Using LLMs in Practice? A Six-Step Ethics Check for Clinicians

A downloaded consent form versus an actual decision process: only one of them tells you what to do the first time a model invents a citation inside a court-ordered report.

In October 2026, Frontiers in Psychology published a conceptual analysis by Fulmer, Xiong, Zhai, Lee, Nkya, and Naugle. The model is a six-step sequence written for working clinicians rather than researchers, and it arrived after the tools were already in use. LLMs are already in note drafting, intake summaries, and case conceptualization, often in practices with no written policy at all.

Treat the model as a scaffold, not a safe harbor. It only holds when paired with your board rules, HIPAA, and specialty guidelines.

Why AI Ethics for Mental Health Professionals Is Now a Scope-of-Competence Question

Scope of competence is not a synonym for a current license. It means you can explain, defend, and correct the work you sign your name to, including the tools that shaped it. When a large language model drafts a note, summarizes an intake, or suggests a differential, the clinician remains the author of record, and the competence question shifts to whether you can evaluate what the model produced.

What the ethics bodies have actually published

APA has the most developed public record: an August 2024 policy statement, Ethical Guidance for AI in the Professional Practice of Health Service Psychology issued in June 2025 and updated that July, a practitioner tool guide first published October 15, 2024 and updated October 25, 2026, plus a health advisory updated August 16, 2026 and a guide to AI-generated advice from June 2026, updated July 25, 2026. How much of it is enforceable rather than advisory varies by document. For ACA, NASW, AAMFT, and NBCC, available public sources did not establish comparable clinical AI standards as of October 2026. That is not proof none exist; check your own association's site and your board.

The gap is yours to close

Most codes predate generative LLMs. They are being interpreted, not amended, which leaves the gap between a code's silence and a model's output to your judgment. The six-step framework is a conceptual analysis, not an empirically validated instrument, and no ethics body has endorsed a single decision model. Treat it as scaffolding, not a safe harbor.

One action, not a slogan

Pick a tool you actually use and run a documented test: feed it a de-identified note and an intake summary, log errors, hallucinations, and privacy behavior, and take the results to consultation. Record the date, the tool version, and what you changed. That file is evidence of competence-building; "I stay informed" is not.

The Six Steps of the 2026 LLM Decision Model, Translated for Practice

The field has moved past asking whether clinicians use LLMs to asking how they can prove they used them responsibly. In October 2026, Frontiers in Psychology published a six-step model by Fulmer and colleagues that gives licensed professionals a repeatable sequence for that proof; the steps translate cleanly into ordinary clinical language.

The Six Steps, in Order

• Problem recognition: Name the specific LLM task and the ethical question it raises. A counselor drafting a progress note realizes the tool is being asked to infer a diagnosis, not just summarize.
• Risk assessment: Weigh harm, confidentiality, and scope before acting. A psychologist considering pasting an intake narrative into a chatbot stops when the narrative contains identifying details.
• Consultation: Seek licensing board guidance, a supervisor, or a colleague. An MFT trainee asks a supervisor whether using an LLM to brainstorm case conceptualization falls within supervised practice.
• Action: Proceed only with the least risky option that meets the need. A social worker uses a de-identified prompt to generate discharge-planning language, then edits it.
• Monitoring consequences: Track what the output changed in the record or the client's care. A clinician notices the model's suggested referral source does not exist and corrects the note.
• Reflective prevention: Update the practice policy so the same question does not recur. The team adds a "no identifiers in prompts" rule to its intake workflow.

Why the Sequence Matters

Risk assessment must come before action; reversing the order turns a decision into a cleanup. Monitoring is the step most practices quietly skip, which is why the model's real deliverable is a documented record of reasoning, not merely a defensible outcome. The authors map the steps to Johnson's components as a conceptual anchor, not a validated instrument.

A conceptual model improves consistency and auditability. It does not transfer accountability away from the clinician, and it should be paired with board rules, HIPAA, and specialty guidelines.

Stratifying Risk: Note Drafting and Intake Screening Are Not the Same Ethical Question

The 2026 six-step model includes an ethical concern stratification that separates low-exposure uses from high-exposure uses. The ordering below follows that logic: the more identifiable, clinical, or legal the material, the stronger the safeguard required before a clinician opens the tool. Because the model is conceptual rather than empirically validated, each safeguard should be checked against licensure board rules, HIPAA, and specialty guidelines.

Client Consent and Disclosure When LLMs Touch the Work, With Sample Language

Two dates matter here: New Mexico's AI informed-consent rule took effect November 18, 2025, and California now requires specific consent before any session is recorded or transcribed. Whether or not your board has acted, consent holds up every other step in this model.

