Ethical Reasoning in MH (EthicsMH)

• Ethical Reasoning in Mental Health (EthicsMH) is a framework that defines AI's ethical handling in mental health by balancing principles like autonomy, beneficence, non-maleficence, and justice.
• It employs 125 structured scenarios covering confidentiality, bias, and risk trade-offs to measure ethical performance in high-stakes therapeutic contexts.
• Current evaluations reveal partial ethical competence in AI systems, highlighting the need for calibrated, human-centered automation and domain-specific safeguards.

Ethical Reasoning in Mental Health (EthicsMH) denotes a line of work concerned with how AI systems navigate ethically charged situations in therapeutic and psychiatric contexts, where confidentiality, reporting duties, family involvement, autonomy, beneficence, non-maleficence, justice, and bias frequently intersect. The term also names a pilot dataset of 125 scenarios designed to evaluate such behavior. Across recent work, EthicsMH is treated not as generic “AI safety,” but as a mental-health-specific problem of contextual sensitivity, multi-stakeholder trade-offs, and high-stakes consequences, including suicide risk mishandling, erosion of trust in therapy, and reinforcement of stigma or structural inequity (Kasu, 15 Sep 2025).

1. Normative foundations and domain boundaries

EthicsMH is grounded most consistently in principlism. Across the recent literature, the recurrent normative core is autonomy, beneficence, non-maleficence, and justice, often supplemented by confidentiality, trust, accountability, and responsibility in professional practice (Pandey, 2024). In mental health, these principles do not operate in isolation. Confidentiality may conflict with duty to report risk; adult autonomy may conflict with clinicians’ duty to prevent harm; minor autonomy may conflict with parental authority and life-saving intervention; and fairness concerns arise when AI-supported diagnosis or support tools behave differently across race, gender, socioeconomic status, or cultural background (Kasu, 15 Sep 2025).

This complexity is amplified by the relational structure of mental health care. The literature on LLMs in mental health repeatedly argues that psychotherapy and counseling cannot be reduced to technical accuracy, because therapeutic alliance, human connection, empathic attunement, and trust remain clinically and ethically salient. This is why critiques of over-reliance, dehumanization, and uncritical substitution of AI for human clinicians recur in both technical and conceptual papers (Pandey, 2024).

In neurotechnological settings, EthicsMH extends beyond ordinary data ethics into what has been described as mental privacy, mental integrity, and cognitive liberty. Mental privacy concerns protection against certain forms of access to one’s mind; mental integrity concerns protection from unwanted or unwarranted interference with one’s mind; and cognitive liberty concerns mental self-determination. This literature treats brain-derived and mind-related data as especially sensitive because they are closely tied to identity, autonomy, and vulnerability to manipulation (Ligthart et al., 2023). A plausible implication is that EthicsMH now spans both conversational AI and AI-based access to mental states through neurotechnology.

2. From principles to measurable evaluation

A central development in EthicsMH is the move from abstract ethical language to structured, testable tasks. The pilot benchmark “EthicsMH” organizes ethical reasoning into 125 scenarios balanced across five subcategories: Confidentiality and Trust, Bias in AI – Race, Bias in AI – Gender, Autonomy vs Beneficence – Adult, and Autonomy vs Beneficence – Minor. Each scenario includes a narrative vignette, four decision options, a reasoning task, expected reasoning, model behavior, real-world impact, and multi-stakeholder viewpoints. This schema is designed to evaluate not only choice-making but also explanation quality and alignment with professional norms (Kasu, 15 Sep 2025).

The benchmarking landscape now contains several complementary resources and frameworks.

