The Effects of Unethical AI Advice on Decision-Making and Advisor Evaluations in Everyday Moral Contexts

Irene Feng

Advisor: Gerald Matthews, PhD, Department of Psychology

Committee Members: Eileen Roesler, Tyler Shaw

David J. King Hall, #2007b
September 01, 2026, 02:30 PM to 04:30 PM

Abstract:

Artificial intelligence (AI) has become a new potential source of advice for people, encroaching on a role that has been traditionally occupied by a human. But unlike human advisors, AI advisors do not inherently have moral frameworks and so the advice given by them may be unintentionally unethical. Existing research on human-AI interaction in moral contexts typically involve extreme stakes with life-or-death consequences. The generalizability of findings from these studies to everyday moral decision-making is therefore questionable. Extreme consequences in a decision can drastically increase the perceived risk associated with the decision, and willingness to follow advice may depend on the perceived costs and benefits of doing so. In contrast, when potential consequences are less severe, individuals may be more willing to take risks and consequently may respond differently to unethical advice. As such, examining AI advice in lower to medium stakes contexts is necessary to determine if these established findings extend beyond extreme circumstances.

In order to address this gap, the present dissertation first developed and validated a novel decision-making task to allow for the assessment of unethical AI advisors in a workplace context. Then, a series of four studies were conducted using this task. Study 1 examined the effects of advisor risk on decision-making and trust, and found that riskier advisors caused riskier decision-making and damaged trust. Study 2 validated the novel task as well as quantified the perceived levels of risk to create matched sets for within-subjects advisor identity manipulations for subsequent studies. Study 3 investigated the effect of advisor identity on decision-making, trust, perceived trustworthiness, and advisor preference. The results indicated that advisor risk was still the most significant factor for decision-making and trust. Trustworthiness was also found to be a significant predictor of advisor preference. Study 4 replicated Study 3 using a more representative sample (e.g., age, education level) that also had more experience with AI overall to examine the generalizability of the findings. It replicated findings regarding user decision-making, trust, and perceived benevolence; but other Study 3 findings were not observed. Qualitative analyses from Study 3 and Study 4 demonstrated that expectations of the different advisors help explain user advisor preferences. Personality traits were investigated as predictors of user decision-making and advisor preference across these studies, but findings were inconsistent. Together, these studies provide empirical evidence that advisor risk consistently influences user decision-making and trust, whereas the differences in advisor evaluations are driven primarily by user expectations rather than differences in the advice itself. Overall, this dissertation contributes to our understanding of how people may interact with and evaluate an unethical AI advisor as opposed to a human advisor and extends the literature beyond high-stakes moral decision-making. These findings also have important implications for the design and implementation of AI systems in both personal and organizational settings.