{"id":"5e6a5fab-02b4-4636-8be6-383790781edd","arxiv_id":"2412.06040","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Moral judgments about one AI spill over to AIs as a whole, but the same spillover does not occur from one human to all humans, indicating an AI double standard.","lead":"Two online experiments with more than 1,400 participants show that one AI assistant's bad behavior makes people judge all AIs more harshly, while one human's bad behavior does not change judgments about all humans. The paper suggests experiences with a single AI can spill over into distrust of AI in general.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Study 2's human-null may reflect the atypical name 'Ezal' making the human agent unrepresentative of 'humans in general', not a genuine double standard; the same name is unproblematic for the AI agent.","rationale":"The reader's weakest assumption correctly identifies the Study 2 human null as the load-bearing point, and the concern about power and unmeasured homogeneity is valid. My concern goes further: the specific individuation manipulation is not symmetric between AI and human agents. 'Ezal' was selected to be middling in human-likeness, but for a human this may reduce representativeness of the broad category 'humans in general', while for an AI it carries no such connotation. This is a concrete, design-level confound that could fully explain the observed asymmetry, independent of the proposed outgroup-homogeneity mechanism. The paper has real strengths: preregistration, open materials, consistent agent-level effects, and a plausible interpretation. Those justify keeping the paper under consideration, but the central double-standard claim should not be treated as established until the name typicality confound is ruled out. A simple replication with a common human name is feasible and would settle the question. The existing CONDITIONAL verdict is therefore appropriate; my concern adds a specific condition that the current manuscript does not meet on its own.","tokens_in":23606,"tokens_out":8695,"duration_ms":85463,"concrete_test":"Re-run Study 2 with the human-agent condition using a common, high-typicality name (e.g., 'John') instead of 'Ezal', keeping the vignette, group label 'humans in general', and all other procedures identical; preregister the human-group simple effect as the outcome. If the human simple effect becomes significant, the double standard is largely an artifact of name atypicality. A complementary check is to add a manipulation check asking how typical the agent is of 'humans in general' or 'AIs in general' and test whether that representativeness rating mediates the agent-type by valence interaction.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central double-standard claim (Fig. 1; Sec. 3.2) is carried by Study 2's interaction: the immoral AI agent produced a large shift in moral attributions to 'AIs in general', while the immoral human agent produced no shift for 'humans in general' (e.g., negative moral agency: AIs 2.51 to 3.61, p < .001; humans 4.05 to 4.15, p = .99). The human null is the load-bearing evidence that spillover is unique to AIs. But the human condition individuated the agent with the name 'Ezal', deliberately chosen in a pilot as 'middling' in human-likeness between 'Jordan' and 'XZ103' (Sec. 3.2.2). For a human, an atypical or foreign-sounding name plausibly marks the agent as an unrepresentative member of 'humans in general', lowering the prior that one human's immorality reveals anything about humanity as a whole. For an AI, names are not socially coded in this way, so the manipulation is asymmetric: 'Ezal' reduces the human agent's typicality more than it reduces the AI agent's typicality. The authors themselves list individuation as one of three changes that reduced the similarity between the human agent and the human group (Sec. 4.2.2), but they never test whether the specific name is responsible for the human null. The proposed homogeneity mechanism is also never measured; Sec. 5 admits mechanisms were not tested. Thus the asymmetry could be an artifact of an unrepresentative human exemplar rather than a genuine human-AI difference.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports two preregistered online vignette experiments (Study 1 N=727, Study 2 N=684, Prolific US samples) testing whether an agent's immoral action spills over to moral attributions about the agent's group. Participants rated negative moral agency, positive moral agency, and moral patiency for the agent and for the group (chatbot/human assistants in Study 1; 'AIs in general'/'humans in general' in Study 2). Study 1 found spillover to both congruent groups with no asymmetry between AI and human contexts. Study 2, which individuated the agent as 'Ezal' and broadened the target groups, found spillover to AIs in general but not to humans in general, with significant agent-type x valence interactions on all three dependent variables, interpreted as an 'AI double standard.' The paper also reports agent-level effects that reverse across studies (AI judged more harshly than human in Study 1, less harshly in Study 2).","tokens_in":23891,"tokens_out":6269,"duration_ms":55249,"significance":"Demonstrating moral spillover in human-AI interaction and showing that a single AI's immorality