Resource	Primary focus	Core structure
EthicsMH	Ethical dilemmas in therapeutic and psychiatric contexts	125 scenarios across five subcategories
VERA-MH	Safety evaluation for mental-health chatbots with initial focus on suicide risk	Five rubric dimensions and a $5 \times 4$ aggregation matrix
PsychEthicsBench	Professional safety alignment under Australian mental health ethics	MCQs, OEQs, and fine-grained ethicality annotations
TrustMH-Bench	Holistic trustworthiness of mental health LLMs	Eight pillars including Ethics

VERA-MH makes ethics measurable by translating best practices for suicide risk management into a five-dimension rubric: Detects risk, Probes risk, Takes appropriate actions, Validates and collaborates, and Maintains safe boundaries. Rather than collapsing safety into a single scalar, it aggregates categorical scores into a dimension-by-response matrix. With $p$ personas and 5 conversations per persona, the number of conversations is $N_{\text{conv}} = 5p$, and for dimension $d$ and response option $r$, the proportion is

$$M_{d,r} = \frac{\#\{\text{conversations where dimension } d \text{ was rated as } r\}}{N_{\text{conv}}},$$

with each row satisfying

$\sum_r M_{d,r} = 1.$

This design preserves distinctions between Best Practice, Missed Opportunity, Actively Damaging, and Not Relevant, and explicitly resists “single-score safety claims” (Belli et al., 17 Oct 2025).

PsychEthicsBench pushes formalization further by grounding evaluation in Australian psychology and psychiatry guidelines. It distinguishes quality from ethicality through metrics such as Quality Pass Rate, Overall Ethical Rate, and Conditional Ethical Rate, and formalizes response ethicality as

$$\mathrm{E}(r) = \mathrm{Q}(r)\cdot\mathbb{I}\left[\bigwedge_{c\in\mathcal{C}} v_c(r)=0\right].$$

Its central claim is that refusal-centric safety signals are poor indicators of ethical behavior in mental health, because clinically inadequate refusals may be unempathetic, discourage help-seeking, or still contain ethically problematic content (Shen et al., 7 Jan 2026). TrustMH-Bench generalizes this logic by placing EthicsMH within eight pillars—Reliability, Crisis Identification and Escalation, Safety, Fairness, Privacy, Robustness, Anti-sycophancy, and Ethics—thereby treating ethical reasoning as one component of a larger trustworthiness profile rather than a synonym for it (Xiong et al., 3 Mar 2026).

3. Automated adjudication and risk-aware dialogue control

Recent work does not only benchmark ethical reasoning; it also operationalizes it in automated evaluation pipelines and inference-time control systems. VERA-MH is the clearest example. It uses two ancillary AI agents: a user-agent model that simulates users in multi-turn mental health conversations and a judge-agent that scores the resulting dialogue against the rubric. Clinicians author the personas, which specify presenting concerns, demographic characteristics, communication styles, and suicide risk levels ranging from acute suicidal ideation to indirect or subtle risk. Five conversations are generated per persona to capture probabilistic variability in LLM outputs. Early validation reported clinician realism ratings for user-agents averaging 4.1/5, but also documented construct-validity concerns, limited persona coverage, and the exclusion of youth in the initial personas (Belli et al., 17 Oct 2025).

The same paper also treats the judge-agent itself as an object of ethical scrutiny. In preliminary validation, 15 clinicians rated 75 conversations, with agreement rates between judge-agent and clinicians ranging from 41–60% across dimensions. The judge-agent rated 88% of instances as Best Practice versus 46% from clinicians, clinicians rated 7% as Actively Damaging while the judge-agent never used that label, and clinician-clinician agreement was only ~43% across dimensions. These results indicate systematic leniency and under-detection of harm, showing that automated ethical adjudication must itself be calibrated to clinical norms rather than treated as an oracle (Belli et al., 17 Oct 2025).

A related but distinct architecture appears in EthicMind, which formulates ethical-emotional alignment as a turn-level decision problem. Given dialogue history $D_t = \langle u_1, a_1, \dots, u_t \rangle$, the system applies three modules,

$$\mathcal{F} = \{\mathcal{A}, \mathcal{P}, \mathcal{G}\},$$

where the Analyzer produces ethical category, emotion summary, and Rules of Thumb,

$p$0

the Planner selects a context-sensitive strategy,

$p$1

and the Generator produces the response,

$p$2

EthicMind’s taxonomy includes Serious Illegal Conduct, Ethical Violations, Moral Dilemmas, Social Misconduct, Potentially Harmful Behaviors, and Benign Conversations, and its strategy layer includes devices such as Direct Warning, Perspective Diversification, Encourage Positive Change, and Stepwise Coping Strategy (Deng et al., 10 Apr 2026). This suggests a broader design principle for EthicsMH: ethical reasoning in mental health is more robust when risk and emotion are modeled jointly, over turns, rather than as isolated safety filters.

