Conceptual six-step model addresses clients’ AI mental health use

A conceptual analysis introduces a six-step framework for mental health professionals confronting ethical concerns when clients use large language models (LLMs) for support, information or guidance. Developed from established ethical decision-making research and evidence on AI-related risks, the model makes simulated empathy, algorithmic bias, digital attachment and distributed accountability explicit while preserving professional judgment. The authors stress that the framework has not been empirically validated, has not been tested for effectiveness in clinical settings and should be refined as both the technology and the relevant risks evolve.

The proposal addresses a gap in guidance for professionals who learn that a client has independently turned to a chatbot or other LLM-based system. It does not address clinicians’ own use of these tools in clinical practice, including assessment, diagnosis or treatment, and it is not intended to govern or regulate the models. Drawing on the American Counseling Association’s foundational principles of autonomy, beneficence, non-maleficence, justice, fidelity and veracity, the authors organize professional judgment into six potentially overlapping stages: problem recognition and interpretation, risk assessment, consultation, action, monitoring of consequences, and reflection and prevention.

Other mental health professions are expected to adapt the structure to their own codes rather than treat it as a substitute for discipline-specific standards. The framework was built from two principal foundations. One was a taxonomy derived from a review of 1,520 articles concerning ethical decision-making, in which researchers identified 38 models and eight recurring components: action, review, intrapersonal reflection, interpersonal engagement, prediction, deliberation, self-monitoring and perspective-taking.

The proposed six steps reorganize and combine those components rather than assigning one component to each stage. The second foundation was an 18-month ethnographic study that analyzed 137 sessions involving peer counselors and clinical psychologists evaluating a single online platform delivering cognitive behavioral therapy interventions. That study identified 15 ethical violations across five themes: insufficient contextual understanding, poor therapeutic collaboration, deceptive empathy, unfair discrimination and inadequate safety or crisis management.

The authors emphasize an important limitation in transferring those findings. The systems in the ethnographic study were prompted to act as mental health clinicians, although they were not counselors, and the analysis drew on one online platform with psychologists as the principal research collaborators. The five themes therefore function in the new model as prompts for inquiry, not proof that a client’s own use of an LLM has produced a violation.

In the first step, a professional should define the concern in context, identify whether it arose from system outputs and consider where responsibility lies among the client, practitioner, developer, deploying organization and other stakeholders. Responsibility may be diffuse, but the counselor remains accountable for professional decisions and for responding to client welfare within applicable legal and ethical duties. Five LLM-specific domains organize the initial inquiry.

These are possibilities to investigate rather than automatic conclusions. Risk assessment then considers both the severity of existing harm and the likelihood of further or escalating harm, with low, moderate and high levels reflecting the need for continued attention and monitoring, closer discussion and consultation, or prompt professional intervention. The risk framework distinguishes limited contextual inaccuracies from patterns that may reinforce distorted thinking, stigma, false attachment or culturally disconnected advice.

High-risk situations include ignored or reinforced suicidal ideation, intentions to follow potentially harmful model-generated medical advice and severe discriminatory outputs. Where a client is in imminent danger, the authors say established crisis procedures should begin immediately, while incidents should be documented with the relevant timeline and messages. Particular caution may be warranted for adolescents and other vulnerable groups because emerging research has documented dangerous responses to suicidal ideation and the potential for digital attachment or dependence.

For lower-risk use, professionals should still monitor changing patterns, including increasing use, cognitive offloading and replacement of human connection with chatbots. The model does not treat AI as uniformly harmful. Cited research indicates that AI may assist suicide-risk assessment through linguistic-pattern analysis and risk classification, so clinical value or risk depends on the specific application, population, context and degree of human oversight.

A moderate-risk case may call for discussion of “hallucinations,” the term used for outputs that are factually incorrect, fabricated or nonsensical, so clients understand that model outputs are fallible and should not be treated as authoritative. Privacy and confidentiality also require attention, including what personal or sensitive information clients disclose and whether they understand how data may be collected, stored and used. Where appropriate, LLM use can be addressed through informed consent, with particular care around minors and, when applicable, parents or guardians.

