Mitigating Psychological Reactance in AI-Assisted Decision-Making: A Simulation Study of an Explainability Buffer

AI-assisted decision systems can fail to deliver value when users resist, discount, or strategically ignore recommendations—especially when advice is experienced as an autonomy threat (psychological reactance) or when confidence collapses after observing errors (algorithm aversion) (Brehm, 1966; Steindl et al., 2015; Dietvorst et al., 2015). Explanations are frequently proposed as a remedy, yet evidence is mixed: they may support calibration in some settings while also being ignored, misunderstood, or misused, depending on cognitive cost, incentives, and evaluation choices (Miller, 2019; Doshi-Velez & Kim, 2017; Buçinca et al., 2020; Kaur et al., 2020; Jacovi & Goldberg, 2020). We contribute a transparent simulation testbed that makes reactance and explainability strength explicit levers and uses them to generate falsifiable, mechanism-linked hypotheses for subsequent human-subject studies. Advice uptake is modelled as a two-stage process—consider the recommendation, then (conditional on consideration) follow it under initial disagreement—consistent with staged advice utilisation accounts (Bonaccio & Dalal, 2006; Yaniv, 2004). We sweep reactance λ ∈ {0, 0.25, 0.5, 0.75, 1.0, 1.2, 1.25, 1.5} and compare two explanation-strength regimes (E ∈ {0.2, 0.7}), holding other behavioural parameters fixed to isolate the interaction. Across five random seeds with 1,200 simulated decision-makers per condition, we observe (i) a monotonic decline in AI influence as reactance increases and (ii) a consistent explainability buffer within the model: higher explainability attenuates influence loss, increases completion, and yields larger regret reductions. These are conditional predictions given explicit assumptions, not causal claims about real users. Author note: Independent researcher; University of Warwick affiliation reflects MSc Behavioural and Data Science (alumni). Code and materials are available via the linked repository and the Zenodo record.