AI for Group Therapy Practices: Notes, Consent and Review Steps

AI can draft group therapy notes without exposing other members only if each member's note is written separately from that member's own information, other members appear as "another member", and a clinician reviews every note before it is filed. Recording or processing a group session also needs each member's informed consent, and one refusal usually changes the plan.

A group is different from individual therapy in one way that matters for notes. In a one-to-one session, the client's own words are the whole record. In a group, a member's note can easily contain what someone else disclosed, and that person never agreed to appear in a stranger's file. The tool does not know that. It will faithfully copy "[first name] described their divorce" into the note of someone who was only listening. The rules below stop that happening.

Follow me on Instagram@sagnikteaches

The per-member note rule

Write the rule on a card, not in your head. Each member's note contains only three kinds of content.

Connect on LinkedInSagnik Bhattacharya

1. That member's own contributions: what they said, did, reported, or declined to share, in their own words where it matters.
2. That member's observed behaviour and engagement: attendance, participation, mood as observed, response to exercises.
3. Interactions, described without identity: "responded to another member's disclosure with support", never "[first name]'s disclosure about their divorce".

What never appears in a member's note: another member's name, initials, job, age, family details, diagnosis, story or any detail that would let someone work out who they are. This is also a good test for your human review: if a member asked to see their own note, would anything in it reveal someone else? If yes, it fails.

Subscribe on YouTube@sagnikteaches

If you keep a separate group-level record, such as the session plan, topic covered and attendance, keep that free of individual clinical content too. Mixed records cause the most trouble when a member requests access to their file, or when notes are subpoenaed or shared with another provider. The details depend on your rules, so ask your compliance adviser how your regulator treats group session records, and whether any of your notes count as a specially protected category of therapy notes that must be kept apart from the main record. One widely used health privacy rule, for example, defines psychotherapy notes as a mental health professional's notes documenting or analysing the conversation in a private, group, joint or family counselling session, kept separate from the rest of the record. It excludes session start and stop times, medication, the type and frequency of treatment, test results and any summary of diagnosis, functional status, treatment plan, symptoms, prognosis or progress. By that test, the per-member progress notes in this tutorial (attendance, engagement, plan) belong in the main record, while a facilitator's separate process jottings might qualify. Whether yours do is a question for your adviser, not an AI tool.

What goes into the tool, and what does not

There are three ways to produce a draft, from safest to least safe. Choose the lowest-risk one that does the job.

Most small practices should start with the first row. It captures most of the time saving, because the tool's value is in turning ten minutes of bullet points into eight clean, structured notes, not in listening to an hour of talk. Recording is a larger decision with its own consent, retention and access questions. If you go that way, the recording versus dictation comparison sets out the trade-offs.

Be careful with claims about specific plans. Business plans of the main chat tools don't train on your content by default, and consumer plans let you switch off training in settings, but neither fact makes any plan appropriate for health information by itself. What decides it is a signed data agreement where your law requires one, your regulator's position, and your insurer's. Ask the vendor in writing, and keep the reply. Don't rely on a web page, including this tutorial, to say a plan is approved for therapy.

A note template that enforces the rule

A fixed template makes the per-member rule mechanical. One template, filled eight times, gives you eight notes with the same shape.

GROUP SESSION NOTE (one member)
Member label: M3 (real name entered in the record system only)
Group: [group name], session [number], [date]
Attended: yes / no / late (minutes)
Topic of the session: [from the session plan]
Member's contributions: [only what M3 said or did]
Observed engagement and mood: [observed, not inferred]
Interactions with the group: [described without identifying anyone]
Risk observations: [none observed / describe; clinician completes]
Plan for next session: [only for M3]
Clinician review: [name, date, time]

"Risk observations" and "Clinician review" are filled by the clinician, never by the tool. A note with a blank review line is a draft and cannot be filed.

The prompt: one member at a time

Run the tool once per member, in a fresh chat each time, with only that member's jottings. This is the key step. If you paste all eight members' jottings into a single prompt and ask for eight notes, one member's details will leak into another's, and you will never be sure you caught every leak.