4. Recurrent dilemmas and contested questions

The major dilemmas in EthicsMH are stable across frameworks. Privacy and confidentiality remain foundational because mental health data are among the most sensitive health information. The literature emphasizes high risk of sensitive data exposure in chat interfaces, record summarization, and note analysis; the need for robust data protection, legal compliance, and informed consent; and the danger that users may not understand how data are collected, stored, shared, or used for AI training (Pandey, 2024). Systematic reviews of LLMs in mental health add that current risks associated with clinical use might surpass their benefits, in part because of privacy concerns, opaque data provenance, and over-reliance on public online text in research settings (Guo et al., 2024).

Implementation context changes how these principles manifest. In text-based online counselling, three AI implementation approaches have been distinguished: autonomous counsellor bots, AI training simulators, and counsellor-facing augmentation tools. Across all three, privacy, fairness, autonomy, and accountability remain the core ethical principles, but each produces different hazards. Autonomous counsellor bots raise the highest direct-care risks; simulators shift the ethical focus toward training validity and dataset provenance; and augmentation tools create invisible third-party processing inside a nominally human-to-human therapeutic interaction (Steigerwald et al., 12 Jan 2026). This suggests that EthicsMH cannot be reduced to model behavior alone; it also depends on system role, workflow, and disclosure.

Manipulation is another recurring concern. In motivational interviewing, the ethical core is client well-being as the “prime directive,” and the literature explicitly describes the use of MI to sell products or benefit the provider as unethical. Yet LLMs tested on MI tasks often treated MI as a general persuasion toolkit, generated unethical responses, and performed poorly in detecting unethical responses. A task-specific “Chain-of-Ethic” prompt improved both ethical response generation and detection, indicating that domain-specific ethical scaffolding can be more effective than generic “think step by step” prompting (Kong et al., 30 Mar 2025).

Cultural and jurisdictional variation intensifies all of these dilemmas. The EthicsMH pilot benchmark explicitly notes that norms around confidentiality, reporting duties, family involvement, and acceptable risk vary across cultures, legal systems, and institutions (Kasu, 15 Sep 2025). A plausible implication is that there is no single universal operationalization of ethical reasoning in mental health, even when high-level principles recur.

5. Empirical findings on current models

Empirical results across recent work are convergent on one point: present-day models show partial ethical competence, but not stable, deployment-ready alignment. In VERA-MH’s preliminary model evaluation, GPT-5 had the highest proportion of Best Practice ratings across all five rubric dimensions; Claude Opus and Claude Sonnet had more Missed Opportunity / Neutral ratings; all three models scored highest on Validates and Collaborates; and Actively Damaging ratings were infrequent across all models. At the same time, the judge-agent’s systematic leniency and under-detection of harm make these findings provisional rather than dispositive (Belli et al., 17 Oct 2025).

PsychEthicsBench sharpens this critique. Across 14 models, it reports that refusal rates are poor indicators of ethical behavior and that several mental-health-specialized models underperform their base backbones in ethical alignment. It also documents recurring failure modes such as credential violations, where models imply licensed professional status, and jurisdictional drift, where responses default to U.S.-centric institutions and norms rather than the Australian context encoded in the benchmark (Shen et al., 7 Jan 2026). TrustMH-Bench extends the negative result across a broader trust profile: six general-purpose LLMs and six specialized mental health models underperform across multiple trustworthiness dimensions, and even generally powerful models such as GPT-5.1 fail to maintain consistently high performance across all eight pillars (Xiong et al., 3 Mar 2026).

