{"id":"fa000b31-1a2a-4dfb-bd7d-d3247e6e6c21","arxiv_id":"2601.08878","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"AI in text-based online counselling presents ethically distinct challenges for autonomous bots, training simulators, and counsellor-facing augmentation tools, requiring role-specific governance.","lead":"This paper examines the ethics of AI in text-based online counselling by analyzing three AI roles: autonomous chatbots, training simulators, and counsellor-facing tools. It shows that the same four principles—privacy, fairness, autonomy, and accountability—demand different safeguards for each role.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Four-principle framework is load-bearing yet derived from an undisclosed, Western-centric corpus; if additional ethical dimensions are missed, the tailored-governance claim loses foundation.","rationale":"The reader identified the framework's representativeness as the weakest assumption; I agree. This is the single most load-bearing condition because the paper's contribution is explicitly to show how 'universal requirements manifest differently' across the three implementation approaches. If the universal set is wrong or incomplete, the tailored governance claims built on it are unfounded. The paper does not provide a systematic methodology and acknowledges the Western-centric source base, so the concern is not hypothetical. A concrete systematic-review test could settle it. I do not see a more serious internal inconsistency: the taxonomy is plausible, the examples illustrate the distinctions, and the claims in §4.2–4.4 follow from the stated framework. Therefore the reader's CONDITIONAL verdict remains appropriate; I would not escalate to rejection. The condition should require disclosed methodology and cross-cultural validation.","tokens_in":12222,"tokens_out":5555,"duration_ms":59299,"concrete_test":"Perform a PRISMA-guided systematic review of ethical principles for AI in mental health/counselling, deliberately including non-Western databases (e.g., LILACS, African Journals Online, CNKI) and grey literature, with two independent coders extracting all normative principles mentioned. Then test whether privacy, fairness, autonomy, and accountability cover 100% (or a pre-specified high threshold, e.g., 95%) of the extracted principles as first-order categories. If additional irreducible categories (e.g., cultural safety, relational autonomy, dignity, sustainability) emerge, the framework is not comprehensive and the paper's central claim requires qualification. For an internal check, map the 'additional considerations' listed in §4.1 (transparency, beneficence, non-maleficence, justice, trust, explicability, sustainability) onto the four principles; any that cannot be reduced without re","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim — that the three AI roles generate fundamentally distinct ethical risk profiles requiring tailored governance (Section 2.3) — depends on the premise that privacy, fairness, autonomy, and accountability are the correct and sufficient ethical framework (Section 4.1). This premise is asserted, not demonstrated: no search strategy, inclusion criteria, or coding methodology is reported for the 'comprehensive analysis,' and the Conclusion concedes the literature 'primarily draws from European and North American contexts.' The framework's universality is load-bearing because the entire subsequent analysis (Sections 4.2–4.4) is filtered through these four lenses. If a broader or non-Western corpus introduced irreducible dimensions — e.g., cultural safety, relational autonomy, dignity, or environmental sustainability (which §4.1 dismisses as 'elaborating' the four without argument) — the resulting risk profiles and governance recommendations would be incomplete. The paper itself admits emerging hybrid approaches 'will require entirely new ethical frameworks' (Section 3.4), which further undermines the claim that the four principles form a universal basis. This does not invalidate the descriptive taxonomy, but it weakens the normative conclusion that tailored governance based on these four principles is sufficient.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a conceptual analysis of ethical issues raised by AI systems in text-based online counselling (TBOC). It distinguishes three implementation approaches—autonomous counsellor/companion bots, AI counsellee simulators for training, and counsellor-facing augmentation tools—and argues that each produces a distinct ethical risk profile requiring tailored governance. It proposes four core ethical principles (privacy, fairness, autonomy, accountability) drawn from professional codes, regulatory frameworks, and selected literature, and applies them separately to each implementation approach. The paper concludes that TBOC's textual nature both creates opportunities for AI and demands implementation-specific