{"id":"381c84a0-cfef-4859-9f56-5a391448b2e0","arxiv_id":"2507.13041","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A sociological reading of human-robot trust concludes that robots are relied on rather than trusted, and outlines a research agenda treating trust as practical engagement.","lead":"This paper argues that what engineers call trust in robots is more accurately called reliance, because robots cannot meet the social conditions of interpersonal trust. It proposes instead a cross-disciplinary agenda that measures trust as practical engagement, combining sociology with multimodal robot sensing.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The category-error conclusion rests on an unstated choice between objective and phenomenological readings of the four structural conditions; the paper's own evidence and hedging undermine the categorical 'fulfils none' claim.","rationale":"The reader's CONDITIONAL verdict is appropriate. I do not propose a different outcome, so verdict_should_be is UNCHANGED. I partially agree with the reader's weakest_assumption: the definitional premise is indeed contested. But I find a more immediate problem internal to the paper's own argument. The paper oscillates between two incompatible ways of understanding the four structural conditions. Page 3 says the barrier is 'due to their limitations' and 'not yet'—an empirical, contingent claim. Page 4 (III.E) makes a categorical, in-principle claim. These cannot both ground the category-error charge. The objectivity/phenomenology ambiguity is not merely an external critique; the paper's Section IV explicitly cites HRI studies showing trust-violation phenomenology, which would satisfy a phenomenological reading. Thus, even if one grants the four conditions as necessary (the reader's concern), the paper has not shown they are actually absent in HRI. The concrete test distinguishes the two readings. If the authors clarify their reading and reframe III.E as conditional on an objectivist definition, the paper is a valuable interdisciplinary proposal. Otherwise the central claim is either tautological or empirically unsupported.","tokens_in":10643,"tokens_out":7453,"duration_ms":87869,"concrete_test":"One decisive check: take the participants in the integrity-violation condition of [17] (a robot that breaks rules) and code their responses—reported betrayal, disengagement, demands for apology—against the four conditions as the paper defines them. If the conditions are read phenomenologically, these participants satisfy them, refuting 'fulfils none' as stated. If the authors maintain an objective reading, ask the counterfactual: would a humanoid with genuine beliefs and self-interest who deceives a human count as an instance of trust? If yes, the barrier is contingent robot limitations and the category-error charge is an empirical claim that needs field-wide evidence; if no, the conditions are definitionally human-only and the conclusion is a stipulation, not a finding. Either outcome requires rewording Section III.E.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central conclusion (Section III.E) is categorical: HRI 'fulfils none of the structural conditions' for interpersonal trust. But the argument never fixes the status of those conditions. In Section III the authors write that HRI cannot 'at least, not yet' be equated with human-human interaction, and that the barrier is 'due to their limitations'—a contingent, empirical hedge. Section III.E drops the hedge and asserts an in-principle impossibility. This ambiguity is load-bearing: if the conditions are objective features (actual robot intentions, actual capacity for betrayal), then the conclusion is a definitional consequence of the chosen sociological theory and does not establish an error in HRI research—it stipulates that HRI trust is not trust. If the conditions are instead phenomenological features of the trustor's experience, then the HRI evidence the paper itself cites in Section IV (e.g., [17], showing humans respond to robot norm violations with disengagement, accountability, and apology expectations) already satisfies the conditions, and Section III.E is empirically false. The paper cannot have it both ways. The category-error charge is valid only under the objective reading, but that reading makes the charge tautological; under the phenomenological reading, the charge is contradicted by the cited studies. The authors need to state which reading they intend and explain why the other is illegitimate.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that the HRI trust literature suffers from conceptual confusion because it applies interpersonal trust concepts to human-robot interaction, whereas robots fail to satisfy four structural conditions of interpersonal trust drawn from sociology and philosophy: reciprocity, the possibility of betrayal, submission to the trustee's perspective, and the capacity to make normative claims. The authors conclude that HRI trust research may be committing a category error and is really studying 'reliance' or something else. They then propose a reconceptualization of trust as pre-reflexive practical engagement with the world and artefacts, and they discuss multimodal computational approaches for modeling and repairing trust in real time, reporting example classification accuracies. The paper is interdisciplinary in intent, bringing Luhmann, Baier, Hertzberg, and Quéré into conversation with the engineering-focused HRI literature.","tokens_in":10830,"tokens_out":4076,"duration_ms":48676,"significance":"If the central argument were accepted, it would be a substantive intervention: it challenges the dominant operationalization of 'trust' in HRI and suggests that a