What Sufficient Consent Actually Requires

Four elements, and a signature line covers only the last one:

• Capacity and comprehension: the client can restate, in their own words, what the tool does and what it does not do.
• Voluntariness: declining carries no penalty, no shortened sessions, no discharge.
• Documentation: the conversation, the version of the disclosure, the date, and the decision.
• Scope: consent names when, how, and which type of AI, per APA's 2026 guidance.

Original Sample Disclosure Language

Adapt freely to your setting:

"I use a large language model, a text-based AI tool, for limited tasks such as drafting progress notes and organizing assessment results. It does not make diagnoses, treatment recommendations, or decisions about your care, and it does not replace our sessions. AI output can be incomplete or entirely fabricated, so I review and edit everything before it enters your record. Only the minimum information needed is entered; I do not enter your name, identifying details, or sensitive disclosures beyond what the task requires. Information is processed by [vendor] under [retention terms]. You may decline AI involvement in part or all of your care, or withdraw consent at any time, without any change to your services or fees. I will not charge you for AI use or reduce our contact time because of it."

Revisit Triggers and the Forensic Exception

Renew the conversation when you add a tool, change vendors, shift to a new use case, or when retention terms change. ACA, AMHCA, and NBCC guidance treats this as ongoing communication, not a one-time form.

In court-ordered and forensic contexts, evaluees generally receive notification, not a veto: you disclose AI involvement in the record, but the evaluation proceeds. Say so plainly, in writing, at the outset.

Related Articles

Confidentiality, Privilege, and HIPAA: What Never Goes Into a Prompt

The confidentiality question around clinical AI has shifted in 2026 from whether clinicians will use these tools to what, precisely, they may paste into them. A practice policy built on the October 2026 Frontiers in Psychology six-step model should start with hard exclusions, not general caution.

Never in a prompt

Some content is categorically out: direct identifiers, obviously, but also contextual identifiers that narrow the pool (a rare diagnosis plus a date plus a small-town locale), any third-party information a client discloses about a partner, employer, or child, raw test data, and anything covered by 42 CFR Part 2.

The BAA, in plain terms

A consumer chatbot tier is not your business associate, whatever the marketing implies. PHI can move to a vendor only under a signed business associate agreement that sets permitted uses, bars the rest, demands appropriate safeguards and Security Rule compliance for electronic PHI, passes identical duties to subcontractors, and requires return or destruction of data at termination³. Breach terms vary more than clinicians expect: the rule allows up to 60 days after discovery¹, HHS model language tightens breaches to 30 calendar days and security incidents to five business days², and the stricter signed agreement governs.

Due diligence before the first prompt

• Retention: the default window, plus what happens to data after termination
• Training: whether inputs improve the model, and whether that is contractually barred
• Deletion: whether on-request deletion exists and is documented
• Subprocessors: who handles the data and in which jurisdictions
• Breach notice: the exact deadline written into your agreement
• Audit logs: whether account-level access logs are available

The de-identification myth

Removing names does not remove re-identifiability, least of all in rural communities or forensic populations where the fact pattern itself identifies the person. And privilege shields the clinician-client communication, not the vendor's separate copy of it.

Documentation, Board Rules, and Who Carries the Liability for an AI-Assisted Decision

Copy-ready note language

A progress note that survives scrutiny answers four questions in plain sentences:

• Tool and purpose: "Drafting assistance for this note was provided by [named tool], used solely to organize the clinician's existing session content."
• Limited scope: state what the tool did not do, e.g. "No diagnostic formulation, treatment recommendation, or risk assessment was generated by the tool."
• Verification: "Clinician reviewed all content against session record and clinical judgment; two statements about client's reported sleep were corrected."
• Attestation: "Clinician is the author of record and accepts full responsibility for the content of this note."

Place the attestation at the end of the note body, directly after the clinical content it covers. "Reviewed and edited" alone is too thin: it asserts a step without showing what was checked or changed. Name the specific elements you verified.

Jurisdiction, liability, and supervision

No single national rule governs AI documentation, so confirm your own board's requirements.¹ Texas Medical Board Rule 163.1 requires every amendment to a record to be clearly identified by date and author, with the original entry still legible, which argues for addenda rather than rewrites.² Mississippi's HB 1717, as introduced, would require clinician attestation of AI output before finalization plus tamper-resistant audit logs³; Arizona's behavioral health board requires documented informed consent before AI-delivered services.⁴

The liability answer is direct: the clinician of record remains the decision-maker and the responsible party.² Vendors do not absorb that duty, and no model becomes the author. In supervised cases, the supervisor owns the standard of care for how AI was used, so review the trainee's AI workflow explicitly. Malpractice carriers are now publishing risk-management guidance on AI documentation, and it rewards exactly this kind of record.

