original source.
Exploring automation bias in human–AI collaboration: a review and implications for explainable AI
Citation. Romeo, G. & Conti, D. (2025). Exploring automation bias in human–AI collaboration: a review and implications for explainable AI. AI & SOCIETY, 41, 259–278. https://doi.org/10.1007/s00146-025-02422-7
Summary. This PRISMA-guided systematic review synthesizes 35 peer-reviewed quantitative studies of automation bias in AI-assisted decision-making published between 2015 and April 2025. It finds that overreliance is shaped by user expertise and AI literacy, cognitive and workload conditions, trust dynamics, verification demands, and explanation design. The authors argue that transparency or explainability alone often does not improve decision accuracy and can reinforce misplaced reliance when explanations act as trust heuristics.
Key claims
- The reviewed studies operationalize automation bias chiefly as agreement with incorrect AI recommendations, across decision tasks in domains including healthcare, finance, national security, public administration, human resources, and mental healthcare.
- Higher verification intensity was associated with better decision accuracy in the reviewed evidence, while responsibility cues alone did not significantly increase verification intensity in the cited HR-dashboard study.
- Explanations, transparency cues, and trust-calibration feedback did not reliably help users recognize incorrect AI recommendations; overly technical, cognitively demanding, or simplistic explanations can reinforce overreliance.
- Domain expertise and AI literacy can support more appropriate reliance, but experienced users can still exhibit automation bias under high cognitive load.
Evidence & limitations
- The review followed PRISMA 2020 procedures and included 35 peer-reviewed quantitative, experimental studies; its findings synthesize heterogeneous designs, measures, and application domains rather than estimating a pooled effect.
- Included literature was restricted to English-language studies published between January 2015 and April 2025 and identified through SCOPUS, ScienceDirect, PubMed, Google Scholar, and supplementary searching.
- The source is a peer-reviewed journal article published online in 2025; its conclusions about intervention effectiveness are limited by the underlying studies and context-specific task designs.