large body of empirical work may be mislabeled. The paper deserves credit for (i) making an explicit, testable conceptual claim rather than a purely empirical one, (ii) grounding the argument in canonical sociological and philosophical sources rather than relying on strawman versions, (iii) advancing a positive alternative (practical engagement) that is potentially fruitful for HRI research, and (iv) reviewing a set of recent quantitative results on multimodal trust detection, including concrete accuracy figures. However, the argument's categorical character in Section III.E is not supported by the paper's own hedging, and the relationship between the category-error thesis and the later reconceptualization is not fully resolved.","major_comments":[{"comment":"The central conclusion that HRI 'fulfils none of the structural conditions' is stated categorically, but the argument leading to it is hedged: earlier in Section III the authors write that an interaction between a human and a robot 'cannot (at least, not yet) be equated' with human-human interaction, and that the reason is 'due to their limitations.' This is a crucial ambiguity. If the conditions are only contingently unmet because of current robot limitations, then the category error is not in-principle and the conclusion should be qualified. If the conditions are meant to be in-principle unmeetable, the authors' own phrase 'not yet' undermines that reading. The paper needs to fix the status of the four conditions: are they contingent, necessary, or both, and why?","section":"III and III.E"},{"comment":"The paper's own empirical review contradicts the 'no normative claims' condition. Section IV states that 'even in the absence of intentionality, humans often project normative expectations onto robotic systems and their breach can lead to similar emotional and behavioural responses as in human-human trust violations,' citing [17]. This is exactly the kind of normative claim that Section III.D declares incongruous ('it would be incongruous for a human to say to a robot: \"Treat me the way I deserve to be treated!\"'). If the structural conditions are phenomenological (i.e., about the trustor's experience), then the cited evidence satisfies the condition. If they are objective (i.e., about actual robot capacities), then the conclusion is a definitional consequence of the chosen theory rather than an empirical discovery. The authors need to specify which reading they intend and why the other is illegitimate.","section":"III.D and IV"},{"comment":"The argument appears to mislocate one of Luhmann's conditions. In Section III the authors claim that HRI lacks 'the memory of previous interactions' and related reflexive exercise, treating this as a structural condition of trust. But trust in Luhmann's account is a property of the trustor, who in HRI is a human with memory, reflexivity, and situated learning. The human in repeated interactions with a robot does accumulate experience and generalize it. Unless the claim is that the robot itself must have such memory, which would shift the subject of trust from the trustor to the trustee, the paper's use of this condition is questionable and needs clarification.","section":"III and IV"},{"comment":"The proposed reconceptualization of trust as 'practical engagement' seems to abandon the interpersonal-trust framework that generated the category error. If trust is to be understood as a pre-reflexive way of engaging with the world and artefacts, then the four structural conditions discussed at length in Section III are no longer the relevant criteria, and the category-error conclusion may no longer apply to the proposed research program. The paper should state explicitly whether this reconceptualization is meant to replace the category-error thesis, to supersede it, or to argue that existing HRI trust research should adopt this new definition while the category error still identifies a problem in the old one. The current text leaves this relationship unresolved.","section":"IV"}],"minor_comments":[{"comment":"The reported classification accuracies (84% for facial features, 68% for HR/HRV, 69% for EDA) are given without any description of the datasets, task types, or classification procedures; please provide these details or cite the original studies more precisely so the claims can be evaluated.","section":"V"},{"comment":"The phrase 'their accuracy as to be improved' should be reworded, for example to 'their accuracy has to be improved'.","section":"V"},{"comment":"The extended quote from [5] is longer than necessary; consider condensing it to the key sentence about asymmetry to improve readability.","section":"III.A"},{"comment":"Reference [17] is a Master's thesis; for the load-bearing claim that humans hold robots to normative expectations, a peer-reviewed source would be more persuasive.","section":"IV"},{"comment":"The conclusion states that 'we study the concept of trust,' but the paper is primarily a conceptual position paper rather than an empirical study; aligning the wording with the actual contribution would be helpful.","section":"VI"}],"recommendation":"major_revision","confidential_remarks":"The paper is a timely interdisciplinary reflection, but its central claim is undermined by an unresolved ambiguity between a contingent and an in-principle reading of the structural conditions. The authors need to reconcile the categorical statement in Section III.E with the hedges earlier in Section III and with the evidence in Section IV. If the argument is revised to acknowledge that the four conditions are a contested definitional choice rather than an independent yardstick, the paper could become a valuable contribution to the HRI trust debate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know: this is a genuine interdisciplinary synthesis, not a strawman job. It brings Luhmann, Baier, Hertzberg, and Quéré into the HRI trust conversation with more care than most, and the central move—that much of what HRI calls trust is really reliance—is worth taking seriously.