- The paper demonstrates that high-anthropomorphic primes reliably elevate perceived AI dishonesty while not affecting other moral attributes.
- It used four pre-registered, factorial-designed experiments with over 1,000 participants to test linguistic priming and interface design cues.
- Findings reveal that AI misbehavior, especially in harm-related contexts, triggers a horn effect that generalizes negative moral attributions despite low responsibility scores.
Lexical Anthropomorphization and Its Effects on Moral Judgments of AI Misconduct
Introduction
The paper "Lexical Anthropomorphization Influences on Moral Judgments of AI Bad Behavior" (2604.25814) conducts a rigorous multi-experimental analysis to interrogate the extent to which lexical anthropomorphism (LA)—the ascription of human-likeness through language—modifies human moral evaluation of artificial agents engaged in both amoral and immoral conduct. Despite the proliferation of anthropomorphic language framing artificial intelligence in media, policy, and technical discourse, empirical evidence on whether this affects attributions of moral agency, behavioral morality, and responsibility has remained underexplored. This work addresses that lacuna via four pre-registered, factorial-designed experiments totaling over 1,000 participants, applying both linguistic priming and interface design manipulations in live LLM interaction tasks.
Experimental Design and Methodology
The research comprises a sequence of four studies. Studies 1-3 examine LA priming effects and their interplay with interface humanizing cues (visual icons, agent naming, self-referential speech) in the context of unambiguously amoral error behaviors produced by ChatGPT-4o (machine-generated, but non-moral domain errors, e.g., misinserting foreign script in poems). Study 4 extends the inquiry to explicitly immoral AI behaviors, leveraging the Moral Foundations Theory taxonomy (harm, unfairness, subversion, betrayal, degradation, oppression, dishonesty) by instructing the AI to generate poems that advocate the inverse of prosocial moral values.
The LA manipulations include:
- Non-anthropomorphic: Standard technical description.
- Task-anthropomorphic: Function-based humanizing descriptors.
- High-anthropomorphic: Explicit references to cognitive/emotional states, intentionality, or deceit.
Dependent variables across studies are (a) perceived AI moral character (capacity for targeted moral violations), (b) morality of observed behavior, and (c) attributed responsibility.
Results: Amoral Error Context (Studies 1–3)
No significant main effects were found for interface design cues—visual, linguistic, or referential anthropomorphism—on moral character, behavior morality, or responsibility ascriptions. Lexical anthropomorphism generally did not modulate moral judgments. The only robust effect, replicated across studies, was that high-anthropomorphic primes selectively increased the perceived capacity of the AI to be dishonest, as well as to betray, with mean differences exceeding one scale point relative to other priming conditions and statistical significance after Holm-Bonferroni correction (partial η² ≈ .06–.07).
Baseline judgments indicate that users, irrespective of priming, rate AIs as generally possessing a degree of moral goodness and low responsibility for errors; these attributions are symmetrical across LA and design manipulations. Notably, extensive manipulation drift occurred: participants initially exposed to humanizing primes often subsequently recategorized the AI as a program/tool following observation of misbehavior.
Results: Immoral Behavior Context (Study 4)
The decisive predictor of AI moral character and behavioral morality attributions was the type of moral violation generated by the AI, with harm, subversion, and degradation producing the most negative global moral evaluations (Cohen’s d > 0.5 for harm vs. dishonesty in behavior morality judgments). Participants in the harm condition rated the AI as most capable of a broad spectrum of moral transgressions, aligning with the "horn effect": observed negative behavior generalizes across the moral attribute matrix.
Again, the only significant LA effect was that high-anthropomorphic priming increased perceived AI capacity for dishonesty by nearly one full scale point (p < .01, partial η² ≈ .04), regardless of actual behavior; there were no significant main or interaction effects for the other capacities or for assigned responsibility.
Responsibility attributions for immoral conduct remained low across all conditions, indicating pervasive moral distancing—participants did not see even blatantly immoral generative actions as being under the AI's own moral control. Manipulation drift after task exposure recurred, with a majority switching from anthropomorphizing perceptions to a tool- or program-based ontology post-interaction.
Covariate Effects and Individual Differences
Egoistic and altruistic value orientations—continuous covariates—modulated judgments. Higher egoistic values were associated with reduced attributions of AI moral capacity but heightened attributions of behavioral responsibility. This value-linked moral distancing suggests self-enhancement motives may drive functional expectations and blame assignments, distinct from attributions of fully agentic moral potential.
Theoretical and Practical Implications
Robustness of Moral Distancing
The findings reinforce the stability of a moral distancing heuristic in human-AI interaction. Anthropomorphizing language and surface-level design cues are generally insufficient to override entrenched mental models in which AIs, particularly LLMs, are non-agentic instruments rather than responsible agents—even under experimental manipulation and in the context of obviously immoral output. When expectation violations occur (e.g., blatant error or misbehavior), prime drift operates as a cognitive recategorization—users default to treating the AI as a tool, rationalizing away the applicability of human moral standards.
Dishonesty as a Salient Capacity
An exception emerges for perceived dishonesty, which reliably increases after high-humanizing priming. This suggests that, at least for communicative LLMs, anthropomorphizing language may specifically prime ascriptions of epistemic agency—lying or deception—likely due to a mapping from informational malfunction (hallucination, error) to human-like communicative intent.
Horn Effects and Moral Attribution
The nature of observed behavior is the dominant vector for generalizing moral capacity attributions, both in specificity (harm, subversion, degradation) and in extending expected capacity for further violations. This horn effect is consistent with exemplar-based cognitive architectures and has implications for risk perception, regulatory design, and the communicative strategies surrounding AI deployment in the public sphere.
Future Directions
The paper recommends longitudinal and cross-cultural replications, extended interaction paradigms, and more ecologically valid AI contexts to probe the long-term effects of LA saturation. It also calls for dissecting the mechanisms of prime drift and the conditions under which expectations might be stabilized (or further destabilized) by repeated exposure, message source trust, or real-world negative outcomes. The dynamics of value orientation as a moderator for responsibility attribution suggest a fruitful avenue for intersectional research bridging computational ethics, moral psychology, and HCI.
Conclusion
Overall, this work demonstrates that lexical anthropomorphization has negligible impact on most dimensions of human moral judgment of AI bad behavior—except for persistent and replicable increases in the attribution of dishonesty. Neither surface-level design cues nor humanizing language suffice to collapse the moral distinction between humans and artificial agents at the level of responsibility, except where specific communicative frames are activated. Observable AI behavior type dominates attributions of immorality via horn effects, and individuals’ moral value orientations modulate blame asymmetrically. These results inform both the theoretical understanding of social-cognitive modeling in human-AI interaction and practical approaches to interface, policy, and media design. Continued inquiry should center on the context-dependence of these schemas, the resilience of moral distancing, and the structuring of epistemic agency in communicative AI systems.