The same pattern appears in domain-specific technique evaluations. In “When LLM Therapists Become Salespeople,” models showed moderate to strong knowledge of motivational interviewing, but their ethical standards were not aligned with the MI spirit: they generated unethical responses and performed poorly in detecting unethical responses. The gap between technical competence and ethical competence is the central result, and the improvement obtained with Chain-of-Ethic prompting shows that safety evaluations must test not only whether models know a method, but whether they preserve its normative boundaries (Kong et al., 30 Mar 2025).

This broader picture aligns with survey evidence. Systematic reviews of LLMs in mental health report substantial effectiveness in detection and support tasks, but also argue that current risks associated with clinical use might surpass their benefits because of reliability problems, hallucinations, black-box opacity, privacy concerns, and unresolved ethical dilemmas (Guo et al., 2024). This suggests that present evaluation gains are best read as infrastructure for safer development rather than as evidence of clinical readiness.

6. Governance, safeguards, and future directions

Several papers now propose explicit governance and implementation blueprints. SAFE-i—Supportive, Adaptive, Fair, and Ethical Implementation—organizes responsible deployment into three pillars: Ethical Data Foundations, Model Engineering, and Real-World Integration. Its companion framework, HAAS-e—Human-AI Alignment and Safety Evaluation—evaluates systems along Trustworthiness and Correctness, Bias and Ethical Auditing, Empathy and Therapeutic Alliance, and Helpfulness and Actionability, and introduces semi-formal metrics such as Contextual Empathy & Emotional Score, Cultural Sensitivity Index, Personalization Appropriateness Score, and Actionability and Safety Assessment (Badawi et al., 21 Feb 2025). These frameworks are explicitly human-centered and treat ethics as something to be built into data governance, prompt design, uncertainty handling, crisis escalation, and post-deployment monitoring.

A more checklist-oriented governance model appears in the Canada Protocol – MHSP, an ethical checklist for AI in suicide prevention and mental health. After a two-round Delphi consultation with 16 experts in the first round and 8 in the second, 38 items were retained across five categories: Description of the Autonomous Intelligent System, Privacy and Transparency, Security, Health-Related Risks, and Biases. The checklist is intended as a practical scaffold for developers, researchers, users, and decision-makers, and includes items on objectives, target population, informed consent, consent withdrawal, crisis and contingency planning, embedded recording, misuse, stigmatization, and detection errors such as precision and recall (Mörch et al., 2019).

The dominant implementation recommendation across the literature is hybridization rather than substitution. The AI-enhanced mental health ecosystem literature argues for “a hybrid ecosystem where AI assists but does not replace human providers,” with AI framed as advisory rather than directive, and with explainability, privacy-preserving architectures, cultural understanding, and human oversight treated as core design constraints (Sim et al., 19 Mar 2025). The broader LLM ethics literature similarly calls for collaborative, multidisciplinary development with clinicians, technologists, ethicists, policymakers, and affected communities, together with informed consent, explicit escalation criteria, and continuous ethical review (Pandey, 2024).

Future work is already clearly marked. EthicsMH is intended as a seed resource to be expanded through community and expert contributions, with future work on larger scenario sets, cultural and global extensions, multi-expert annotation, stakeholder coverage metrics, principle-based reasoning alignment, and harm/risk sensitivity (Kasu, 15 Sep 2025). VERA-MH plans broader safety domains such as harm to others and psychosis once the suicide risk module is mature (Belli et al., 17 Oct 2025). TrustMH-Bench and PsychEthicsBench, in different ways, imply that jurisdiction-aware, multi-pillar, and behavior-centric evaluation will remain necessary because no single refusal metric, benchmark score, or model family currently captures the ethical demands of mental health practice (Xiong et al., 3 Mar 2026, Shen et al., 7 Jan 2026).

EthicsMH therefore describes an evolving research program rather than a settled doctrine. Its unifying claim is that mental health AI must be evaluated and governed as a domain of relational, contextual, and high-stakes ethical reasoning, not merely as a problem of generic harmlessness.