Write a group therapy session note for ONE member using the template below.
Use only the jottings I give you. If something is not in the jottings, write "not recorded".
Do not mention any other group member by name, label or any identifying detail.
Refer to others only as "another member" or "the group".
Do not infer feelings, diagnoses, causes or risk. Do not add advice or a plan.
Quote the member's own words only where I have put them in quotation marks.

TEMPLATE:
[paste template]

JOTTINGS FOR THIS MEMBER ONLY (label M3):
- Arrived 5 min late, apologised to the group
- Spoke about week at work being "exhausting", mentioned sleeping poorly
- Declined the breathing exercise, watched
- Offered a supportive comment after another member spoke
- Asked whether next week's session can start earlier
- Facilitator note: seemed quieter than last week

Illustrative output: a filled template in which M3's note says they arrived five minutes late, described a tiring week at work with poor sleep, declined the breathing exercise but stayed engaged as an observer, offered support to another member, and asked about an earlier start. "Risk observations" says "Not recorded: clinician to complete". "Seemed quieter than last week" is carried as the facilitator's observation, attributed as such, not turned into "low mood".

What you would fix: the tool might write "a supportive comment after another member's disclosure about their job". That adds a detail (the topic of the other member's disclosure) the jottings did not give, and it could identify someone in a small group. Cut it to "after another member spoke". Rule of thumb: if a phrase in an interaction line could only be true for one person in the room, delete it.

Consent: what to ask and how to word it

Consent for using AI in group notes has two layers, and they are separate. Layer one covers the member's own information in their own notes. Layer two, if you record, covers the member's voice and disclosures being captured and processed alongside everyone else's. Treat them as two questions, with two answers on file.

Keep the wording plain and specific. Illustrative wording, to adapt with your own adviser, not to copy as legal text:

"After each group session I write a short private note about each person's participation. To save time, I sometimes use an AI writing tool to tidy my own jottings into the note. I use labels instead of names and I never include other members' details in your note. I read and correct every note myself before it is filed. You can say no, and saying no will not affect your place in the group. I'll write your note by hand instead. You can change your mind at any time."

For recording, use a separate form that states: what is recorded, who can hear or read it, which tool processes it, how long recordings and transcripts are kept, who can access them, what happens if a member leaves the group, and that a single refusal means the session will not be recorded. Never make a recording the default, and never ask the group in the room to raise hands. Group pressure is not consent; each person should answer privately, in writing, before the first session, and again if the tool or the terms change.

Illustrative case: a group of eight, seven of whom consent to recording. The eighth says no on a private form. The facilitator can switch to jottings-only for everyone, which keeps the group whole and spares the eighth member from being singled out. Telling the seven that "we've decided not to record after all" without saying why is both kind and sufficient. That decision, and the reason, go in the group's file. The tutorial on telling clients you use AI has more on how to raise it without alarming people.

The review routine: six checks before filing

1. Identity sweep. Search each note for any other member's label, name or initials. Then read it with the question: could anyone in this group recognise another member from this line?
2. Source check. Put the jottings beside the note. Each statement in the note must come from the jottings. Delete any extra.
3. Inference check. Find every verb of feeling or cause ("seemed anxious because", "clearly struggling with"). Keep only if it is your observation, and word it as one.
4. Risk line. The clinician completes it from their own judgement. If anything in the session raised concern, it is written by you, not the tool.
5. Similarity check. Compare this note with the member's last two. Notes that are almost identical week after week are a record-keeping problem in their own right. The cloned notes tutorial shows how to spot and avoid them.
6. Sign-off. Name, date and time in the review line. Only then does the note go into the record.

Budget three minutes a note for review. An eight-member group means about 25 minutes, which is not trivial, and any time saving needs to be set against it. If you find yourself skipping checks to save time, stop using the tool for that group.

Worked example: an eight-member weekly group

Illustrative figures for a practice that runs one eight-member skills group weekly for 12 weeks.

