Robots are both anthropomorphized and dehumanized when harmed intentionally

The study investigates whether witnessing intentional harm to a robot increases attribution of mind—particularly the capacity to feel pain—to the robot (harm-made mind), and whether this effect is amplified when the robot can detect and simulate emotions. Grounded in the theory of dyadic morality, moral events are construed as agent–patient dyads, linking moral agents to agency (intent) and moral patients to experience (feelings, notably pain). Prior work showed intentional harm augments mind attributions to entities with ambiguous or absent minds (robots, PVS patients, the dead). As social robots increasingly recognize and display emotions, ethical concerns arise that such features may enhance perceived moral status and manipulate users. The paper aims to replicate the harm-made mind effect and test moderation by robots’ emotion-detection/simulation, and to explore whether a human who harms a robot is seen as less prosocial than one who treats it kindly. Hypotheses: H1, harm increases perceived capacity for pain; H2, this effect is stronger for emotion-simulating robots; H3, perceived pain mediates effects of harm on broader mind perception; exploratory H4, harming a robot reduces perceived prosociality of the moral agent; H5, this prosociality effect is stronger when the robot simulates emotions.

Prior research (Ward et al., 2013) demonstrated that intentional harm increases mind attribution to liminal-mind agents via dyadic completion. Subsequent studies with robotic/humanlike avatars replicated the effect with smaller sizes. Emotion-related capabilities and socially interactive behaviors generally increase perceived robot mind and moral status; attributing affective states to robots can reduce willingness to sacrifice them and increase reluctance to harm. Detailed facial articulacy and humanlike expression amplify mind attribution when robots are mistreated. Concerns exist that emotion-simulating robots may be manipulative. The social robot–human virtue link posits that cruelty to robots could relate to indifference toward cruelty to living beings; evidence is limited but watching robot abuse can elicit anger toward the abuser. Together, literature suggests harm should elevate perceived experience (pain), potentially more so for emotion-simulating robots, and that observers may judge harmers as less prosocial.

Design: Two close replications of Ward et al. with a 2 (Harm: yes/no) × 2 (Emotions: yes/no) between-subjects design. Dependent variables: perceived capacity for pain (robot), mind perception (overall, and subscales agency/experience), perceived prosociality of the moral agent; manipulation check of moral wrongness. Both studies preregistered on OSF. Participants: Experiment 1 recruited 452 via Prolific; final N=429 (51% men, 48% women, 1% other; mean age 28, SD 9.33, range 18–66) after exclusions for very short completion times and timeout. Sensitivity: detectable f≈0.136 (ANOVA main/interaction), f≈0.026 (LMR, 3 predictors) at 80% power. Experiment 2 recruited 700; final N=677 (38% men, 61% women, 1% other; mean age 34.59, SD 12.27, range 18–75) after exclusions; English first language required. Sensitivity: f≈0.108 (ANOVA), f≈0.016 (LMR). Stimuli: Emotions manipulation described robot “George.” Emotions=yes used Ward et al.’s vignette: complex social robot able to detect/show emotions (happiness, surprise, fear); Emotions=no described similar robot lacking emotional abilities. Pretest (n=75) confirmed higher perceived emotional experience for Emotions=yes (M=3.25, SD=1.53) vs Emotions=no (M=1.48, SD=0.97), p<0.001, d=1.38; no significant difference in perceived capacity to think/act. Harm manipulation: Experiment 1 used Ward et al.’s text. No harm: caretaker Dr. Richardson re-oils circuits; Harm: caretaker abuses robot by stabbing a scalpel into sensors due to envy of the creator. Experiment 2 clarified intentionality and dyad: Dr. Richardson as robot’s creator harms the robot out of boredom with maintenance; explicitly stated awareness of harm (harm vs no harm), and slightly adjusted wording to be less suggestive about emotion sensitivity. Procedure: Participants read Emotions and Harm paragraphs (same page), then completed questionnaires. Random assignment to conditions. Compensation £0.63. Measures: Manipulation check—perceived moral wrongness (1=extremely wrong to 7=extremely right). Perceived capacity for pain—single item (7-point agreement). Mind perception—18 items from Gray et al. excluding pain; 7-point agreement. PCA yielded two factors (agency: 5 items; experience: 13 items) in both experiments; reliability: overall mind α=.94 (both), agency α=.81/.84, experience α=.94/.97. Perceived consciousness measured but reported in supplement. Perceived prosociality—adapted 16-item adult prosocialness scale (1–5), averaged; α=.96/.97. Analyses: Two-way ANOVAs with Harm and Emotions on manipulation check, pain, prosociality; for pain, log10 transform checked but untransformed means reported. Bayesian ANOVAs (JASP) compared models. Mediation analyses used Hayes PROCESS model 4 (5000 bootstraps) with Harm→Pain (mediator)→Mind (overall and subscales), including Emotions as covariate; reported indirect (ab), direct (c′), and total (c) effects. Assumptions (e.g., homogeneity) evaluated; where violated, results noted; transformations attempted when relevant.

