{"id":"26b96da7-925d-4d44-ba81-c640b6df5b20","arxiv_id":"2601.09869","paper_version":2,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A PRISMA-ScR scoping review of 22 studies finds convergence on attribution-based definitions of anthropomorphisation but divergence in operationalization, a risk-heavy normative framing, and limited empirically grounded governance guidance.","lead":"This paper is a scoping review of 22 studies on the ethics of making conversational AI agents seem human. It maps how researchers define anthropomorphisation, which ethical risks and benefits they emphasize, and what methods they propose for governance.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Gate C and Table 3's empirical-to-normative bridge eligibility rule make the headline finding of a limited empirical-to-governance bridge partly a selection artifact; the limitation is acknowledged (§7) but load-bearing for the abstract's central gap claim.","rationale":"The reader's weakest assumption is exactly the Gate C selection effect, and I agree with that identification. The paper is transparent: it follows PRISMA-ScR, reports inter-rater agreement, pre-registers the protocol, and explicitly acknowledges in Section 7 that the ethics-oriented inclusion criteria may exclude empirical studies without explicit normative framing. For a scoping review explicitly aimed at ethically oriented work, this acknowledged boundary is a limitation rather than a fatal flaw. However, the abstract's headline claim about 'limited empirical work that links observed interaction effects to actionable governance guidance' is worded broadly enough that the selection criterion and the conclusion become partially circular. A supplementary screening pass would settle whether the gap is merely a framing gap in the literature or an actual absence of empirical work relevant to governance. If the test shows that many non-ethics-labeled empirical studies exist, the authors should rephrase the conclusion to make the 'ethically oriented corpus' scope explicit in the abstract; this would be a clarity revision, not a change to the substantive synthesis of the 22 included studies. Therefore I do not change the reader's accept verdict, but I would flag the need for that wording check.","tokens_in":19538,"tokens_out":9451,"duration_ms":100650,"concrete_test":"Re-screen the 133 full-text records and re-run the search without the ethics block. Concretely: (1) classify all 111 excluded full-text records and record how many were excluded solely by Gate C or the Table 3 empirical-to-normative bridge requirement; (2) run the anthropomorphism × technology query for 2021–Oct 2025 with the ethics block omitted, screen with Gates A and B only, and code the newly eligible empirical studies for whether they report measurable effects on trust, overreliance, disclosure, or attachment and whether they offer design or governance implications. If a substantial set of additional empirical studies with concrete design implications appears, the authors should add an explicit 'within the ethics-framed corpus' qualifier to the abstract's 'limited empirical work' sentence; otherwise the claim stands as currently scoped.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the field shows limited empirical work linking observed interaction effects to actionable governance guidance (Abstract; §7). This claim cannot be assessed independently of the eligibility rules that built the corpus. Gate C (§4.3) requires records to contain explicit normative analysis or frame anthropomorphisation in ethical terms, and Table 3 further states that empirical studies are eligible only if they articulate a defensible empirical-to-normative bridge. Because the corpus is admitted only when the very bridge the review is measuring is present, the finding that such bridges are rare is at least partly determined by the inclusion criterion. The review's own Section 7 acknowledges that this strategy may exclude adjacent empirical studies that measure anthropomorphic effects without making normative claims explicit. That is not just a scope note: if a large body of such studies exists—e.g., HCI/consumer experiments measuring trust, overreliance, or disclosure without an ethics vocabulary—the correct conclusion is that the gap is one of explicit normative translation, not of empirical evidence. The review's conceptual synthesis and risk taxonomy survive this point, but the headline empirical-governance gap as worded does not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a PRISMA-ScR scoping review of ethically oriented research on anthropomorphisation of LLM-based conversational agents published 2021–2025. From 910 records, 22 studies were retained after screening and charted for conceptual definitions, ethical challenges/opportunities, and methodological approaches. The review reports convergence on attribution-based definitions, divergence in operationalization, a predominantly risk-focused normative framing, and limited empirical work connecting observed interaction effects to actionable governance guidance. It closes with design and governance recommendations for future research.","tokens_in":19803,"tokens_out":11903,"duration_ms":127192,"significance":"If accepted, the review provides a useful consolidated map of a fragmented and fast-growing area. Its strengths include a preregistered protocol, transparent database-specific search strings, dual full-text screening with reported inter-rater reliability, and a clear distinction between anthropomorphism as attribution and anthropomorphic cues as design features. The proposed taxonomy of epistemic, affective, and social/relational pathways is a conceptually helpful organizing device. The inclusion of the authors' own study [26] is disclosed and does not appear to structurally determine the synthesis. The paper's main contribution is its governance-oriented synthesis, which identifies the missing empirical-to-normative bridge as a key gap.","major_comments":[{"comment":"The Gate C eligibility rule and Table 3's requirement that empirical studies articulate a 'defensible empirical-to-normative bridge' exclude empirical studies that measure anthropomorphic effects without explicit normative framing. The Abstract's