Before: the facilitator writes eight notes after the session, taking about 12 minutes each, so roughly 96 minutes a week, often at the end of a long clinical day. Over 12 weeks that is 19 hours.

With the routine above: jotting during and just after the session takes about 15 minutes for the whole group. The tool drafts eight notes in about 10 minutes of back-and-forth, with a fresh chat for each member. Review takes about 3 minutes per note, so 24 minutes. The total is about 49 minutes a week, so about 47 minutes saved weekly and 9.4 hours over 12 weeks. That is real, and it is also about half of what many sales pages imply, because review time does not vanish. Time your first three sessions before you decide whether the routine is worth keeping.

A realistic slip from week four: in the second draft for M5, the tool wrote "supported another member who spoke about caring for a parent". M5's jottings said only "supported another member". The tool had borrowed the detail from earlier in the same chat, where the facilitator had asked a clarifying question using M2's information. The review caught it because the identity sweep reads for any added detail. The fix was procedural: a new chat for every member, and no clarifying questions with another member's information in the same thread.

Before and after: a leaky note and its repair

Seeing the failure beats reading the rule. Illustrative: here is a draft the tool produced when it was given the whole group's jottings at once, for the member labelled M6.

"M6 attended and was engaged. M6 listened as M2 talked about the death of a partner and said that it reminded them of their own loss last year. M6 became tearful when M4 described a hospital stay and left the room for five minutes. M6 thanked the facilitator."

Three problems. It names M2 and M4 by label, which is identifying within the practice's own system. It carries M2's bereavement and M4's hospital stay into M6's record. And "became tearful" and "reminded them of their own loss" are inferences, unless the facilitator saw or heard them. The repaired version, from M6's own jottings only:

"Attended the full session apart from a five-minute break (left the room during discussion; returned unprompted). Contributed once, saying the topic was close to their own experience of loss. Responded to another member's disclosure with a supportive comment. Thanked the facilitator at the end. Risk observations: clinician to complete."

Notice what happened to the "tearful" line. It is gone, because the facilitator's jottings say only that M6 left for five minutes. If the facilitator did see tears, they add that observation to the jottings, in their own words, and the tool carries it. The tool should never be the first to notice a feeling.

Sharing with referrers, funders and other clinicians

Group notes often travel: a progress letter to the referring GP, a report for a funder, a summary for a co-working clinician. The per-member rule applies at every step, and the tool needs more care here because letters are written for an outside reader.

Build these from the member's filed note, not from the group's raw material. Give the tool the member's own signed-off note, with the label and not the name, and a short instruction on what the letter is for. Ask it to use only that note, to add no clinical interpretation, and to leave a clearly marked gap where a clinical judgement is needed. Then you write the judgement yourself. Before sending, check what the member has agreed can be shared and with whom, since consent to group participation is not consent to disclosure outside the practice.

Two illustrative cases. A funder asks for attendance and engagement across 12 sessions: the tool can turn a table of attendance dates and the facilitator's one-line engagement notes into a short paragraph, and it needs no clinical detail at all. A referrer asks how a member is responding: that letter needs the clinician's own view, and the tool's job is limited to tidying structure and checking that no other member is mentioned. If you can answer a request with attendance and a single line the facilitator wrote, never involve the tool with anything more.

Monthly audit: ten notes, fifteen minutes

Once a month, the practice lead pulls ten AI-assisted notes at random across groups and facilitators, and checks four things: no other member is identifiable, every statement traces to the jottings, the risk line was completed by a clinician, and the review line is signed. Record the results in one line per note. Count failures by type, and change the prompt or the template for the commonest one.

A realistic finding after the first audit: the identity sweep passed on all ten, but four notes ended with the same sentence, "Plan for next session: continue to attend and engage." That is a signal the template's plan line is inviting boilerplate. Reword the field as "Plan for next session: only if the facilitator has a specific plan; otherwise write 'none specific'." A prompt that allows a specific "none" prevents invented plans.

