The Symbiosis of MathPsych, ICCM, and Human Factors Psychology

Increasingly, people work with automated advice that assists their decisions. Recently, an evidence accumulation model (EAM) has been developed to describe the cognitive processes by which humans incorporate automated advice. This model posits two mechanisms through which advice can influence decisions: an advisory process that inhibits evidence accumulation toward decisions that disagree with the advice, and a reliance process that excites accumulation toward decisions that agree with it. However, as is typical of EAMs, it assumes decisions are independent and identically distributed across trials. In practice, the reliability of automated advice can vary over time and across contexts, and people adapt their behavior to such changes. We propose an automation-reliability-learning EAM (ARL-EAM) that models how users update their estimate of automation reliability on each trial via a delta rule, with corresponding adjustments to advice-related accumulation parameters. We tested this model in an air traffic conflict detection task in which participants (N = 96) judged whether pairs of aircraft were in conflict. Participants completed both a manual block (without a decision aid) and an automation block (with a decision aid). Automation reliability began at 90%, dropped to either 70% or 50% (between-subjects), and subsequently recovered to 90%. The ARL-EAM provided a close account of behavioral adaptation to changing automation reliability. Learning about reliability dynamically modulated the parameters governing how advice influenced evidence accumulation. Furthermore, key automation-use parameters were associated with subjective trust in automation. Our findings clarify the cognitive mechanisms supporting flexibility in human-automation teams.

This is an in-person presentation on July 18, 2026 (10:40 ~ 11:00 EDT).

Task difficulty significantly influences decision-makers’ confidence and subsequently dictates how they interact with automated suggestions. In this presentation, we review our recent studies examining the interplay between task difficulty and automation accuracy, utilizing the Single-Target Self-Terminating (STST) capacity framework within Systems Factorial Technology. In our first study, using a basic categorization task, we manipulated automation accuracy and task difficulty. Our results indicated that while participants did not achieve super capacity overall, the benefits of automated aids were highly evident in difficult tasks. Interestingly, when different levels of automation accuracy were intermixed, the impact of overall automation accuracy diminished, whereas the significance of trial-by-trial information accuracy increased. In our second study, we deconstructed “difficulty” by manipulating it via either increased variance or reduced discriminability between judgment categories. We found that low-accuracy automation resulted in limited capacity regardless of the difficulty type. Conversely, high-accuracy automation yielded super capacity, except when category variance was low. This suggested that decision uncertainty induced by category variance amplifies the gained efficiency when participants interacted with automated suggestions. These findings highlight that human decision-making efficiency with automation is highly sensitive to contextual factors. Finally, to extend our basic-level findings to an applied domain, we recruited 12 physicians to diagnose osteoporosis from medical images, both with and without AI-generated probabilities. Here, the physicians achieved unlimited capacity when provided with AI assistance. Together, this review demonstrates how STST capacity can effectively uncover the underlying processing dynamics when decision-makers interact with varying automation accuracies across diverse task difficulties.

This is an in-person presentation on July 18, 2026 (11:20 ~ 11:40 EDT).