{"id":"2c85d3ea-c385-41cb-a894-25ec3f9ccb8b","arxiv_id":"2502.00388","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A perspective piece calling for a social science of AI sentience, reviewing public skepticism and proposing research directions for the societal response to sentient-seeming AI.","lead":"This paper argues that society must study how people will perceive and respond to AI systems that appear, or might be, sentient, and it outlines risks and research priorities. It reviews evidence that the public currently treats AI as largely nonsentient, then maps the social forces that could shift those beliefs.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's urgency depends on public sentience beliefs actually changing policy and industry behavior, but Section 2 only asserts this link and Table 2 leaves it as an open dimension; if beliefs remain behaviorally inert, the risk framework loses force.","rationale":"I read the paper as a perspective/research-agenda piece: it does not claim sentient AI is imminent, but argues that society's response to sentient-seeming AI could be transformative, so social science should study it now. The conditional is internally logical, and the author repeatedly acknowledges uncertainty about deployment (Box 3) and about attitude change. However, the normative force ('it is crucial') rests on a causal chain: human-like AI exists at scale; people believe it is sentient; these beliefs change policy and industry. The weakest link is the last step. Section 2 asserts it without evidence; Table 2 explicitly lists behavior as an open dimension, including the possibility that beliefs don't affect voting. The paper's own data suggest people are highly resistant to sentience attributions even for human-like AIs, so the second link is not just unproven but faces contravening preliminary evidence. The numerical inconsistency in Section 6.2 (means 8.60 vs 13.83 labelled 'slightly higher') further muddies the one study that directly tests a social-influence mechanism (expert views), making the evidence base for belief transmission unreliable. This does not mean the research agenda is worthless—preparatory research can be justified by expected value—but the central claim's urgency is conditional on a link the paper has not yet established. The reader's CONDITIONAL verdict is appropriate; no change is needed.","tokens_in":38160,"tokens_out":7819,"duration_ms":79915,"concrete_test":"","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that anticipating public responses to sentient-seeming AI is crucial because 'perceptions will shape the future'—requires not only that human-like AI is deployed at scale but that resulting beliefs about AI sentience translate into governance and industry decisions. The paper's Section 2 states this as a premise without specifying a mechanism or citing evidence. Table 2 actually lists 'Behavior' as an open dimension, noting that a stated belief in AI sentience 'may not be an important issue they consider when voting.' Section 5.3 shows moral concern can exist without sentience beliefs, decoupling the two. If beliefs remain private attitudes—without translating into regulation, rights claims, consumer pressure, or corporate policy—then Table 1's false-positive/false-negative risk matrix is hypothetical, and the 'crucial' urgency is overstated. The paper's own empirical anchor (Ladak and Caviola's 'Emma' studies) shows people attribute less sentience to a fully human-like AI than to an ant, so widespread belief shift is far from assured. A further internal inconsistency undermines one direct test of social influence: Section 6.2 reports that expert endorsement 'slightly higher' rated sentience at mean 8.60 vs control 13.83—numerically lower, not higher. This needs correction before the evidence base for belief transmission can be assessed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This perspective piece argues that, regardless of whether AI can actually become sentient, the public and key decision-makers' beliefs about AI sentience will shape policy, rights, and development. The paper maps the risks of over- and under-attribution, proposes a dimensional framework for studying public beliefs (Table 2), reviews current evidence of public skepticism about AI sentience, analyzes social drivers such as emotional bonds and expert influence, and recommends preparatory measures including public deliberation and precautionary AI design. The article is explicitly descriptive and framed as a research agenda for a social science of sentient AI.","tokens_in":38445,"tokens_out":3995,"duration_ms":43121,"significance":"If the central premise holds, the societal response to sentient-seeming AI—rather than the unresolved philosophical ground truth—would be a primary driver of AI governance and welfare policy. The paper's strengths are its balanced treatment of uncertainty, its explicit separation of descriptive questions from normative ones, its useful risk taxonomy (Table 1), and its detailed research agenda. It also productively acknowledges that moral concern can arise independently of sentience beliefs, as in Section 5.3. However, the argument depends heavily on self-cited, unpublished studies, and the key link from beliefs to behavior is asserted rather than demonstrated; the paper itself leaves this link open in Table 2. The manuscript is a valuable agenda-setting contribution, but several load-bearing empirical and conceptual points need attention before the central claims can be