Forensic Psychology and AI Records: Discoverability, Test Data, and the APA Specialty Guidelines

Forensic practice is now the hardest case for AI policy in mental health. The APA Specialty Guidelines for Forensic Psychology are deliberately medium-neutral, and as of 2026 they say nothing explicit about prompts, chat logs, or model outputs, leaving practitioners to reason from principles rather than labels.

Assume it will be requested

Assume the prompt, the output, and every draft between them may be discoverable, and that opposing counsel will eventually ask. No 2026 rule creates a categorical privilege or work-product exemption for AI drafts. Guideline 10.06 is best read functionally: material created or exchanged in connection with the matter, considered in forming an opinion, or needed to understand your conclusion falls within documentation. Guideline 1.01 supplies the integrity anchor, requiring accuracy, honesty, and a clear line between what you verified and what the model produced.

What never enters a consumer model

Test items, raw scores, and evaluee-identifying detail stay out of consumer LLMs. Test security and data integrity are separate problems from HIPAA. Under APA Ethics Standard 9.04, test data (raw scores, responses, recordings) may be released on subpoena; manuals, instruments, protocols, and questions need a court order. De-identification is weakest where the stakes are highest, because evaluees from small populations (a rural court, a rare presentation, one clinic in a small county) can be re-identified from context alone.

If the subpoena arrives

Preserve rather than delete: platform, dates, prompts, uploaded sources, outputs, iterations, edits, and version histories. There is no AI-specific protocol, but 10.08 and EPPCC 6.01 require records that are accurate, current, and pertinent. Disclose what is responsive, and let the record already show that AI assistance was used and human-verified. A note written afterward to explain the gap is worse than the gap.

The model never holds the license; the clinician does, and the only place that accountability becomes visible, reviewable, and defensible is the note a human read before signing.

Hallucinations and Fabricated Citations: How to Verify Before It Reaches the Record

Fast draft or defensible draft: the difference is not speed, it is whether every sentence you keep will survive a records request.

The numbers are not marginal

A 2026 analysis of 300 references produced by three leading models found that 55.0% carried a verifiable citation inaccuracy and 28.3% were wholly fabricated.¹ On mental health topics specifically, a GPT-4o study of 176 citations found errors in 56%, with roughly one in five hallucinated. Junior clinicians in a 2026 multicenter study caught only 15.8% of hallucinations, and 13.1% caught none.² Fluency is the trap: a summary reads cleanly precisely because the model smoothed over what it did not understand. In one 2026 evaluation, answers matched the correct binary decision 90% to 98% of the time, yet evidence supported only 60% to 65% of them.

A verification protocol that fits a workday

• Citations: resolve every DOI or primary link before it enters a note or report.
• Statutes, codes, and board rules: check the section number against the official source.
• Diagnostic claims: compare against current criteria, not the model's paraphrase.
• Session summaries: confirm each fact against your own notes.

Where the boundary sits

The model may draft. It may not originate a clinical judgment, a diagnosis, or a risk formulation; those stay yours. If you cannot trace a statement to a source or to your session notes, it does not go in the chart.

If flawed output already influenced a decision

Document the catch, consult a supervisor or colleague, and correct the record. State what the tool produced, what you found, and what changed. An addendum is defensible; a silent edit is not.

A Copy-Ready Practice Policy Outline and the 'Jenny' Case as Supervision Material

A practice policy for large language models is a short written document that tells everyone in the office which tools are allowed, what may go into a prompt, and who signs off when something goes wrong.

A Nine-Part Outline to Adapt

Draft it with these headed sections: purpose and scope; permitted uses; prohibited uses; approved tools and vendor requirements; consent and disclosure; documentation expectations; supervision duties; incident response; and annual review. Each section should name a responsible person, not just an expectation. Keep it to two pages, because a policy nobody reads is not a policy.

Using the Jenny Case in Supervision

The paper's hypothetical Jenny case walks a clinician through LLM-specific ethical choices rather than presenting one correct answer. Use it in supervision, then ask:

• Which step of the six does this decision trigger first?
• What would Jenny be told, in plain language, before any tool is used?
• Which parts of the record would the tool see, and which must stay out?
• If the output is wrong, who catches it, and where is that check documented?
• Does a board rule or agency policy already answer this, or is it genuinely new?

A Peer Consultation Exercise

Bring one de-identified use case to a peer group and walk it through the six steps out loud. Compare where members disagree. The disagreements are the training targets.

Any draft policy needs review by counsel and against your own board rules before adoption. The model is a scaffold, not a substitute for either.