When not to use an AI scribe
An AI scribe is a tool, and like any tool it has a wrong setting. Knowing when not to use an AI scribe is part of using one well. The honest version of this technology is not “always on, always helping.” It is a device you reach for in some sessions and deliberately leave in the drawer in others, the way you decide whether to take a phone call during a session or let it ring. This post is the list most vendors skip: the moments where the right clinical choice is to keep the microphone off and write the note by hand later, or not generate a draft at all.
The client has not freely consented
Consent is not a checkbox you clear once at intake. A client can agree to recording in March and feel differently in June, when the material turns shameful or raw. If a client declines, hesitates, or asks you to turn it off mid-session, that is the end of the conversation about whether to record. No clinical upside outweighs a client feeling surveilled in the one room that is supposed to be safe.
Two practical notes. First, consent given under perceived pressure is not really consent. If a client senses that declining will inconvenience you, annoy you, or mark them as difficult, they may say yes while meaning no. Build the offer so the easy answer is “no problem either way.” Second, some clients will consent to transcription but not to audio retention. Honor the narrower permission. A tool like CouchNotes, which processes everything on the device and can auto-delete audio after a note is finalized, makes that promise easier to keep, but the discipline is yours, not the software’s.
The deeper question of who gets to decide, and what clients are actually agreeing to, deserves its own treatment. We go further in the ethics of AI in therapy notes.
High-acuity and crisis sessions
When a client walks in actively suicidal, dissociating, disclosing abuse, or in acute panic, your entire instrument is your presence. These are the sessions where a half-second of divided attention is felt. Setting up a recording, glancing at a screen, or holding any part of your mind on “is this capturing” is attention you cannot spare.
There is also a documentation argument, and it cuts against the scribe. Risk documentation is the most scrutinized writing you produce. It needs your specific clinical reasoning: what you assessed, what you ruled out, why you chose the disposition you chose. A draft built from a transcript summarizes what was said. It will not reliably capture what you thought and decided, and those are exactly the elements a reviewer, a board, or a court will look for. For high-acuity work, write it yourself, contemporaneously, in your own clinical voice. We lay out what belongs in that note in documenting risk and suicidal ideation.
A reasonable middle path exists: be fully present during the crisis, and if you want support afterward, dictate a structured summary into the tool once the client has left and you have stabilized your own thinking. The draft then serves your reasoning rather than replacing it.
Forensic and other legal contexts
Some work is shaped by the courtroom from the first minute. Court-ordered evaluations, custody assessments, fitness-for-duty work, and any session you can reasonably foresee being subpoenaed all change the calculus.
| Context | Why an AI scribe is a poor fit |
|---|---|
| Forensic evaluation | The transcript and any draft become discoverable; opposing counsel can scrutinize both |
| Custody / high-conflict | One parent may later contest what was recorded or how it was summarized |
| Anticipated litigation | A machine-assisted draft you edited can muddy what you actually observed versus what the model inferred |
The issue is not that a draft is inaccurate. It is that you now hold two artifacts, a transcript and a generated summary, and an opposing party gets to interrogate the gap between them. In forensic work, fewer intermediate artifacts is usually safer. None of this is legal advice, and the rules vary widely by jurisdiction and engagement type, so confirm the specifics with your attorney or the retaining party before any recording tool comes near the room.
More reasons when not to use an AI scribe: protecting rapport
This one is harder to name because the cost is invisible until it compounds. Some clients, and some moments, are sensitive to anything that signals you are managing a process rather than sitting with them.
Consider keeping the scribe off when:
- The therapeutic relationship is new and trust is still being earned.
- The client has a trauma history involving being watched, recorded, or monitored.
- The work is heavily transference-based or psychodynamic, where the room’s privacy is itself the intervention.
- You notice the tool changing how you listen, narrating for the transcript instead of tracking the person.
That last point is about you, not the client. If you catch yourself speaking more for the record than for the human across from you, the tool has begun to shape the session, and that is a reason to stop.
Candor is the trust-building move
The clinicians clients trust most are the ones who tell them where the edges are. It is tempting to undersell the technology to clients, or to oversell it to yourself. Both are mistakes. The honest framing is simple: this produces a draft that you read, correct, and sign; you are the author of record; and here are the sessions where I will not use it. Saying that out loud does more for the relationship than any feature.
A note on the privacy claim that often gets stretched here. Meeting privacy obligations is necessary but not the whole story, and it is never a reason to record a session a client does not want recorded. On-device processing changes the question, in that the audio of a hard session can stay on your machine and never touch a server, but it does not change whether you should be recording in the first place. That judgment stays with you.
A scribe earns its place by sparing you the documentation grind on routine sessions, so you have more attention for the ones that demand all of it. Used that way, the moments you turn it off are not a limitation of the tool. They are the clearest sign that you are still the clinician in the room, deciding case by case where a draft helps and where it has no business being. That decision is clinical work, and it stays yours.