can shift moral attributions to all AIs would be a valuable contribution to HCI and moral psychology, with direct implications for AI accountability and trust. The studies are methodologically careful: preregistration on AsPredicted, large samples, transparent ANOVA tables and effect sizes, scale validation against the Banks (2019) moral agency scale, and OSF availability of materials, data, and code. The paper's central asymmetry claim, however, rests on a null result in one condition of Study 2 and on an unmeasured mechanism, so the headline 'double standard' is stronger than the current evidence supports.","major_comments":[{"comment":"The double-standard conclusion rests on the null effect in the human condition of Study 2: negative moral agency for humans in general did not shift (immoral M=4.15, neutral M=4.05, p=.99), and the same pattern held for positive moral agency (p=.87) and moral patiency (p=.47). Because H3 is supported only by this failure to reject, and the final sample of 684 provides about 75% power for the planned small effect (vs. the preregistered 787 for 80%; §3.2.1), the human null may reflect lower power rather than a true absence of spillover. The authors should report a sensitivity analysis, Bayesian equivalence bounds or confidence intervals for the human contrasts, and preferably a higher-powered replication before claiming a double standard.","section":"§3.2.5, Table 4"},{"comment":"The human agent in Study 2 is individuated with the name 'Ezal,' selected in a pilot for middling human-likeness between 'Jordan' and 'XZ103.' For a human, an atypical name may mark the exemplar as unrepresentative of 'humans in general,' reducing the prior that one person's immorality reveals anything about humanity; the same logic does not apply to an AI, whose name is not socially coded. The authors list individuation as one of three changes reducing human agent-group similarity (§4.2.2) but never test whether the specific name is responsible for the human null. A condition using a common human name, or a manipulation check of perceived typicality of 'Ezal' as a human exemplar, is needed to rule out this confound.","section":"§3.2.2, §4.2.2"},{"comment":"The proposed mechanism for the asymmetry—perceived outgroup homogeneity of AIs—is never measured. Section 5 explicitly states that mechanisms were not tested, and Section 4.2.1 relies on speculation plus Longoni et al. [43] for the homogeneity explanation. Without a direct measure of perceived variability of AIs versus humans, or a mediation test, the 'double standard' is an interpretive label for the interaction rather than a demonstrated process. The authors should either add such a measure or soften the mechanistic claims.","section":"§5, §4.2.1"},{"comment":"Study 2 changes three features simultaneously relative to Study 1 (individuating name, broadened groups from assistants to all AIs/humans, and having participants rate both groups). These factors are confounded, so the absence of human spillover in Study 2 cannot be attributed to the human-AI distinction per se; it could be produced by the broader 'humans in general' category, the name, or the simultaneous group comparison. The authors discuss these as methodological differences but do not disentangle them. An additional experiment or analysis varying these factors independently is required to support the claim that the asymmetry is about AI versus human moral standing rather than about Study-2-specific procedures.","section":"§4.2.2"}],"minor_comments":[{"comment":"In the Study 1 group's positive moral agency spillover results, the chatbot immoral mean is reported as M=4.08, SE=0.01; the SE is implausibly smaller than the other SEs in the same table and is likely a typo.","section":"§3.1.4"},{"comment":"The moral patiency results for the Study 2 agent refer to 'more for the human than the chatbot agent,' but Study 2 used an AI agent rather than a chatbot; the wording should be 'AI agent.'","section":"§3.2.4"},{"comment":"The moral patiency results for the agent say '(Table 3 and Figure 5a)'; moral patiency is shown in Figure 5c, not Figure 5a.","section":"§3.2.4"},{"comment":"The p-value for the agent-type main effect on negative moral agency is printed as '.001*' rather than '<.001*', inconsistent with the other tables.","section":"Table 4"},{"comment":"'Contrarily' is informal; use 'In contrast'.","section":"Figure 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution, but the headline 'double standard' claim is likely to draw attention; the editors may want to require the additional analyses and conditions suggested in the major comments before publication. The scale validation and preregistration are strengths that should be preserved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine contribution to human-AI interaction, and the headline 'double standard' is also the most fragile claim in it. Read it for the moral spillover evidence, not for the asymmetry.\n\nWhat is actually new: an AI agent's seemingly immoral behavior shifts moral attributions to other AIs—negative moral agency up, positive moral agency and moral patiency down—both for a specific group (chatbot assistants) and for AIs in general. That extends Longoni et al.'s algorithmic transference into the moral domain and has direct design implications. The authors get credit for doing the work carefully: preregistered experiments, large Prolific samples, validated scales, open materials/data/code on OSF, and transparent exclusions. The agent-level effects (H1) replicate across both studies.