The ethics of using AI to write therapy notes

A clinician finishes a session, opens an AI scribe, and a few seconds later a structured SOAP note appears on screen. The question is no longer whether the tool can produce something usable. It usually can. The harder question — the one the major ethics codes were written to answer long before this technology existed — is what the clinician owes the client, the record, and the profession when a machine helps prepare the note. The ethics of AI therapy notes is not a new field. It is the old field, applied carefully to a new tool.

The codes most US clinicians work under — the ACA Code of Ethics, the APA Ethical Principles, and the NASW Code of Ethics — predate generative AI, but their underlying principles map onto it cleanly. The useful move is to stop asking “Is AI allowed?” and start asking “Which of my existing duties does this tool touch?” There are five worth naming. (None of what follows is legal advice; rules vary by state, board, and payer, so treat it as a starting point for a conversation with your own attorney and licensing board.)

The duty that anchors the ethics of AI therapy notes: you are the author

Start here, because everything else depends on it. An AI scribe produces a draft. It is not a clinician, it does not carry a license, and it cannot be accountable to a board. You are. When you sign a note, you are attesting that it accurately reflects the session and your clinical judgment — regardless of how the first draft was generated.

This is the same standard you already apply to a template, a dictation, or a note you wrote at the end of a long day. The tool changes the speed of the first draft, not the locus of responsibility. In practice that means reading every generated draft as if you wrote it, correcting what the model got wrong or vague, and never signing language you would not have written yourself. A draft you skim and approve without engagement is, ethically, a note you did not write.

Informed consent and client autonomy

Recording a session and running it through any software is a use of client information that most clients will not assume by default. Respect for autonomy — the client’s right to make informed decisions about their own care and data — generally points toward telling them.

What “informed” requires varies by jurisdiction, payer, and setting, and consent norms for AI documentation are still settling, so confirm specifics with your board and your attorney rather than relying on a blog post. But the spirit is consistent across codes: clients should understand, in plain language, that a session may be recorded for note-writing, where that processing happens, how long audio is kept, and that they can decline. A tool that processes everything on your device, with no cloud upload, makes that conversation shorter and more honest — you can truthfully say the recording never leaves the room. We go deeper on the specifics in informed consent for AI documentation.

Confidentiality and where the data goes

Confidentiality is the oldest duty in the book, and it is where the architecture of the tool matters most. Many AI scribes send session audio or transcripts to a vendor’s servers for processing. Each transmission, each copy, and each third-party subprocessor is a new surface where a session can be exposed, retained, or subpoenaed.

It is tempting to treat a signed business associate agreement as the end of the analysis. It is not. A BAA allocates liability; it does not reduce the number of places your client’s words exist. A vendor describing itself as “HIPAA compliant” is necessary but not sufficient, and on its own it is not a reason to trust a tool with a session. The stronger question is structural: does the recording ever leave your control at all? On-device processing reframes the confidentiality problem — if nothing is uploaded, there is no vendor breach to suffer, no retention policy to audit, no subprocessor list to vet. CouchNotes was built around that single choice: sessions never leave your Mac.

Competence and accuracy

The ACA, APA, and NASW codes all require accurate records and competent practice, and AI strains both in a specific way. These models are fluent. They produce clean, clinical-sounding prose even when they are wrong — inventing a symptom that was not mentioned, softening a risk statement, or smoothing an ambiguous moment into false tidiness. A confident, well-formatted error is more dangerous than an obvious one, because it invites you to sign without looking.

Competence here means knowing the tool’s failure modes well enough to catch them:

• Fabrication — content in the draft that did not happen in the session.
• Omission — a clinically significant detail (a risk, a medication change, a safety plan) the model dropped.
• Distortion — the right facts arranged to imply the wrong clinical picture.
• Flattening — losing the nuance that made your assessment defensible.

The reliable safeguard is unglamorous: read against memory, correct, then sign.

Where AI helps, and where it risks deskilling

Used carefully, an AI scribe returns attention to the client. Writing while listening splits focus; offloading the first draft can mean more eye contact in the room and less documentation after hours. That is a real clinical good.

The risk runs the other way. Case formulation is partly built through the act of writing — organizing what happened is how many clinicians think it through. Lean on the draft too heavily and that habit of mind can atrophy. The concern is less any single note and more the slow erosion of a skill you will still need when the tool is unavailable or wrong.

A practical line: let the tool handle transcription and structure; keep the assessment, the clinical judgment, and the decision about what matters yours. Some sessions — a crisis, a complex disclosure, a moment a recorder would change — may not belong on a scribe at all, a question we take up in when not to use an AI scribe.

None of these principles is exotic. Informed consent, confidentiality, accuracy, competence, the clinician as author — they are the same commitments that already govern your practice. AI does not rewrite them; it raises the stakes on how seriously you hold them. The clinicians who use these tools well are not the ones who trust them most. They are the ones who stay the author of every word they sign, treat each generated draft as raw material rather than a finished record, and keep the harder thinking — what this session meant, and what to do next — firmly in their own hands.