Therapy session notes with AI, without breaking confidentiality
Can a psychologist hand session notes to an AI? Yes, as long as it does not know who they are about and keeps nothing. A week in a private practice, followed closely.
"Am I actually allowed to put my session notes into an AI?" The question comes up whenever psychologists meet for peer consultation, often in a lowered voice. Because the notes pile up: a few words scribbled between two clients, abbreviations you no longer understand three weeks later, and in the evening, the wish to write it all up in ten minutes rather than an hour.
The short answer: yes, if the AI does not know who you are talking about, if it processes the text in a setting you control, and if it keeps nothing. Rather than a list of principles, here is a working week of a fictional psychologist, Claire, a clinician in private practice, with the requests she actually makes and what she does with the results.
What a psychologist's duty of confidentiality covers
What is said in a session is probably the most intimate part of a person's life: their fears, their sexuality, their family conflicts, sometimes things they have never told anyone. The profession's codes of ethics make confidentiality a central principle, and criminal law punishes its breach: in Switzerland, Article 321 of the Swiss Criminal Code covers psychologists, and most other countries have equivalent professional secrecy rules. This information is also health data, and therefore sensitive data under both the GDPR and the Swiss FADP.
Pasting session notes into a consumer AI tool means handing them to a third party that often keeps the history, sometimes for a long time, and that may fall under foreign law. Even without a name, a session account can point to someone: an unusual job, a mid-sized town, a specific family event are enough.
One point needs to be clear before we start: IA Confidential is neither client record software nor a health data host. It keeps nothing from your sessions, and that is precisely why it should not become your archive. The record stays where you keep it today; the AI helps you write, nothing more.
A week of session notes in Claire's practice
Monday, 8 pm: setting things up once and for all
Claire creates an account and a space she calls "Sessions". In the space's Assistant tab, she writes her instructions, which will apply to every conversation she opens there: the outline of her notes (topics discussed, experiences reported, observations, working hypotheses, leads for the next session), a plain style, in the third person, with no unnecessary jargon. She adds a rule that matters to her: always keep the client's own words, in quotation marks, separate from her own interpretations.
In her profile, under Constraints, she states that she is a psychologist bound by professional secrecy. This profile is kept in her browser and is never sent to external models.
Finally, she checks the model settings: by default, her requests will be handled by open models installed on servers in Switzerland, which pass nothing on to third parties. To format notes, she needs nothing else.
Tuesday, 7 pm: three sessions to write up
It has been a busy day. For each session, Claire has a few lines in her notebook. Before typing them, she sorts through them: no name or initials, but a marker only she can link to her record, "client A". She also removes anything that would make the person recognisable without a name: the exact job becomes "works in the hospital sector", the town disappears, the sister's first name becomes "her older sister".
Then she opens a conversation in the "Sessions" space and writes the request she will reuse all week, the one you can copy:
Here are my raw notes from the sixth session with a client in their thirties, seen for panic attacks. Write them up following the space's outline. Keep word for word, in quotation marks, the sentences I noted in quotation marks. Do not add any hypothesis, diagnosis or lead I have not written. If an abbreviation is ambiguous, do not guess it: flag it at the end.
The structured note comes back within seconds, with two abbreviations flagged, which Claire fills in herself. She rereads it, corrects a phrase that had turned the client's hesitation into a statement, then copies the note into her record, where the person's identity is kept. She does the same for the other two sessions, each in its own conversation.
Wednesday, noon: preparing the afternoon session
Client B is coming back at 3 pm, for the tenth time. Claire wants the thread fresh in her mind. In a new conversation, she pastes her last four notes, already written up and with no identifying details.
Based on these four session notes, list the themes that recur from one session to the next and those that have disappeared. For each theme, say which sessions it appears in. End with three open questions I could keep in mind, without any clinical interpretation.
The overview reminds her that a conflict with a parent, very present at the start of the therapy, has not come up for three sessions. She may do nothing with it, but she knows it as she walks into the session. The suggested questions are ordinary; she keeps one, rephrased in her own way.
Thursday, 6 pm: the letter to the GP
A client has asked her to write to her GP to support a referral to a psychiatrist. Claire has her consent to share the relevant information, and only that.
Write a short letter from a psychologist to a GP, supporting the referral of a client to a psychiatrist. Information to include, and nothing else: in therapy for four months, persistent sleep problems, significant fatigue, request for an opinion on possible medication. Collegial tone, one page at most, no detail about the content of the sessions.
The assistant writes the letter as a document, which she can edit and then export to Word or OpenDocument. The layout is rebuilt from the text: Claire transfers the content into her own letter template and adds the client's identity, the date and her signature. The client's name never went through the AI.
Friday morning: preparing for supervision
On Fridays, Claire presents a case to her supervisor. For this presentation, she goes further in removing details: the age becomes a range, the profession disappears, life events are summarised without dates.
Based on these notes, prepare a case presentation for supervision: context of the therapy, how it has evolved over six months, what I find difficult in the therapeutic relationship, and the question I want to ask. One page, with no detail that could identify the person. If an element seems identifying to you, flag it instead of rewording it.
The assistant points out that a detail about the family's holiday spot could be enough, in a small community, to recognise the client. She removes it. That is exactly the kind of extra vigilance she wanted.
Saturday: leaving nothing lying around
The notes are in her record, the letter has been sent, the presentation is printed. Claire deletes the week's conversations with "Delete conversation": they only existed in her browser, and there is no reason for them to stay there. She keeps the space and its instructions for the following week.
The limits of AI-assisted session notes
AI is useful for structuring, harmonising and keeping track of a thread. It is not useful for doing your clinical thinking for you.
It can be wrong, and confidently invent things: attribute to the client a sentence they never said, harden a nuance, slip in a diagnostic term you never used. That is why Claire's request forbids any addition and requires word-for-word quotations, and why she rereads every note before transferring it. The note remains yours, and so does the responsibility.
Nor does it listen the way you do. What it picks up in a text are repetitions and absences; what plays out in the relationship, the silences, the transference, it does not see.
Finally, some situations gain nothing from going through an AI, even a confidential one: a child at risk, a disclosure of abuse, ongoing legal proceedings. In those cases, you write it yourself and get help from a colleague or a supervisor.
What stays confidential
With a confidential model, the notes you paste and the answers are processed on servers located in Switzerland, under the Swiss data protection act, and nothing is kept on our side. The history lives in your browser: we could not hand it over if asked, because we do not have it. And a deleted conversation no longer exists anywhere.
Nothing goes to an external model unless you have chosen it. If you select one and your message contains sensitive data, the Confidentiality filter steps in and offers to answer with a confidential model, message intact, or to send an anonymised version that you review first. For session notes, stay on the confidential models. Your profile is never sent to an external model.
The best protection is still the one Claire applied on Tuesday evening: what travels cannot identify the client. And if someone else also uses your computer, at the practice or at home, use a personal browser profile: the history there can be read by anyone who opens it.
Further reading
Doctors follow the same logic for their letters: see our guide on writing medical reports with AI. To know how far to strip down an account before handing it to a tool, read the difference between anonymisation and pseudonymisation, and keep the list of data never to paste into an AI close at hand.
To try Tuesday's request on a made-up session: open the chat.