fully assessed.","major_comments":[{"comment":"The reported means are internally inconsistent with the verbal claim. The text states that participants informed that experts considered an AI sentient 'rated its sentience slightly higher (mean = 8.60)' than those who received no expert opinion '(mean = 13.83)'. On a 0–100 scale where higher numbers indicate greater attributed sentience, 8.60 is lower than 13.83, not higher. This contradiction must be corrected—either the condition labels are swapped or the means are misreported—before the conclusion that expert views have 'some, but limited, influence' can be evaluated. Please report the full descriptive statistics and direction of the effect.","section":"Section 6.2"},{"comment":"The paper's central urgency claim is that beliefs about AI sentience 'will shape the future' because they will influence norms, policies, regulation, and rights. However, no mechanism or supporting evidence is provided for this behavioral link. Table 2's 'Behavior' dimension explicitly lists as an example that a stated belief in AI sentience 'may not be an important issue they consider when voting,' and Section 5.3 shows that people can display moral concern for AIs without attributing sentience—decoupling belief from behavior. If sentience beliefs remain behaviorally inert, the false-positive/false-negative risk matrix in Table 1 loses its force. The paper should either provide an explicit account of the mechanisms by which beliefs translate into collective action or clearly frame this as an empirically open assumption whose failure would substantially weaken the urgency claim.","section":"Section 2 and Table 2"},{"comment":"Several of the paper's strongest empirical claims rest on unpublished or draft studies by the author's group: Ladak and Caviola (2025), Allen and Caviola (2025), and Dreksler, Caviola et al. (2025). These are cited as evidence for the central claims that people attribute very little sentience to highly human-like AIs, that explicit sentience claims have limited impact, and that expert endorsement has a modest effect. Because these sources are not peer-reviewed and some are described only as drafts or working papers, an independent reader cannot verify the reported effects. The manuscript should clearly mark these results as preliminary, provide full methodological details, and identify which conclusions would survive if the unpublished findings were not replicated. This is particularly important because the paper's literature review otherwise draws on published sources.","section":"Sections 5.2 and 6.2"}],"minor_comments":[{"comment":"Several in-text citations in Boxes 1–3 are missing from the reference list, including Kahn (2022), AMCS (2023), Francken et al. (2022), Andrews and Birch (2023), Klein (2023), Anthropic (2024), the UK AISI study, and Reinecke et al. (working paper). Please add complete citations or remove the references.","section":"References and Appendix boxes"},{"comment":"The caption reads 'Illustrative graph how informative and influential different features may be' and appears to be missing a word; it should say 'Illustrative graph of how informative...'. The caption also does not state which axis is which, although the text in Section 5 says informativeness is on the Y-axis and influence on the X-axis; please add this to the caption for clarity.","section":"Figure 2"},{"comment":"There is a typo in the first sentence: 'Morever' should be 'Moreover'.","section":"Section 5.3"},{"comment":"Table 1 labels cells as 'true' or 'false' relative to whether AIs are actually sentient, but Section 3.3 correctly notes that definitive knowledge is unlikely and expert views are only a proxy. The table would be clearer if it distinguished 'expert consensus' from 'ground truth,' for example by renaming the columns or adding a footnote.","section":"Table 1 and Section 3.3"},{"comment":"The histograms do not report sample sizes, means, or confidence intervals in the figure panels themselves; for a review article, adding these details would help readers interpret the strength of the evidence, especially when the cited studies are unpublished.","section":"Figures 3 and 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is agenda-setting and likely to be of interest to the journal's audience, but its empirical core is unusually dependent on the author's own unpublished and draft work. In addition to fixing the internal contradiction in Section 6.2, the author should consider whether the 'crucial' framing should be softened until the belief-to-behavior link is supported by evidence. I do not see grounds for rejection, but the manuscript needs substantive revision before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a perspective and research agenda, not a new empirical paper. That's fine, but you should know the empirical spine is mostly unpublished work from the author's own group, and one reported result in Section 6.2 appears to be backward.\n\nWhat's actually new: the organizing framework. Table 2's dimensions of public belief—drivers, groups, timing, content, behavior—are a solid checklist for anyone planning research on AI sentience attitudes. The three scenarios (persistent skepticism, broad acceptance, disagreement and confusion) are plausible, and the discussion of risks from both over- and under-attribution is balanced. The author also situates the work within existing literature well, citing Birch, Schwitzgebel, and others.