How to Use AI as a Therapist: Uses, Limits, Risks

A therapist's week also splits into two different kinds of work filed under one job title: the writing that documents what happened, and the work itself — reading a client in the room, holding the relationship over months, and making a risk call that someone has to be licensed and accountable for. A general AI tool is genuinely useful at the first: a progress note drafted from what was actually covered in session, a coping-skills handout rewritten in plainer language, an intake form filled from information already decided. It has no access to the second, and treating a fluent answer as a clinical one is the mistake worth avoiding from the start — a widely cited survey of hallucination in large language models covers why confident and correct are not the same property in these systems, which matters most exactly where the question is clinical.

Will AI replace therapists covers that split at the level of the whole profession — what is genuinely shrinking and what is not. This is the practical version: what to actually hand over this week, and exactly where a fluent-sounding answer needs a human read before it reaches anyone.

Start with what makes therapy different from most office work

Two things narrow the options before a tool is even open. Session content carries the same confidentiality obligations as any other clinical record, and pasting identifiable details about a client's history, disclosures or diagnosis into a consumer chat tool is not a grey area — it is the kind of mistake a practice policy exists to prevent. The second is that a therapist's real output is not the note, it is the judgement behind it: what a client's silence or tone actually means in that specific session, and what to do when something disclosed looks like risk. A tool has no access to either, and no licence backing the call either way.

Where AI genuinely helps

The pattern across everything below is the same: the tool drafts text from something you already decided in session, and you check it against what was actually said before it goes out under your name.

1. 01Drafting a progress note from bullet points of what was actually covered, instead of writing the narrative from a blank page between clients.
2. 02Rewriting a coping-skills or psychoeducation handout in plainer language for a specific client, from a technique already taught in session.
3. 03Turning treatment-plan goals already agreed with a client into the written summary a chart or an insurer expects.
4. 04Summarising a long clinical article or practice guideline against a specific question, so you can go straight to the relevant section instead of rereading the whole thing.
5. 05Drafting a first-pass reply to a routine scheduling or billing message, which you review and send yourself.

A worked example

Say a session just covered a specific grounding technique, and the client would benefit from a short handout to reread during the week.

The bad version: describe the client's history or what they disclosed in session and ask a general tool what might be going on with them, or ask it to recommend which technique this particular client needs. That treats a system with no access to the relationship, and no licence, as if it had both — and a fluent-sounding answer gives no sign it is wrong.

The version that works: take the technique already decided in session and ask the tool only to turn it into a short, plain-language handout — nothing about the client's case included, nothing clinical decided by the tool. That is the same instinct OpenAI and Anthropic both give in their own prompting guidance: name the exact task and the exact constraint, rather than asking an open question and trusting whatever comes back. Read the handout against what was actually taught before it leaves the room.

A model can turn a coping technique into a take-home handout faster than a therapist can retype it. It cannot be the one who decided which technique this client needed.

Where it will quietly get you into trouble

• —Describing a client's symptoms or history to a general tool and treating a confident-sounding read as a clinical opinion. The same hallucination research applies here with more at stake than almost anywhere else on this list.
• —Pushing back on your own correct read after a tool disagrees, and watching it fold instead of hold its ground. Research on sycophancy in language models found these systems often agree with whatever the user pushes toward, which is the opposite of a second opinion worth trusting.
• —Pasting identifiable client details into a consumer tool with no agreement covering how that data is handled. Responsible AI examples that actually come up at work covers the redact-first habit that applies here before anything else does.
• —Letting a rewritten handout drift from what was actually taught in session. How a worked method catches an added detail in a summary is the general version of the check worth running on anything a tool drafts before it goes out under your name.

Getting started without a project plan

1. 01Check your practice's or employer's policy on AI tools and client data before using any general tool on real client information, even informally.
2. 02Pick one recurring, checkable task first — a progress-note draft from a session you already remember, or a handout rewrite of a technique already taught — because the output is instantly checkable against what happened.
3. 03Find the repetitive part of your own week if the first task is not obvious. For most therapists it is the writing around sessions, not anything closer to the session itself.
4. 04Draw one hard line and keep it: nothing that assesses risk, suggests a diagnosis, or responds to a disclosure goes near a general-purpose tool. The relationship and the judgement call stay with the therapist.

The part that actually protects the role

Task-level research mapping real AI usage onto occupational data keeps finding the same shape across clinical and caring work: a tool scores highest on drafting, summarising and rewriting — the documentation wrapped around the job, not the job itself. Microsoft's own follow-up note on that research is explicit that a high score there is a claim about assistance, not about replacing the relationship and the judgement a licence is actually paying for. How to use AI as a pharmacist and how to use AI as a nurse work through the same split for two other licensed, hands-on roles: the drafting layer shrinks, the verification and the judgement call do not.

The useful stance is neither refusing the tools nor leaning on them for anything clinical. It is being the therapist who gets a note or a handout drafted in a fraction of the usual time, and never once asks a chat tool to assess a client or respond to a disclosure.

The short version

Use AI for progress-note drafts, handout rewrites and routine paperwork, on a practice-approved tool with nothing identifiable about a client pasted into a consumer product. Never let it near a diagnosis, a risk assessment or a disclosure. Check your practice's policy before you start, pick one checkable task, and keep the hard line exactly where it is.

That checking habit is also the one the labour market has started pricing in more broadly: PwC has measured a real wage premium for workers with demonstrable AI skills across the wider economy — 62%, up from 57% the year before. Coursium teaches the practical layer underneath that — using the tools well on the paperwork, and knowing exactly where to stop. Stay ahead of AI by learning them on your phone.