Manipulation check: Strong main effect of Harm on moral wrongness in both experiments; harm rated more morally wrong (Exp1 Harm M=2.33, SD=1.36 vs No harm M=5.14, SD=1.18; Exp2 Harm M=2.72, SD=1.31 vs No harm M=5.15, SD=1.16). No main effect of Emotions; no significant Harm×Emotions interaction (both frequentist and Bayesian favored main-effects-only models; BF10>100 vs null; interaction models not supported). Perceived capacity for pain: Main effect of Harm (small) in both experiments: harm increased attributed pain capacity. Main effect of Emotions (medium–large): emotion-simulating robots received higher pain attributions. No Harm×Emotions interaction (H2 not supported). Bayesian ANOVAs favored main-effects-only models; interaction models disfavored (Exp1 BF10=0.095; Exp2 BF10=0.238 vs best model). Descriptives (Table 4) show pain ratings below midscale across all conditions. Mediation (H3): Harm→Pain (a path) positive and significant; Pain→Mind (b path) positive and significant. Positive indirect effects of Harm on overall Mind (Exp1 b=0.13, 95% CI [0.01, 0.25]; Exp2 b=0.20, 95% CI [0.10, 0.30]). Total effects of Harm on Mind not significant; direct effects negative (Exp1 c′ b=−0.18, 95% CI [−0.34, −0.02]; Exp2 c′ b=−0.21, 95% CI [−0.33, −0.09]), indicating suppression: positive indirect via pain and negative direct effect. Subscales: Agency—positive indirect via pain (Exp1 b=0.05, 95% CI [0.002, 0.11]; Exp2 positive via pain), no significant direct effect of Harm on agency (Exp1 b=−0.15, p=0.25; Exp2 b=−0.19, p=0.08). Emotions had significant indirect effects via pain on agency in both experiments; no clear direct effects. Experience—positive indirect via pain (Exp1 b=0.15, 95% CI [0.009, 0.30]; Exp2 b=0.245, 95% CI [0.12, 0.38]) and significant negative direct effects (Exp1 b=−0.19, p=0.03; Exp2 b=−0.22, p<0.001), consistent with suppression at the experience level. Emotions had strong positive direct and indirect effects on perceived experience in both experiments (e.g., Exp1 direct b=1.26; Exp2 direct b=1.19; all p<0.001). Perceived prosociality (moral agent): Strong main effect of Harm in both experiments—harmers perceived as less prosocial (supports H4). No Emotions main effect in Exp1; small positive Emotions main effect in Exp2 (higher prosociality ratings when the robot simulated emotions). Interaction: none in Exp1; significant in Exp2, but opposite to H5—Emotions increased perceived prosociality only in the no-harm condition (M=3.40 vs 3.06; difference=0.34, SE=0.08, p<0.001). Bayesian model comparisons favored main-effects-only in Exp1; in Exp2, interaction emerged but not in hypothesized direction. Overall: H1 and H3 supported; H2 and H5 not supported; H4 supported.

Findings replicate the harm-made mind effect: intentional harm to a robot increases perceived capacity for pain, consistent with dyadic morality and dyadic completion, even for robots explicitly described as lacking emotional capabilities. However, contrary to expectations, the robot’s ability to detect and simulate emotions did not moderate the harm-made mind effect or perceived moral wrongness of the act, suggesting the effect’s robustness across differing baseline experience cues. Mediation results reveal a complex pattern: harm increases mind via pain attributions (indirect), but simultaneously exerts a negative direct influence on overall mind—driven particularly by reduced perceived experience—yielding suppression. This mirrors dehumanization dynamics wherein observers may downplay an entity’s mind to rationalize harm. Alternatively, it may reflect tension between an automatic visceral attribution (harm-made mind) and reflective post-hoc rationalization (denial of mind), or coping with uncanny mind-like cues by dehumanizing robots. Prosociality judgments of the agent robustly decreased with harming a robot, indicating that observers’ moral evaluations of humans are influenced by their treatment of robots; in one study, caring for an emotion-simulating robot slightly elevated perceived prosociality when no harm occurred. These outcomes inform ethical design and policy by showing that mind attributions and third-party judgments can shift with context and robot features, while cautioning that harm can both humanize and dehumanize robots in observers’ minds.

Across two preregistered experiments with large online samples, witnessing intentional harm to a robot increased perceived capacity for pain and, via this, broader mind attributions, replicating the harm-made mind effect. This effect did not depend on the robot’s ability to detect and simulate emotions. Simultaneously, harm exerted a negative direct effect on mind—particularly experience—indicating concurrent anthropomorphization and dehumanization. Observers consistently rated a human who harms a robot as less prosocial than one who treats it kindly. These findings contribute to theories of dyadic morality and HRI ethics, suggesting context-sensitive and sometimes conflicting mind attributions. Future work should use more naturalistic stimuli (e.g., videos/embodied robots), include control conditions with no action, examine person-level variability with within-subjects designs, manipulate reflective processes and uncanny feelings, and incorporate physiological/behavioral measures to clarify mechanisms.

Primary limitations include reliance on text-based vignettes, which, while enabling close replication and internal validity, may lack ecological realism compared with videos or real robot interactions. Wording choices could have interacted with manipulations (e.g., mention of sensor damage). The studies lacked a pure inaction control (no maintenance), as the no-harm condition involved active care (re-oiling), which some may view as moral consideration. Analyses were at the group level, potentially obscuring person-level heterogeneity in mind perception and HRI responses (group-to-person generalizability). Effects of simulated emotions might be stronger with physically present robots; mediated versus copresent differences could alter empathy and mind attributions. More naturalistic setups would also enable physiological and behavioral measures of automatic responses.