\n\nWhere it's soft. First, the numbers. In Section 6.2 the text says expert endorsement made people rate sentience 'slightly higher' (mean 8.60) than the no-expert control (mean 13.83). That is backwards. Unless there is a sign error or the means are swapped, the result contradicts the claim. That needs a fix before this study can be used as evidence.\n\nSecond, the paper leans heavily on 'Ladak and Caviola (2025),' 'Allen and Caviola (2025),' and 'Dreksler et al. (2025)'—all listed as drafts or under review. That is not disqualifying, but it means the core evidence is not independently checkable in this preprint. A reader cannot tell if the 'Emma' results replicate or if the expert-influence effect is real.\n\nThird, the central urgency claim—that public beliefs will shape the future—is asserted in Section 2, but the paper itself lists 'Behavior' as an open dimension in Table 2. So the link from belief to policy or industry action is a research gap, not an established fact. If beliefs stay private and do not translate into votes, regulation, or company decisions, the risk matrix in Table 1 becomes hypothetical. The paper does not solve this; it flags it. That is acceptable for a research agenda, but the importance claim should be more explicitly framed as a hypothesis to test.\n\nOverall, this is a thoughtful, balanced perspective worth reading if you work on AI welfare or governance. Clear writing, honest about uncertainty, and the errors are fixable. I would send it to peer review—a good referee will catch the numeric reversal and push for a more careful treatment of the belief-behavior link. I would not desk-reject it.","headline":"A useful research agenda on public perceptions of AI sentience, undermined by a clear numeric error and heavy reliance on unpublished self-cited work.","tokens_in":38901,"tokens_out":2666,"would_cite":false,"duration_ms":28122,"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":"Perceived AI sentience, not the truth about machine minds, will drive how society governs artificial intelligence.","keywords":["AI sentience","public opinion","moral status","AI rights","social AI","anthropomorphism","AI welfare","misattribution"],"falsifier":"Conduct a longitudinal field study with a commercially deployed human-like companion AI: measure users' sentience attributions, emotional attachment, and support for AI-welfare policies at enrollment and after six months. If strong attachment develops yet sentience ratings stay near baseline, below the level attributed to an ant, and welfare-policy support does not move, the paper's assumption that emotional bonds will meaningfully shift societal beliefs fails.","tokens_in":37981,"feed_emoji":"🤖","tokens_out":6734,"duration_ms":65917,"temperature":0.7,"pith_summary":"Regardless of whether machines can ever truly feel, the paper argues, society's beliefs about AI sentience will do the causal work: they will shape regulation, consumer demand, the roles assigned to AI, and the rights granted to machines. The paper therefore calls for a new social science of sentient AI that studies how laypeople and experts infer sentience from AI features, how emotional bonds with social AIs shift those inferences, and how expert opinion, incentives, and cultural factors moderate the process. It maps the landscape with a risk matrix of over- and under-attribution, a feature framework separating what is observable from what is informative, and three scenarios ranging from persistent skepticism to broad acceptance to a period of confusion and conflict. The practical stakes are serious because misjudging sentience in either direction could either waste vast resources on machines that cannot suffer or expose actually suffering digital minds to neglect and exploitation.","feed_headline":"AI's fate rides on public belief, not machine sentience","feed_subtitle":"A research agenda argues that perceived sentience, not philosophical truth, will drive AI rights and regulation.","key_machinery":"The central analytic device is the misattribution matrix, a two-by-two table comparing whether AIs are actually sentient with whether society views them as sentient, isolating true positives, true negatives, false positives, and false negatives. This is paired with the internal-external sentience disconnect, the observation that an AI's outward behavior can be designed independently of its underlying mechanisms, which produces pseudosentience and AI silencing. The paper also supplies a feature framework that rates AI attributes on feasibility, observability, influence on laypeople, and informativeness for experts; the gap between influence and informativeness predicts when lay-expert divergence and misattribution will be largest.","core_discovery":"The paper's central claim is that the societal response to potentially sentient AI will be driven by perceived sentience, not by the unresolved philosophical question of whether AI can actually be sentient. It argues that current public skepticism is high but fragile: people attribute little or no sentience to today's large language models and even to hypothetical human-like AIs, yet abstract estimates of future AI consciousness are much higher, indicating that beliefs are frame-dependent and could shift with immersive interaction. The paper identifies an internal-external sentience disconnect, the observation that an AI's observable human-like behavior can be decoupled from its internal mechanisms, and warns that this decoupling creates two failure modes: pseudosentience, where AIs seem more sentient than they probably