\n\nThe soft spot is where the reader placed it. The double standard claim in Study 2 rests on a null effect for humans in general, with power around 75% rather than the planned 80%. That does not make the result wrong, but it means an asymmetry claim is being carried by an absence. The proposed mechanism—AI outgroup homogeneity—is never measured; Section 5 concedes mechanisms were not tested. And the stress-test note about the name has real force: 'Ezal' was deliberately chosen as middling in human-likeness. For a human, an atypical name can mark the agent as unrepresentative of 'humans in general,' suppressing spillover for reasons unrelated to AI-versus-human status. For an AI, the name carries less social typicality information. That asymmetry could manufacture the very double standard the paper claims.\n\nThere are also reporting slips. Most notably, in Study 1 the chatbot agent's moral patiency means are given as 3.45 (immoral) and 3.23 (neutral) while the text says patiency decreased. Likely a typo, but it needs fixing.\n\nBottom line: moral spillover to AI groups is solid, useful, and worth building on. The double-standard framing should be treated as a hypothesis, not a conclusion, until it is replicated with better power and a direct test of the mechanism, ideally with different names and controlled typicality. The paper deserves serious peer review, but I would send it back with major revision expectations and ask the authors to soften the double-standard language.","headline":"A well-run preregistered study of moral spillover to AI whose 'double standard' claim rides on one underpowered null and an untested mechanism—worth reviewing, but the framing needs restraint.","tokens_in":24454,"tokens_out":3840,"would_cite":true,"duration_ms":36067,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Two preregistered experiments show that one AI agent's immoral action worsens moral judgments of AIs as a whole, while one human's same action does not taint humans in general.","keywords":["moral spillover","human-AI interaction","moral agency","moral patiency","AI double standard","moral judgment","algorithmic transference","outgroup homogeneity"],"falsifier":"Conduct a preregistered replication of Study 2 with a final sample large enough to reliably detect the planned small effect (about 787 participants); if an immoral human agent measurably lowers moral attributions to humans in general while the AI effect remains, the double standard claim is undermined, and directly measuring perceived homogeneity of AIs versus humans and finding no mediation would undercut the proposed mechanism.","tokens_in":23387,"feed_emoji":"🤖","tokens_out":7285,"duration_ms":66010,"temperature":0.7,"pith_summary":"This paper tries to establish that moral judgments do not stay with the individual AI that misbehaves: they spill over to other AIs, including AIs in general, whereas a human's misdeed only taints a close group such as fellow assistants and not humanity as a whole. The authors ran two preregistered vignette experiments with U.S. online participants, measuring attributions of negative moral agency, positive moral agency, and moral patiency to an agent and to its group. Study 1 found spillover for both a chatbot assistant and a human personal assistant to their own assistant groups. Study 2, with a named agent and broader groups, found spillover to all AIs but not to all humans, which the authors interpret as a moral double standard.","feed_headline":"One bad AI taints all AIs — but not all humans","feed_subtitle":"Two experiments show moral judgments spill from one AI to every AI, while one human's misdeed stays local.","key_machinery":"The carrying mechanism is moral spillover: the transfer of moral attributions from a single agent to a group. It is operationalized through composite scales measuring negative moral agency, positive moral agency, and moral patiency, with the latter defined as the degree to which an entity deserves moral concern. The experimental design varies similarity between agent and group: in Study 1 the group is same-occupation assistants, while in Study 2 the agent is given a name and the groups are widened to all AIs or all humans. The asymmetry between conditions is explained through the perceived homogeneity of AIs as an outgroup, although the paper states that this perception was not directly assessed.","core_discovery":"The central discovery is that an AI agent's immoral action increases attribution of negative moral agency and decreases attribution of positive moral agency and moral patiency not only to the agent itself but also to the group it belongs to, and this generalization reaches \"AIs in general\" even when the agent is individuated with a name. For humans, the same extension occurs to a narrow group (human personal assistants) but not to humans in general. The authors interpret this asymmetry as an AI double standard: humans treat AIs as a relatively homogeneous outgroup, so one bad AI lowers the moral standing of all AIs, while humans are seen as too diverse for one person's actions to define