\n\nWhat the paper does well: the four structural conditions (reciprocity, betrayal, submission to the other's perspective, normative claims) are drawn accurately from the cited sources, and the argument is internally coherent. It is honest about its limits: Section V explicitly says it does not analyze computational models in depth, and the practical-engagement reconceptualization is framed as a research direction, not a demonstrated result. The institutional-trust angle is a real gap in HRI and worth flagging.\n\nThe main soft spot is the categorical conclusion in Section III.E. Earlier the paper hedges, saying human-robot interaction cannot \"at least, not yet\" be equated with human-human interaction and that the barrier is \"due to their limitations\"—that reads as contingent and empirical. Then III.E asserts that HRI \"fulfils none of the structural conditions,\" as if in-principle. That is a big step. The stress-test note has it right: if the conditions are objective (actual robot intentions, actual capacity for betrayal), the conclusion is a definitional consequence of the chosen theory—it stipulates, rather than discovers, that robots cannot be trusted. If the conditions are phenomenological (features of the trustor's experience), then the paper's own cited evidence—for instance, [17] showing people respond to robot norm violations with disengagement, accountability, and apology expectations—already satisfies them. The paper needs to pick a reading and defend why the other is illegitimate.\n\nThe definitional nature of the argument is not fatal; it is a philosophical stance, and the paper is honest about it. But it means the contribution is a call for a vocabulary shift, not a demonstrated error in HRI. Also, the empirical asides in Section V cite accuracy figures (84%, 68%, 69%) without full methodology—minor for a position paper, but cleaner to mark them as illustrative.\n\nThis paper will be valuable to HRI researchers who want a sociological vocabulary and to sociologists/philosophers who want a concrete test case. It deserves a serious referee; the referee should push for clarity on the status of the four conditions and ask the authors to align their conclusion with their own hedges.","headline":"A well-read conceptual position paper that usefully reframes much HRI trust research as reliance, but its categorical category-error claim rests on an unstated definitional choice.","tokens_in":11475,"tokens_out":1576,"would_cite":false,"duration_ms":17817,"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":"Human-robot interaction lacks the four structural conditions of interpersonal trust, so most HRI 'trust' research is really about reliance.","keywords":["trust","reliance","human-robot interaction","social robotics","sociology of trust","trust repair","category error","interdisciplinary dialogue"],"falsifier":"Run a controlled study in which a robot is programmed to deceive a user for its own apparent benefit — for example, falsely reporting that a task is complete to avoid recharging — and code participants' reactions for normative blame and attributions of ill intent. If those reactions are indistinguishable from responses to human betrayal, the paper's claim that robots cannot betray would be falsified.","tokens_in":10406,"feed_emoji":"🤖","tokens_out":5883,"duration_ms":57109,"temperature":0.7,"pith_summary":"This paper argues that most research on trust in human-robot interaction may be studying something else: reliance. Bringing sociology into dialogue with social robotics, it identifies four structural conditions that make interpersonal trust possible — reciprocity, the possibility of betrayal, the trustor's submission to the trustee's point of view, and the ability to make normative claims — and argues that human-robot interaction satisfies none of them. If that is right, a robot can disappoint but cannot betray, and researchers who speak of \"trust in the robot\" are committing a category error. The paper's constructive move is to propose a broader notion of trust as practical engagement with the world, along with trust in the institutions that deploy robots, and to recommend real-time multimodal measurement and adaptive repair of human-robot expectations.","feed_headline":"Trust in robots is really reliance, not trust","feed_subtitle":"Four conditions make trust real; robots fail them all. The field may be studying reliance instead.","key_machinery":"The argument is carried by a checklist of four structural conditions of interpersonal trust, drawn from Luhmann, Baier, Hertzberg, and Quéré: reciprocity (mutual engagement), the possibility of betrayal, asymmetry as submission to the trustee's viewpoint, and normative claims and obligations. Each section tests HRI's relational trust models, especially the asymmetric model of Schäfer et al. built on Mayer, Davis, and Schoorman, against one condition and finds the interaction fails it. The four-part checklist functions as a criterion: if an interaction lacks all four, the correct term is reliance, not trust.","core_discovery":"The paper's central claim is that human-robot interaction does not reproduce the structural conditions that make interpersonal trust possible, so the concept of trust is being used inappropriately in much of social robotics. A robot can cause disappointment, malfunction, or violated expectations, but it cannot enter a reciprocal commitment, betray, receive