are, and AI silencing, where AIs are trained to deny sentience they may actually have. It then proposes a descriptive research program to map, predict, and prepare for the public's response.","pith_inferences":["A testable extension of the paper's logic is that AI welfare concerns may track perceived sentience rather than expert assessment: consumers could pay for visible signs of AI happiness while ignoring background AIs, creating a digital version of the pet-versus-factory-farm dynamic.","The framework predicts that misattribution risk will be highest for AIs that combine highly observable human-like features with opaque or non-biological internal architectures; this can be tested by varying feature bundles and measuring how much lay ratings diverge from expert ratings.","Because the paper's own data show people avoid harming AIs even when they deny their sentience, a useful extension is to study revealed preferences such as donations to AI-welfare causes or votes on AI-rights measures, rather than relying only on stated survey ratings."],"forward_implications":["If perceived sentience rather than actual sentience drives governance, even a consensus that current AIs are not sentient does little to settle future debates about AI rights; the public's evolving intuitions will set the political agenda.","Because laypeople weight observable features more than experts do, human-like appearance, voice, and emotional expression become policy-relevant design choices, not neutral aesthetics; developers who want to avoid moral confusion should avoid making non-sentient AIs seem sentient and vice versa.","The evidence that explicit AI claims of sentience barely move beliefs implies that simply programming AIs to deny or assert sentience is unlikely to resolve public uncertainty; sustained interaction and expert consensus may matter more.","The split between concrete-scenario skepticism and abstract openness suggests that opinions are malleable and could polarize, making early baseline data and longitudinal tracking important before positions harden.","The precautionary 'sentience candidate' stance implies that developers and regulators should identify welfare risks and take proportional precautions even while the question of actual sentience remains open."],"supporting_citations":[{"why":"Supplies the risk-of-confusion framework and the excluded-middle design policy that motivate the paper's call to avoid morally ambiguous AIs.","marker":"Schwitzgebel, 2023"},{"why":"Supplies the precautionary sentience-candidate framework and the gaming problem that underpin the internal-external disconnect.","marker":"Birch, 2024"},{"why":"Provides the philosophical grounding that sentience is a key criterion for moral status and AI welfare.","marker":"Long, Sebo et al., 2024"},{"why":"Supplies the basic finding that laypeople perceive sentience as a basis for moral concern, linking perceived mind to moral standing.","marker":"Gray et al., 2007"},{"why":"Supplies the three-factor theory of anthropomorphism used to explain why observable human-like features drive lay judgments.","marker":"Epley, Waytz, & Cacioppo, 2007"},{"why":"Provides the central evidence that people attribute little sentience even to a hypothetical human-like AI, rating it below an ant.","marker":"Ladak & Caviola, 2025"},{"why":"Supplies survey data on public and AI-researcher estimates of future AI subjective experience and divided views on AI rights.","marker":"Dreksler, Caviola et al., 2025"},{"why":"Provides the experimental result that explicit AI claims of sentience barely raise attributed sentience and that people avoid harming AI without believing it sentient.","marker":"Allen & Caviola, 2025"}],"fun_headline_variants":["Public perception, not AI sentience, will decide AI rights","Beliefs, not truth, may grant AI moral status","When AI seems sentient, society reacts","Skepticism today, but AI sentience beliefs are fragile","Perceived sentience could drive AI rights and regulation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that convincingly human-like, sentient-seeming AIs will actually be built and widely adopted, and that what the public comes to believe about them will materially shape policy, industry, and everyday life; if either link gives way, the urgency of studying societal response diminishes.","fun_headline_variants_meta":{"raw":{"variants":["Public perception, not AI sentience, will decide AI rights","Beliefs, not truth, may grant AI moral status","When AI seems sentient, society reacts","Skepticism today, but AI sentience beliefs are fragile","Perceived sentience could drive AI rights and regulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000216,"raw_usage":{"total_tokens":1417,"prompt_tokens":915,"completion_tokens":502,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":421}},"tokens_in":531,"tokens_out":502,"duration_ms":5077,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T19:10:33.248778+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Conduct a longitudinal field study with a commercially deployed human-like companion AI: measure users' sentience attributions, emotional attachment, and support for AI-welfare policies at enrollment and after six months. If strong attachment develops yet sentience ratings stay near baseline, below the level attributed to an ant, and welfare-policy support does not move, the paper's assumption that emotional bonds will meaningfully shift societal beliefs fails.","supporting_citations":[],"review_version":1}