I participate in a Google Group on which a clinician recently asked, basically, “What do folks think of recording sessions and using AI to summarize those sessions for the purpose of generating session notes?”
First, a note on notes: my session notes tend to be 2-5 words. A typical session note: “Greed. Envy. Desire. Parents. Shame.” I had a career, prior to becoming a therapist, in finance, and one thing I learned quite early in my career is that everything I write down potentially may serve as evidence against me – or someone else. As a clinician, I don’t “use” session notes. I remember what my patients tell me, or, if I forget, that’s important. I never go back to my notes. Never, ever. Like, I literally don’t think I’ve ever done so. So. The sole purpose of my session notes is to satisfy the state, and/or insurance companies, in the event that one or the other ever asks to see them. I conceive of them as evidence a session happened, and that something happened in the session. No more, no less.
I once had my notes subpoenaed. The subpoena came from my patient’s attorney, oddly. I told my patient, “Um, you know, the notes belong to you. If you want them, you may simply ask me for them.” I tried to discourage my patient from doing so, and described them roughly as I just described them above. “They can’t possibly be useful to anyone, except for someone seeking to establish that my notes are somehow inadequate.” I failed. My patient asked for them. And was disappointed. My patient had hoped my notes would somehow be helpful in the context of ongoing litigation. They were neither helpful nor harmful, and never made their way into court.
Maybe I have hubris (I probably do), but I don’t fear regulatory sanction. The worst case I can imagine? I get audited by some regulatory agency, they tell me my notes suck, and tell me to do a better job going forward. ¯_(ツ)_/¯
Back to AI.
Because of how I keep notes, the question I repeated at the top of this post holds little interest to me. I will, though, describe a few of the ways I’ve found AI useful in my practice – at least one of which relates orthogonally to the question asked. They’re listed in descending order of the frequency with which I’ve used it.
Recording my dreams
I dream a lot, and I’m in four-day-a-week analysis. For years, I’ve recorded voice memos of my dreams in the middle of the night, and reviewed those recordings prior to my analytic sessions. Some years ago, I started using the Google recorder app to transcribe those dreams, and to send them to the (private) journal I maintain. That app does a passable job of transcription, but only passable. When OpenAI made its “Whisper” transcription service available to those using its API, I began using Whisper to transcribe the dreams, saving the recordings to a Google Drive folder, using Pipedream (a workflow automation tool) to check that dedicated folder, and having it submit the recordings to OpenAI for transcription. It did a much better job than my voice memo app of transcribing. Somewhere along the line, I refined my instructions. Initially, the instructions were:
Write a Title for the transcript that is under 15 words. Write a summary of the provided transcript, from the perspective of the speaker, in the first person, in bullet points, with each bullet point containing no more than 10 words. Then return a list of the people mentioned in the provided transcript. Then return a list of the emotions mentioned. Then return a list of actions mentioned. Then return a list of places mentioned. If and only if the speaker mentions any associations, return a list of those associations. Interpretation: Offer 3 tentative psychoanalytic interpretations of the dream- one Freudian, one Kleinian, and one that is developmental/genetic. Then, provide the entire transcript.
Predictably, the interpretations sucked, but the rest of the instruction was super-helpful. I’ve dispensed with the “interpretation” function, but I still use the rest nearly nightly.
Clinical notes
I don’t submit recordings of sessions to AI (but see below for what I do do in this regard). I do, though, often record brief voice memos after meaningful, interesting, or difficult sessions – notes to self about patients. I maintain a private “blog” on which my clinical thoughts reside, and these voice memos are – as are my dreams – automatically transcribed and submitted to that blog/journal. I don’t often look back at the journal (though I do, occasionally). Mostly, I use the voice memo as a sort of lower-transaction-cost journal – a place where, instead of typing, I speak my thoughts, and think as I speak.
Using ChatGPT generally
I often want to use ChatGPT while I’m walking, or while I’m nowhere near a computer. In these instances, I record a voice memo, and use a similar Pipedream script to submit my spoken instructions to the API to get a response. Recent examples included:
Brainstorming for an upcoming presentation I did – “What follows are some disorganized, rambling thoughts on a subject. Please organize my thoughts thematically, into an outline.”
Similarly: “I’m preparing a presentation on topic x. The presentation will be an hour long. These are some of the points I hope to make: [list]. These are some of the questions I want to raise: [list]. Please respond to this with a list of at least twenty additional questions that might be interesting to consider.”
Organizing a to-do list – “These are some of the tasks I want to accomplish today. Organize them thematically, into numbered lists.”
Preparing for clinical presentations
I had one patient whose sessions – with my patient’s permission – I transcribed verbatim, in real time, during session. I found it interesting, several times, to feed ChatGPT (via the Enterprise version, which is confidential and which OpenAI doesn’t use to train its models) the transcripts, and to ask:
You are a helpful assistant. What follows is a detailed transcript of several separate discussions between a patient and his therapist, on separate days.
First, provide a one-paragraph summary of the transcript.
Then, for each day of the transcript, provide a one-paragraph summary of the session, in the form:
Tue 2/13: [summary]
Wed 2/14: [summary]
And so on.
Then, itemize and describe briefly (in 1-3 sentences):
Each dream mentioned during the period of the transcript.
Each childhood memory recounted.
My patient’s general thoughts and feelings about their partner.
My patient’s general thoughts and feelings about work.Here’s the transcript:
I found the responses super-helpful to me in reviewing about a week’s worth of sessions at a time, particularly when thinking about presenting my work in a clinical context. It didn’t replace thought, but it certainly augmented it.
Thinking through questions
I often begin sessions with ChatGPT (or Claude, or Gemini – ChatGPT has been my preferred AI chatbot until quite recently, but Claude and Gemini definitely are catching up) by saying something like, “I’m a classically trained psychoanalyst. You are a very experienced analyst, and I want your help in thinking through a thorny question. Here’s a paragraph describing a clinical problem about which I’m thinking – the situation, and my thoughts at this point. Interview me, one question at a time, asking thoughtful questions about my patient and my thinking, and keep asking me questions until I tell you to stop, at which point, I’d like you to provide a brief summary of my thinking.”
This approach often pushes me to ask myself – and to answer – questions that might not otherwise occur to me. And, it often helps me flesh out my thinking. Note: I’m not asking ChatGPT to think for me; I’m asking it to spur me to think more.