finding of 'limited empirical work that links observed interaction effects to actionable governance guidance' is therefore a finding about the explicitly ethics-oriented corpus, not about the broader empirical literature. Section 7 acknowledges the exclusion, but the Abstract wording overstates the result. Please rephrase the finding as limited within the ethically oriented literature, and note in §7 that a separate synthesis of non-ethically framed HCI/consumer empirical studies would be needed to determine whether the empirical base itself is thin or only its normative translation.","section":"Abstract and §4.3/Table 3"}],"minor_comments":[{"comment":"The row for Dennett's intentional stance lists '[26, 44]' as the included studies citing it, but [44] is the primary source; the text (§5.2.1) correctly names [26,45]. Please correct the inconsistency.","section":"Table 1"},{"comment":"Manzini et al. is listed under year 2025, while reference [43] gives the 2024 AIES proceedings; please align the year and venue across the table and reference list.","section":"Table 4"},{"comment":"The disciplinary percentages sum to more than 100% (59% HCI, 32% social sciences, 23% philosophy, 18% ethics). State explicitly that the categories are not mutually exclusive.","section":"§5.1"},{"comment":"The 'Frictional design and social transparency methods' entry cites [33,54], but [54] is not among the included studies; distinguish external supporting references from corpus references in the figure caption or legend.","section":"Figure 4"},{"comment":"The sentence 'See Table 2 in [31] for more details' should be 'Table 2 of [31]' to avoid confusion with the review's own Table 2.","section":"§5.2.2"},{"comment":"The convergence claim in the Abstract ('convergence on attribution-based definitions') should be qualified, given the finding in §5.2.1/Table 1 that 55% of the included manuscripts provide no explicit theoretical anchor; the convergence appears to hold for the subset of studies that offer a definition, which the paper's own Discussion already notes.","section":"Abstract / §7"}],"recommendation":"minor_revision","confidential_remarks":"The paper is a methodologically transparent scoping review whose main claims are defensible after a modest scoping clarification. The self-citation of [26] is disclosed and does not appear to drive the synthesis; no integrity issues. The manuscript fits the journal's scope and should be acceptable after the abstract is aligned with the review's explicit ethics-oriented scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi,\n\nThis paper is worth reading if you work on AI ethics or human-AI interaction. It is a scoping review, not a new empirical result, but it does what a good scoping review should do: it gives you a reliable map of a fragmented literature. I'd suggest sending it to peer review rather than desk rejecting.\n\nThe genuinely new part is the three-way synthesis: conceptual foundations, normative stakes, and methodological approaches in LLM-era work on anthropomorphising conversational agents. The distinction between anthropomorphism as attribution and anthropomorphic cues as design features is clear, and the finding that most ethically oriented work leans risk-forward while opportunities are thinly argued is a fair reading of the corpus. The authors also follow PRISMA-ScR, report search strings and dates, calculate inter-rater reliability (κ=0.76), and put their protocol on OSF. That transparency earns credit.\n\nThe main soft spot is the one the authors themselves flag in Section 7 but under-weight: Gate C and Table 3 require records to contain 'explicit normative analysis' or an 'empirical-to-normative bridge' to be included. Because the corpus is admitted only when the bridge is already present, the finding that such bridges are rare is partly determined by the inclusion criteria. If there is a large body of empirical studies measuring trust, overreliance, or disclosure effects without an ethics vocabulary—and my guess is there is—those studies are excluded. The correct reading is probably that the field lacks explicit normative translation, not necessarily empirical evidence. This doesn't invalidate the conceptual or normative synthesis, but it blunts the abstract's claim about 'limited empirical work that links observed interaction effects to actionable governance guidance.' That claim needs a qualifier like 'in the explicitly ethically oriented literature.'\n\nThe self-citation point is minor. One included study is the authors' own (Ferrario et al. [26]), but the synthesis draws on 22 studies and doesn't structurally depend on it. Not a problem.\n\nNet: this is a solid contribution for AI ethics researchers and policymakers who want a map of the ethical debate around anthropomorphic LLMs. The selection bias is real but manageable if the authors tighten the abstract and discussion. I'd happily cite it as a reference point.\n\nRecommendation: send it to peer review; with revisions to separate the filtered-corpus finding from the broader field claim, it can be a useful published review.","headline":"A transparent, well-run scoping review whose mapping and research agenda are genuinely useful, though the headline empirical-governance gap is partially an artifact of its own inclusion criteria.","tokens_in":20273,"tokens_out":3750,"would_cite":true,"duration_ms":39362,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A review of 22 ethically oriented studies finds that the field agrees on attribution-based definitions of anthropomorphism but lacks shared operationalization, is risk-forward, and has a thin empirical-to-normative bridge.","keywords":["anthropomorphism","conversational agents","large language models","AI ethics","scoping review","deception","trust","governance"],"falsifier":"A reader could audit the 111 excluded full-text records or run a parallel search without the ethics keyword block to check whether a substantial share of empirical studies report anthropomorphic effects plus design recommendations without ever using words like 'ethics' or 'moral.' If such studies exist in numbers, the claim that governance guidance is not empirically grounded would need