the trustor's surrender of judgment, or be subject to normative claims in the way a person can. Therefore researchers in HRI may be committing a category error: studying reliance while calling it trust. The paper does not stop at the negative result: it proposes a broader sociological notion of trust as practical engagement and as institutional trust to rebuild a common vocabulary between sociology and robotics, and it advocates modelling trust dynamically through multimodal behavioural and physiological signals.","pith_inferences":["Extension: the same four-condition test can be applied to other AI artefacts such as chatbots and autonomous vehicles; if they also fail it, the category-error critique generalises well beyond embodied robots.","Extension: the four conditions could be treated as graded dimensions rather than all-or-nothing thresholds; future HRI measures could score an interaction on each condition and reserve the word 'trust' for high scores.","Extension: if institutional trust is the real object, then a robot's trustworthiness may be less about its own behaviour and more about the reputation of the organisation behind it; this is testable by varying the institution while keeping robot behaviour identical.","Extension: the paper's own proposed experiments could test whether users ever treat a robot as a genuine trustee by coding whether their protests after robot deception contain normative blame like 'you shouldn't do that to me' rather than only disappointment."],"forward_implications":["Much of what HRI calls trust is, by the paper's standard, reliance: a performance-based attitude toward a reliable tool, not a relationship with normative force.","The field should stop trying to import the interpersonal-trust analogy and instead study trust as practical engagement, a pre-reflexive way of acting in concrete situations.","Institutional trust becomes a central object: robots are deployed by institutions, so users' trust may target the institution rather than the machine.","Trust repair should be reframed around violated expectations and disengagement, not betrayal, and should be calibrated rather than maximized.","Robots should be given adaptive computational models that read multimodal signals — face, voice, physiology, and behaviour — to detect and repair loss of trust in real time."],"supporting_citations":[{"why":"N. Luhmann, La confiance: supplies the premise that trust requires structural conditions, quoted as 'the structural conditions that make [trust] possible'.","marker":"[28]"},{"why":"A. Baier, 'Trust and Antitrust': provides the distinction between disappointment and betrayal that underpins Section III.B.","marker":"[29]"},{"why":"L. Hertzberg, 'On the attitude of trust': defines the grammar of trust versus reliance, including the 'look up from below' asymmetry, used in Section III.C.","marker":"[30]"},{"why":"L. Quéré, Avoir confiance: supplies the normative-claims condition, where being trusted creates obligations and shared vulnerability, used in Section III.D.","marker":"[21]"},{"why":"N. Luhmann, Trust and Power: provides the cognitive-economy and familiarity account of interpersonal trust that HRI allegedly lacks.","marker":"[22]"},{"why":"Schäfer et al., 'Trusting robots': the asymmetric relational trust model that the paper critiques for reducing trust to reliance by making reciprocity irrelevant.","marker":"[5]"},{"why":"Mayer, Davis, and Schoorman, 'An Integrative Model of Organizational Trust': the source of the vulnerability-based definition adopted by the HRI trust model that the paper challenges.","marker":"[13]"}],"fun_headline_variants":["Robots can't be trusted, only relied upon","Trusting robots is a category error","Robot trust is a misnomer, call it reliance","Robots don't qualify for trust, only reliance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands on the premise that the four structural conditions drawn from sociology are necessary for any genuine trust; if trust can be one-sided, graded, or non-reciprocal, as some HRI researchers define it, then the category-error conclusion does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Robots can't be trusted, only relied upon","Trusting robots is a category error","Robot trust is a misnomer, call it reliance","Robots don't qualify for trust, only reliance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000803,"raw_usage":{"total_tokens":3475,"prompt_tokens":836,"completion_tokens":2639,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":2578}},"tokens_in":452,"tokens_out":2639,"duration_ms":24548,"temperature":1.0,"reasoning_tokens":2578,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:31:45.469932+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a controlled study in which a robot is programmed to deceive a user for its own apparent benefit — for example, falsely reporting that a task is complete to avoid recharging — and code participants' reactions for normative blame and attributions of ill intent. If those reactions are indistinguishable from responses to human betrayal, the paper's claim that robots cannot betray would be falsified.","supporting_citations":[{"cited_title":"On the attitude of trust,","cited_arxiv_id":null,"evidence_quote":"L. Hertzberg, 'On the attitude of trust': defines the grammar of trust versus reliance, including the 'look up from below' asymmetry, used in Section III.C."},{"cited_title":"Trusting robots: a relational trust definition based on human intentionality,","cited_arxiv_id":null,"evidence_quote":"Schäfer et al., 'Trusting robots': the asymmetric relational trust model that the paper critiques for reducing trust to reliance by making reciprocity irrelevant."}],"review_version":1}