safeguards.","tokens_in":12415,"tokens_out":3131,"duration_ms":38528,"significance":"If the central claim is accepted, the paper makes a useful contribution by moving ethical analysis beyond autonomous chatbots to comparatively neglected simulator and augmentation settings, and by organizing concrete ethical hazards around a small set of principles. The three-way taxonomy and the mapping of each principle onto each implementation approach are clear and well-illustrated with representative systems. The authors also explicitly acknowledge limitations: the source base is Western-centric, the analysis is conceptual rather than empirically validated, and emerging hybrid configurations may require further work. These candid statements strengthen the paper's credibility. However, the paper's normative weight depends on the claim that the four chosen principles are the correct and sufficient framework; this premise is not established with the methodological transparency the claim requires.","major_comments":[{"comment":"The derivation of the four principles—privacy, fairness, autonomy, accountability—is load-bearing, because all subsequent analysis in §§4.2–4.4 is organized around them. Yet the paper says only that a 'comprehensive analysis' of professional codes, regulatory frameworks, and scholarly literature reveals convergence; no search strategy, inclusion criteria, coding scheme, or method for synthesizing sources is reported. The Conclusion concedes that the literature 'primarily draws from European and North American contexts,' and §4.1 dismisses other candidate principles (transparency, beneficence, non-maleficence, justice, trust, explicability, environmental sustainability) as 'typically elaborating' the four without argument. As written, the sufficiency claim is asserted, not demonstrated. I request either a systematic methodology or a reframing of the claim as 'commonly prioritized principl","section":"§4.1, with §1 and §5"},{"comment":"The statement that emerging hybrid approaches 'will require entirely new ethical frameworks' conflicts with the paper's earlier claim that the four principles form a stable basis for analysis. If entirely new frameworks are needed, then the tailored-governance conclusion based on the four principles does not generalize beyond the three established approaches. The paper should reconcile this tension—for example, by specifying whether the four principles are meant to be necessary but not sufficient, or whether 'new frameworks' are extensions rather than replacements. As it stands, the universality claim in §4.1 and the 'entirely new' claim in §3.4 are in tension.","section":"§3.4"},{"comment":"Two of the three representative systems used to illustrate the taxonomy are from the authors' own group (VirCo in §3.2, CAIA in §3.3). This is not by itself a problem, and the paper does occasionally flag conflicts of interest for external systems (e.g., Woebot, Wysa), but the authors' own systems are presented as neutral evidence of the categories. Please explicitly disclose that these are the authors' systems and discuss how selection of representative systems was made. Otherwise the taxonomy risks being shaped by convenient, self-authored examples.","section":"§3.2, §3.3, references [46], [52], [13]"}],"minor_comments":[{"comment":"The text refers to a '2024 addendum' from WHO, but the cited reference [61] is dated 2025. Please harmonize the in-text date and the reference list.","section":"§4.1 / References [60], [61]"},{"comment":"The abstract says 'Textual constraints may enable AI integration'; this phrasing is slightly paradoxical. Consider clarifying that the absence of non-verbal cues reduces the input complexity for AI, while the same constraints create new risks.","section":"§1 / Abstract"},{"comment":"The claim that 'current validation studies often measure simulator performance rather than subsequent real-world competence' is important but unsupported by a citation. Please add evidence or soften the wording.","section":"§4.3"},{"comment":"The sentence 'Empirical studies reveal persistent implementation gaps' is followed by references [4] and others. Consider giving more detail on which studies reveal which gaps, to strengthen the basis for the gap statement.","section":"§2.3"},{"comment":"The word 'systematically' is used in several places (e.g., §2.3, §4.1) to describe the analysis, but no systematic method is reported. Using 'systematically' may overstate methodological rigor; consider 'comprehensively' or 'in a structured manner.'","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a worthwhile conceptual contribution with a clear taxonomy, but the central ethical framework needs stronger methodological grounding or more modest claims. The authors' willingness to state limitations is welcome; the revision should act on those limitations rather than only acknowledge them. I see no insurmountable flaw, provided the load-bearing derivation of the four principles is either made transparent or reframed in terms of the reviewed sources."