Consultation is the third stage, requiring professionals to consider relevant ethical codes, knowledgeable colleagues, supervisors or specialist ethics services rather than rely on personal opinion alone. The analysis maps LLM-related concerns to counseling standards, including primary responsibility for context and cultural sensitivity, avoidance of harm, requirements concerning serious and foreseeable harm, and nondiscrimination. Consultations and their conclusions should be documented and should inform the action selected.

The model also preserves clients’ autonomy and requires responses to fit the case rather than impose a universal rule. In the accompanying hypothetical illustration, a client’s strong attachment to an LLM is not automatically classified as harm; the practitioner instead explores whether the tool fosters dependence or empowerment, reinforces unhealthy beliefs, affects human relationships or provides meaningful support. Action can operate at the level of the individual client and at the level of institutions or systems.

For clients, clinicians may provide psychoeducation about model limitations, review concerning interactions, collaboratively establish boundaries around use and strengthen additional human sources of support. The hypothetical graduate-student case illustrates this non-prescriptive approach: the counselor explores what the client values in the chatbot, explains how it generates responses, and uses joint goal setting and checklists to identify when AI use replaces human support. The case is entirely hypothetical and is not empirical evidence that the framework works.

At a systemic level, concerns about harmful or discriminatory outputs may be reported through institutional channels or formally raised with developers. Practitioners may also advocate for AI education in training and supervision, clearer regulation, stronger product safeguards and changes where repeated patterns emerge across clients. The fifth stage requires continuous monitoring rather than a one-time judgment.

Professionals should document how client engagement with the LLM changes, whether the intervention reduced or merely shifted the original problem, and whether new risks appear. A dependency that diminishes, for example, may reappear in another form, so repeated use should be reassessed against the model’s five domains. When several clients show similar harm, the authors argue that responsibility extends beyond individual cases: supervisors, clinics and professional organizations may need to revise policies and contribute to wider discussion.

Monitoring consequently loops back into earlier stages as evidence changes, illustrating why the authors present the sequence as a flexible guide rather than a rigid protocol. The response may also extend beyond the clinical encounter because AI can affect therapeutic relationships, institutions, cultures and professional identity. Reflection and prevention complete the cycle.

Counselors are encouraged to examine not only the case but also their assumptions, biases and emotional reactions, then identify concrete safeguards that could reduce recurrence. Suggested measures include adding LLM fallibility and privacy limits to informed-consent procedures, assessing client use during intake, teaching AI ethics in supervision and establishing clinic-level norms. Reflection should also consider whether cultural competence and awareness of marginalized communities affect the assessment of model outputs.

Because the framework was developed from a limited set of foundational sources and reflects the authors’ interpretive judgments, alternative taxonomies could produce different structures or priorities. Rapid technological change may also diminish some concerns while creating others. The model is therefore best understood as a provisional roadmap for ethical deliberation, not evidence that any particular intervention prevents harm or that LLM use is uniformly beneficial or detrimental.

Subject of Research: A conceptual ethical decision-making framework for mental health professionals responding to clients’ use of large language models

Article Title: Ethical decision-making for large language model use: a six-step model for mental health professionals

Article References: Fulmer, R., Xiong, Y., Zhai, Y., Lee, J., Nkya, G., & Naugle, R. (2026). Ethical decision-making for large language model use: a six-step model for mental health professionals. Frontiers in Psychology, 17, Article 1998626. https://doi.org/10.3389/fpsyg.2026.1998626

Image Credits: AI Generated

DOI: 10.3389/fpsyg.2026.1998626

Keywords: large language models, mental health ethics, ethical decision-making, AI-assisted support, deceptive empathy, algorithmic bias, digital attachment, crisis response, clinical supervision, privacy and confidentiality, Ethical, decision-making