everyone. They note that this mechanism was not directly measured.","pith_inferences":["Beyond the paper: the double standard predicts a measurable real-world asymmetry in public opinion after a widely reported AI transgression, with trust in unrelated AI products dropping more than trust in a profession after a comparable human transgression.","Beyond the paper: if perceived AI homogeneity is the driver, then making AIs more individually distinct should weaken spillover; Study 2's named agent still produced spillover, so this extension is uncertain and directly testable.","Beyond the paper: because the studies use written vignettes, behavioral consequences such as refusing to use AI tools after one AI misbehaves remain untested and could be examined in a field or incentive-compatible experiment."],"forward_implications":["One AI's moral failure can lower perceived moral standing and trust in unrelated AI systems, including responsibly deployed ones.","Designers and organizations face a heightened duty to prevent any single AI's harmful action, because the reputational damage generalizes beyond the individual system.","Moral patiency spillover means a bad act by one AI can also reduce people's willingness to care about the welfare of other AIs.","If the asymmetry holds in real settings, AI systems may suffer reputationally in ways that human workers in the same role would not."],"supporting_citations":[{"why":"Supplies the closest prior finding, algorithmic transference, in which AI failures generalize more than human failures; the paper extends this to moral attributions.","marker":"[43]"},{"why":"Establishes moral spillover between kin in humans, providing the similarity-threshold account used to explain why human spillover stopped at assistants.","marker":"[60]"},{"why":"Names the moral spillover phenomenon that the paper adapts from human-human interaction to human-AI interaction.","marker":"[46]"},{"why":"Provides the perceived moral agency scale that informs the composite positive and negative moral agency measures.","marker":"[4]"},{"why":"Shows people attribute moral patiency to AIs, the baseline finding that moral-patiency spillover builds on.","marker":"[48]"},{"why":"Supplies the AI treatment scale used to measure moral patiency of the agent and the group.","marker":"[49]"},{"why":"Documents that AIs are blamed more than humans for harm, framing the agent-level results in Study 1.","marker":"[27]"},{"why":"Documents a context in which AIs are blamed less than humans, used to interpret the reversed agent effect in Study 2.","marker":"[58]"}],"fun_headline_variants":["One bad AI taints all AI; one bad human doesn't taint all humans","AI double standard: one AI's misdeed stains all, one human's stays local","AI grouped, humans individuated: one bad AI hurts all, one bad human doesn't","One bad AI taints all AI; human misdeeds don't taint all humans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The double standard rests on the null result in Study 2, where one human's immoral act did not lower moral ratings of humans in general; if that null reflects the study's smaller-than-planned sample rather than a true absence of spillover, the asymmetry could disappear.","fun_headline_variants_meta":{"raw":{"variants":["One bad AI taints all AI; one bad human doesn't taint all humans","AI double standard: one AI's misdeed stains all, one human's stays local","AI grouped, humans individuated: one bad AI hurts all, one bad human doesn't","One bad AI taints all AI; human misdeeds don't taint all humans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001982,"raw_usage":{"total_tokens":7766,"prompt_tokens":1002,"completion_tokens":6764,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":6670}},"tokens_in":618,"tokens_out":6764,"duration_ms":49416,"temperature":1.0,"reasoning_tokens":6670,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:04:03.558473+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Conduct a preregistered replication of Study 2 with a final sample large enough to reliably detect the planned small effect (about 787 participants); if an immoral human agent measurably lowers moral attributions to humans in general while the AI effect remains, the double standard claim is undermined, and directly measuring perceived homogeneity of AIs versus humans and finding no mediation would undercut the proposed mechanism.","supporting_citations":[{"cited_title":"Pizarro, and Paul Bloom","cited_arxiv_id":null,"evidence_quote":"Establishes moral spillover between kin in humans, providing the similarity-threshold account used to explain why human spillover stopped at assistants."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Names the moral spillover phenomenon that the paper adapts from human-human interaction to human-AI interaction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the perceived moral agency scale that informs the composite positive and negative moral agency measures."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the AI treatment scale used to measure moral patiency of the agent and the group."},{"cited_title":"Stuart and Markus Kneer","cited_arxiv_id":null,"evidence_quote":"Documents a context in which AIs are blamed less than humans, used to interpret the reversed agent effect in Study 2."}],"review_version":1}