Edge cases worth deciding in advance

• A member discloses a risk to themselves or others. The clinician writes that part of the record by hand, following the practice's risk protocol. The tool never summarises a risk disclosure.
• Co-facilitated groups. Agree which facilitator writes jottings, so you don't get two versions of the same session. Both read the final note if they both observed the member.
• A member asks for a copy of their note. Know in advance how your rules handle access, and whether the AI-assisted drafts you used on the way count as part of the record. Ask your adviser before the first request arrives.
• A member leaves. Their consent record, jottings and drafts are dealt with under your retention policy. Delete chat histories in the tool on the schedule your policy says, not whenever you remember.
• A member is under 18 or lacks capacity to consent. Consent comes through the appropriate parent, guardian or representative as your rules require; check before using any tool at all.
• A new tool or a changed setting. Re-run consent if the processing materially changes.

Practical guidance on an overall notes workflow is in the confidentiality checklist for note tools, and the individual-session angle is in the tutorial on using ChatGPT for session notes. If you are still setting up the group itself, the group launch tutorial covers filling and structuring it.

A written policy in ten lines

Put this on one page and have every facilitator sign it. Adapt it with your own adviser.

1. AI tools may be used only to tidy facilitator-written jottings.
2. No real names or identifying details are entered; labels are used and the key stays in the records system.
3. One member, one chat, one note.
4. No recording of sessions unless the practice has approved the specific tool, signed any data terms required, and every member has consented in writing.
5. Members may refuse AI use without any effect on their place in the group.
6. Risk observations are written by the clinician.
7. Every note passes the six checks and carries a review line before filing.
8. Chat histories are deleted on the retention schedule.
9. Errors and near misses are logged and reviewed monthly.
10. The policy is reviewed every six months and whenever the tool or its terms change.

If you'd rather talk the set-up through, that is what my AI implementation consultation covers, and the decisions here, labels, consent and review, are the same ones we would work through.

Questions group facilitators ask about AI notes

Can I record a group session and have AI write the notes?

Only with each member's informed, recorded consent, and only if your tool and its data terms are cleared by your practice and, where relevant, your regulator or insurer. A recording carries every member's voice and disclosures, so one refusal usually means no recording. Many groups decide a facilitator's own dictation after the session is safer.

What if one member says no to AI but the others agree?

Then AI is not used for the session, or not for that member's content. The safest default is to write that member's note by hand and keep them out of any shared prompt. Never make the group's decision depend on one person feeling pressured to agree.

Should the AI see other members' names?

No. Use member labels such as M1 to M8 in anything you give a tool, and keep the key in your records system. Each member's note should mention other members only as 'another member', never by name or by identifying detail.

Which AI plan is approved for therapy notes?

Check, in writing, with your vendor and your own compliance adviser. A plan's marketing page is not an agreement. Whatever you choose, the signed data terms, your regulator's rules and your insurer's position decide it, not a tutorial. Keep the vendor's written reply on file with the date.

Further reads

• AI for Speech Therapy Reports: Draft, De-Identify, Review — Draft, de-identify and review, applied to reports.
• AI Tools for Psychotherapists: Diary, Invoices and Referral Replies — Admin jobs around the sessions that AI can take on.
• AI in Psychotherapy Practice: Enquiry Replies and Fee Policies — Enquiry replies and fee policy wording.
• AI Scribe for Occupational Therapists: A Buyer's Checklist — Buyer's questions for any scribe tool.
• AI Notes for Chiropractors: Consent and Data Checks Before Recording — Consent and data checks before recording.
• How Physiotherapists Use AI to Write Treatment Notes Faster — Another profession's note workflow, for comparison.
• Therapy Practice Management Software: What the AI Add-Ons Cost — SimplePractice, Jane and Carepatron base plans against their AI add-ons, with a monthly sum for a three-therapist practice and the traps in each quote.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI Data Controls FAQ; ChatGPT Business and Claude plans pages; the psychotherapy-notes definition in 45 CFR 164.501 (Cornell LII copy) (checked 4 November 2026). Check your own regulator and insurer.