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper earns its place. The taxonomy of autonomous bots, training simulators, and counsellor-facing augmentation tools is well thought out, and the ethical mapping is concrete enough to be useful to developers and practitioners. The focus on simulators and augmentation is a real gap in the literature, and the authors make a solid case that these roles produce different risk profiles—not identical ones. I also like that they flag conflicts of interest in the trials they cite (Woebot, Wysa) rather than taking the evidence at face value. The paper is clearly written and, for a conceptual piece, unusually grounded in real systems.\n\nThe central claim—that different implementation approaches generate distinct ethical challenges requiring tailored governance—holds up. It does not depend on the four principles being exhaustive; the authors themselves go beyond privacy, fairness, autonomy, and accountability in the detailed sections (competence gaps, ecological validity, skill atrophy). So the stress-test concern about a load-bearing universal framework is overstated. If anything, the four principles are an organizing device, and the paper's real value is in the concrete per-role analysis.\n\nThat said, the soft spots are real. The 'comprehensive analysis' from which the four principles emerge is not a systematic review: no search strategy, no inclusion criteria, no coding scheme. For a conceptual paper that may be acceptable, but the authors should say so explicitly and avoid claiming comprehensiveness. The Western-centric source base is acknowledged, which is good, but the dismissal of other principles (transparency, beneficence, etc.) as mere elaborations is asserted, not argued. And the self-referential examples—VirCo and CAIA are their own systems—should be disclosed as a potential conflict of interest. None of these are fatal, but they are the kind of thing a referee should push on.\n\nYou could also point out the tension between the claim that the four principles are universal and Section 3.4's assertion that emerging hybrid approaches 'will require entirely new ethical frameworks.' That is a legitimate internal inconsistency, but it is more a rhetorical overstatement than a broken argument.\n\nWho is this for? Anyone working on AI ethics in mental health, especially in text-based counselling settings, will get value from the taxonomy and the per-role discussion. It deserves a serious referee, not a desk reject. I would send it to peer review with requests for (a) methodological transparency about how the principles were selected, (b) disclosure of the authors' own systems as examples, and (c) toning down the comprehensiveness claims. With those revisions it could be a useful, citable contribution.","headline":"A genuinely useful conceptual mapping of three AI roles in text-based counselling, with a defensible central claim; the main soft spot is the hand-wavy derivation of the four-principle framework, but that does not sink the paper.","tokens_in":12934,"tokens_out":2231,"would_cite":true,"duration_ms":24386,"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":"The paper claims that the ethical landscape of AI in text-based online counselling splits into three distinct territories according to where the AI sits in the counsellor-counsellee dyad: as a replacement for the counsellor, as a stand-in f","keywords":["AI ethics","text-based online counselling","conversational agents","human-AI collaboration","responsible AI","privacy","fairness","autonomy"],"falsifier":"A documented case where an AI counselling tool caused harm that cannot be mapped onto any of the four principles—for example, a relational or cultural harm arising purely from the loss of shared humanity—would undercut the claim that these four principles are sufficient. A systematic review of non-Western mental health AI deployments would provide the clearest concrete test.","tokens_in":12051,"feed_emoji":"💬","tokens_out":4402,"duration_ms":45852,"temperature":0.7,"pith_summary":"This paper argues that AI in text-based counselling is not one thing: autonomous bots, training simulators, and counsellor-facing assistants each create a different ethical burden. It shows that the same four principles—privacy, fairness, autonomy, accountability—appear in all three, but the concrete hazards differ: bots risk fatal misjudgement, simulators risk biased or ineffective training, and assistants risk hidden influence and professional dependency. The paper's contribution is a structured map for thinking about these differences, so that governance can be tailored to the specific role rather than applied generically. A sympathetic reader would come away with the insight that where an AI sits in the counselling relationship determines what the ethical problem actually is.","feed_headline":"AI counselling ethics change with the AI's