revision.","tokens_in":19438,"feed_emoji":"🤖","tokens_out":3879,"duration_ms":39347,"temperature":0.7,"pith_summary":"This paper is a scoping review of 22 ethically oriented studies on anthropomorphizing large language model-based conversational agents, published in the LLM era (2021–2025). It tries to establish that the field has converged on attribution-based definitions of anthropomorphism but diverges sharply in how it operationalizes the concept, is dominated by risk-focused normative analyses, and rarely connects observed interaction effects to concrete governance actions. The authors argue that this fragmented state leaves design and policy recommendations under-specified. For a reader, the review matters because it identifies what would need to be true—shared definitions, validated measures, longitudinal evidence—before responsible anthropomorphic design can be operationalized.","feed_headline":"Human-like AI ethics is fragmented, 22-study review finds","feed_subtitle":"Researchers agree on what anthropomorphism is, but not on measuring or governing it.","key_machinery":"The analytical engine is the distinction between anthropomorphisation as a psychological attribution and anthropomorphic cues as manipulable design features, together with a three-dimensional taxonomy of attributions (cognitive/epistemic, affective, behavioural/social). This taxonomy organizes the corpus into three risk pathways—epistemic expertise/authority, empathy/care, and relational/normative role attributions—and connects them to normative stakes such as autonomy, dignity, justice, and privacy. The review uses this machinery to diagnose the missing 'empirical-to-normative bridge': observed effects like trust shifts or over-disclosure are rarely tied to testable governance actions such","core_discovery":"Across the included corpus, anthropomorphisation is most often defined as the attribution of human-like mental states or social qualities to non-human systems, while anthropomorphic cues are treated as design features that invite such attributions. The review finds three recurring attribution dimensions—cognitive/epistemic, affective, and behavioural/social—and three pathways through which risks arise: attributions of expertise and authority, attributions of care and empathy, and relational/normative attributions such as friend or partner role-play. Ethically, the literature is primarily risk-forward, centering on deception, overreliance, and dependency, with benefits treated as context-depe","pith_inferences":["Editorial inference: Because the search required explicit normative framing in titles or abstracts, empirical HCI studies that measure anthropomorphic effects without using ethical keywords may have been systematically excluded; including them could narrow or widen the reported 'empirical-to-normative' gap.","Editorial inference: The review's implied hierarchy—simple usability cues ethically lighter than virtue-laden role simulation—suggests a concrete design rule: regulators could tier requirements by the complexity of the human capacity being simulated.","Editorial inference: The convergence on attribution-based definitions invites a measurement program that adapts existing psychological anthropomorphism scales to LLM interaction, which would directly test the review's claim that operationalization is fragmented.","Editorial inference: If the field matures along the suggested lines, one testable extension is a benchmark that evaluates whether de-anthropomorphising interventions actually reduce overreliance and disclosure without reducing usability."],"forward_implications":["If the review's synthesis is correct, future work cannot simply add more ethical warnings; it first needs a shared operationalization of anthropomorphism and validated instruments.","Governance recommendations—AI-identity disclosure, de-anthropomorphising, sandboxing—need to be turned into testable hypotheses with defined outcome metrics.","Ethical permissibility is use-case dependent; low-stakes humanization may be tolerable while simulating complex capacities like empathy or authority is riskier.","Longitudinal and interaction-level studies are needed to capture cumulative effects such as reliance trajectories, dependency, and norm displacement.","Regulatory frameworks should be lifecycle-sensitive, without assuming anthropomorphisation alone constitutes a high-risk feature."],"fun_headline_variants":["Anthropomorphic AI ethics: a fragmented, risk-forward scoping review","Human-like AI cues: ethics review finds definition consensus, governance gap","Anthropomorphism of AI: risks dominate ethical literature, review shows","AI anthropomorphism: three attribution paths, one risk-heavy ethical field","Scoping review maps ethical terrain of anthropomorphised conversational AI"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that screening out studies without explicit normative framing does not systematically miss empirical work that links anthropomorphic effects to governance guidance; if it does, the review's headline gap is an artifact of the search.","fun_headline_variants_meta":{"raw":{"variants":["Anthropomorphic AI ethics: a fragmented, risk-forward scoping review","Human-like AI cues: ethics review finds definition consensus, governance gap","Anthropomorphism of AI: risks dominate ethical literature, review shows","AI anthropomorphism: three attribution paths, one risk-heavy ethical field","Scoping review maps ethical terrain of anthropomorphised conversational AI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1584,"prompt_tokens":744,"completion_tokens":840,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":749}},"tokens_in":488,"tokens_out":840,"duration_ms":9148,"temperature":1.0,"reasoning_tokens":749,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:26:17.591994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could audit the 111 excluded full-text records or run a parallel search without the ethics keyword block to check whether a substantial share of empirical studies report anthropomorphic effects plus design recommendations without ever using words like 'ethics' or 'moral.' If such studies exist in numbers, the claim that governance guidance is not empirically grounded would need revision.","supporting_citations":[],"review_version":1}