17.3 Artificial Intelligence in Systemic Practice: Applications & Ethics

New Content, Directly Testable

The examination blueprint updated for 2026 administrations, built on the 2024-2025 role delineation, introduced artificial intelligence as explicit content. It appears in four places:

• Domain 2 description: assessing and diagnosing "including use of artificial intelligence (AI) as appropriate."
• Domain 3 description: designing and conducting treatment "including use of artificial intelligence (AI) as appropriate."
• Domain 4 description: evaluating process and terminating "and use of artificial intelligence (AI) as appropriate."
• Domain 6 title and tasks: task 06.01 requires professional responsibility and competence regarding legal issues and regulations across delivery modalities "including use of artificial intelligence (AI)," and task 06.09 requires written professional disclosures regarding providing technologically assisted services "including use of artificial intelligence (AI)."

Knowledge areas 71 through 75 supply the content: application of AI in clinical practice, implications of AI on clinical practice and systemic therapy, AI-powered tools in assessment, treatment planning, and administrative tasks including automated progress tracking, and ethical considerations including privacy, consent, and algorithmic bias.

An important framing point: the AAMFT Code of Ethics revised effective January 1, 2026 does not contain a standard specific to artificial intelligence. AI use is governed by the existing standards, principally Standard VI on technology-assisted professional services, alongside confidentiality, competence, and informed consent. Do not attribute an AI-specific standard to the Code.

Where AI Actually Shows Up

Note the pattern: the least risky uses are administrative and the riskiest are those that touch clinical judgment or that interface directly with a client in distress.

The Governing Standards

Standard 6.1 — Technology-Assisted Services. Before using technology to support treatment or supervision, therapists must determine that the technologically assisted service is reasonably appropriate for the client or supervisee, inform them of the potential risks, make reasonable efforts to secure confidential information transmitted or stored, and use the technology only after obtaining appropriate education, training, or supervised experience with it. That last clause makes AI competence a Standard III competence question, not merely a technology preference.

Standard 6.2 — Technology and Treatment Decisions. Therapists do not use technology in place of their own independent treatment decisions and decision-making process. This is the single most exam-relevant provision for AI. An AI system may draft, summarize, flag, or suggest; the clinical decision remains the therapist's, and the therapist is accountable for it. An option in which a therapist follows an algorithmic recommendation because the system produced it is a distractor.

Standard 6.3 — Technology and Consent to Treat. Informed consent must be obtained before treating through technology-assisted means, and clients must be informed in writing of the associated risks.

Standard 6.5 — Technology and Documentation. Technology used to generate, transmit, or maintain documentation containing confidential or identifying information must adhere to relevant professional and legal standards.

Standard 1.11 — Written Consent to Record. Informed consent and written authorization are required before recording images, video, or audio, or using any transcription services. An ambient AI scribe records and transcribes, so it falls squarely within this standard. Consent must be obtained from each individual competent to give it — which in couple and family work means everyone in the room.

Standard 2.2 and 2.5. Disclosure of confidences requires written authorization, and records must be stored, safeguarded, and disposed of in compliance with applicable law. Sending session content to a third-party AI vendor is a disclosure.

The Specific Risks

Confidentiality and the business associate problem. Any vendor that creates, receives, maintains, or transmits protected health information on behalf of a covered entity is a business associate, and a business associate agreement is required before use. Consumer-grade general-purpose AI tools typically do not offer one. Pasting session content into a public chatbot is a disclosure to a third party without authorization or safeguards, and it is the most common real-world violation in this area.

Fabrication. Large language models produce fluent, confident output that can be factually wrong, including invented clinical detail, misattributed statements, and citations to research that does not exist. In a clinical record, a fabricated detail is a permanent inaccuracy that may later be read by a court, a payer, or another clinician. Task 06.14's requirement of accurate records means every AI-drafted note must be reviewed and corrected before it is signed.

Algorithmic bias. Knowledge area 74 names bias in algorithms directly. Models trained on non-representative data can under-detect or over-detect risk in populations under-represented in training, and can encode diagnostic disparities present in historical records. A risk-flagging tool that was validated on one population may perform differently on another, and the therapist's obligation is to treat the output as one input, weighted by knowledge of the tool's validation.

Automation bias and deskilling. Clinicians tend to over-trust automated output and to under-question it, and the absence of a flag is easily read as reassurance. A negative AI risk screen never substitutes for the direct suicide assessment that tasks 02.03 and 05.02 require.

Client-facing tools and crisis. A chatbot recommended for between-session support may respond inadequately to a disclosure of suicidal intent. Before recommending any client-facing tool, know how it handles crisis content and tell the client explicitly what it is not.

Systemic reductionism. Most AI clinical tools are built on individual models of psychopathology. A tool that produces an individual diagnosis and an individual treatment plan can quietly pull a systemically trained therapist toward an individual formulation, which is a competence issue specific to this profession.

Disclosure: What Task 06.09 Requires

Task 06.09 lists what must be provided in written professional disclosures: fees, credentials, how to report complaints, confidentiality, privileged communication, and providing technologically assisted services including use of artificial intelligence.

A defensible AI disclosure states what tools are used and for what purpose; what data they receive and whether that includes session content; whether a business associate agreement is in place; whether sessions are recorded or transcribed, with separate written authorization under Standard 1.11; that the therapist reviews all AI-generated content and retains clinical decision-making authority; and that the client may decline AI-assisted documentation without affecting their care. That last element matters: consent that cannot be refused is not consent.

Two related obligations. Under Standard 6.6, technology may not be used to practice in a jurisdiction where the therapist is not legally allowed to practice, and state regulation of AI in health care is developing quickly, so licensees must track their own board's rules. And under Standard 4.6 and task 06.15, supervisees' confidences are protected too, so a supervisor putting a supervisee's session material through an AI tool has the same authorization obligations.