With the rapid advancement of artificial intelligence (AI) technology, public acceptance of its cognitive capabilities has grown, particularly in domains like medical data analysis where AI may rival or surpass human experts. Despite this growing acceptance, public resistance remains high when AI leads decision-making processes, such as in healthcare, which may stem from concerns about ethical boundaries (Longoni et al., 2019; Bigman & Gray, 2018). This highlights the need to understand how the public evaluates AI behavior in complex moral scenarios, a question that cannot be addressed solely by technical regulations. Moral dilemmas, a classic paradigm in moral psychology, provide a framework to explore how individuals judge AI versus human agents when faced with conflicting choices, such as utilitarian (outcome-focused) versus deontological (principle-focused) actions. Although moral dilemma research on human agents is well established, whether these findings extend to AI agents remains to be determined, as mind perception and responsibility attribution may differ for non-human agents (Waytz et al., 2010; Moor, 2006). This study investigates whether public moral judgments of AI diverge from those of humans, and how gender and AI-induced anxiety moderate these differences.

Three online experiments were conducted using an embodied AI agent (robot) to enhance ecological validity (Malle & Scheutz, 2015). Study 1 employed a 2 (agent: human / AI robot) × 2 (moral behavior: utilitarianism / deontology) ×2 (gender: male / female) mixed design with 474 participants (309 females, M±SD = 29.65 ± 5.75). Participants evaluated 14 moral dilemma scenarios (e.g., footbridge dilemma) on an 11-point morality scale (-5 to +5). Study 2 (N = 290) introduced an anxiety manipulation: participants read five negative (anxiety group) or positive (non-anxiety group) AI-related texts before moral judgments. Anxiety was self-rated on a 7-point scale. Study 3 (N = 196) added a control group reading AI-irrelevant materials to disentangle AI-induced anxiety from baseline anxiety. All studies measured demographic variables, gender, and pre-existing attitudes using scales like the Oxford Utilitarianism Scale (Kahane et al., 2017). Data were analyzed using repeated-measures ANOVA with SPSS, applying Greenhouse-Geisser correction when necessary.

Study 1 showed a significant agent × moral behavior interaction. Compared with human agents, AI robots received higher moral ratings for utilitarian actions (M_AI_utilitarian =.26 ± 1.67; M_human_utilitarian = −.67 ± 2.01; t(473) = 16.74, p <.001, Cohen’s d =.77), whereas moral ratings for deontological actions did not differ between AI and human agents (p >.05). Gender moderated this effect: males exhibited a stronger deontological tendency when evaluating human agents (p<.05), whereas females showed comparatively more flexible judgments. Study 2 replicated the agent × moral behavior interaction and confirmed successful anxiety induction (anxiety: M_anxiety = 5.70 ±.81; non-anxiety: M_anxiety = 3.34 ± 1.63; p <.001). Induced anxiety was associated with more conservative moral evaluations and slightly reduced the positive evaluation of AI’s utilitarian actions, but the relevant interaction effect was not significant. Study 3 further indicated a gender-specific effect of induced anxiety: relative to the control condition, females in both the anxiety and non-anxiety groups showed reduced moral acceptance of the AI robot’s utilitarian actions, whereas males maintained higher moral acceptance of AI utilitarian actions. In the control group, baseline anxiety was low, and for both males and females, moral ratings of AI utilitarian actions were significantly higher than ratings of AI deontological actions.

Taken together, the public appears to regard AI as a relatively rational and efficient executor and to apply an evaluation orientation that prioritizes utility maximization and overall efficiency for AI. Relative to humans, outcome-maximizing decisions made by AI are more likely to be judged as morally acceptable, resulting in more favorable evaluations of AI’s utilitarian choices. The results also reveal gender differences: females are more susceptible to the influence of anxiety, which may relate to higher empathy and risk sensitivity, whereas males’ judgments are comparatively more stable and place greater emphasis on utilitarian considerations. Overall, moral judgments of AI are context-dependent and dynamically moderated by emotional states and individual differences. These findings provide evidence for the psychological mechanisms underlying moral evaluations in human-AI interaction and may inform the design of responsible AI systems that align with public moral expectations.