Melanie Bancilhon¹ (), Audrey Siqi-Liu², Victor Platzer², Dwight J. Kravitz^2,3, Kelvin S. Oie¹, Stephen R. Mitroff²; ¹US Army DEVCOM Army Research Laboratory, ²The George Washington University, ³US National Science Foundation

Computer-aided detection and artificial intelligence (AI) cues—systems that provide information on the possible presence and/or location of targets—have the potential to enhance performance across various tasks, including visual search environments (e.g., aviation security screening, medical diagnostics). Such tasks, which often require rapid, accurate decision-making under pressure, can result in high error rates, making the integration of assistive AI tools highly valuable. Prior work demonstrates that the effectiveness of human-AI collaboration depends on both the accuracy of the AI and the human's ability to appropriately rely on AI recommendations. This raises an important question—might some people be more amenable to the use of AI-assisted decision-making tools? The current project investigated how individual differences mediate reliance on AI in AI-assisted visual search. We conducted a large-scale online experiment where we administered self-report surveys (e.g., perception of AI, trust in automation, subjective workload, personality traits, need for cognition) and a single-target (50% target prevalence) behavioral “Ts and Ls” visual search task incorporating AI-assistance. To mimic real-world AI-assisted decision scenarios wherein humans have comparable capabilities to the AI, we calibrated the AI’s accuracy (hits, correct rejections, false alarms, and misses) to each individual user’s performance on an initial block undertaken without AI assistance. Participants then completed counterbalanced experimental blocks with and without AI-assistance. In the AI-assisted block, participants received AI recommendations indicating whether a target was present or absent prior to, and during, stimulus presentation. The relationship between individual differences and the adherence to the AI-assistance will be discussed in light of implications for both theoretical accounts and for designing adaptive AI systems that can personalize interactions based on individual user characteristics. More broadly, the project aims to highlight the role of individual differences in AI-human interactions to inform future research in adaptive AI and human-AI collaboration.

Acknowledgements: ORISE, ARL Cooperative Agreements #W911NF-23-2-0210 and #W911NF-24-2-0188