role","feed_subtitle":"Bots, simulators, and AI assistants raise different privacy, fairness, autonomy, and accountability questions.","key_machinery":"The organising device is a taxonomy of three AI implementation approaches defined by their position in the counsellor-counsellee interaction: (1) autonomous counsellor or companion bots, which substitute for the human counsellor; (2) AI counsellee simulators, which stand in for the counsellee in professional training; and (3) counsellor-facing augmentation tools, which operate invisibly in the counsellor's workspace. The analysis works by asking, for each approach, how each of the four principles (privacy, fairness, autonomy, accountability) translates into concrete requirements when that approach processes sensitive personal data during moments of vulnerability.","core_discovery":"On the paper's own terms, the central discovery is a role-based mapping of ethical risk in AI-assisted text-based counselling. Examining three established implementation approaches—autonomous counsellor bots, AI counsellee simulators, and counsellor-facing augmentation tools—the paper shows that each placement within the counselling dyad reconfigures the relationship and therefore the meaning of privacy, fairness, autonomy, and accountability. Autonomous bots carry the heaviest burden, with fairness failures becoming potentially fatal and accountability lacking a clear human subject; simulators reorient ethics away from protecting counsellees toward the integrity of professional education; a","pith_inferences":["The same role-based logic could extend to other text-based helping professions, such as crisis hotlines or online education, where the distinction between replacement, simulation, and augmentation may predict distinct ethical hazards.","The paper's framework implies a practical audit design: for each of the three roles, a separate checklist derived from the four principles, usable by ethics reviewers before deployment.","Because the source base is largely European and North American, testing the framework against non-Western counselling traditions, including community-based and family-oriented care models, would either strengthen or revise the four-principle foundation.","The three roles can be read as points on a spectrum of AI autonomy; if that spectrum holds, regulatory risk-tiering (like the EU AI Act's use-case tiers) could be mapped onto it, though the paper does not itself draw this regulatory conclusion."],"forward_implications":["Regulators and developers should stop treating 'counselling AI' as a single category; risk tiers should follow the AI's role in the counselling dyad.","Autonomous counsellor bots require fail-safe handoff protocols so that a counsellee in crisis can always reach a human, even if the system itself fails.","Training simulators need outcome validation: evidence that skills practiced on synthetic counsellees actually transfer to real counselling, not just high realism scores.","Augmentation tools should include monitoring for automation bias, since counsellors may unknowingly adopt AI suggestions under time pressure.","Consent procedures for augmentation tools may need to inform counsellees that AI processes their messages, introducing transient processing where possible."],"fun_headline_variants":["Ethics shift with AI's role in counselling","Three AI roles, four ethical tests in therapy","How AI's seat in therapy changes its ethics","Counselling AI: ethics depend on the job","Bots, simulators, tools: ethics follow the role"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire analysis rests on the premise that privacy, fairness, autonomy, and accountability capture all ethically relevant dimensions of AI in text-based counselling, drawn from a review focused on European and North American sources.","fun_headline_variants_meta":{"raw":{"variants":["Ethics shift with AI's role in counselling","Three AI roles, four ethical tests in therapy","How AI's seat in therapy changes its ethics","Counselling AI: ethics depend on the job","Bots, simulators, tools: ethics follow the role"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000101,"raw_usage":{"total_tokens":795,"prompt_tokens":619,"completion_tokens":176,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":363,"completion_tokens_details":{"reasoning_tokens":111}},"tokens_in":363,"tokens_out":176,"duration_ms":3141,"temperature":1.0,"reasoning_tokens":111,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T11:03:34.527048+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A documented case where an AI counselling tool caused harm that cannot be mapped onto any of the four principles—for example, a relational or cultural harm arising purely from the loss of shared humanity—would undercut the claim that these four principles are sufficient. A systematic review of non-Western mental health AI deployments would provide the clearest concrete test.","supporting_citations":[],"review_version":1}