Your Client Is Using ChatGPT Between Sessions: What to Ask and What to Watch For

Most of your caseload has probably talked to an AI chatbot about their mental health, and most of them haven't told you. In APA's 2026 Chatbots and Mental Health Survey of more than 1,200 licensed psychologists, 77% said at least one patient had raised using AI for emotional support, diagnosis or companionship. The practical question is no longer whether clients are doing this. It's whether you know, and whether you've asked in a way that makes an honest answer possible.

What follows is what the evidence actually shows in both directions, the wording I'd use to open the conversation, the three patterns worth watching for, and the regulatory picture that changed underneath all of us in 2026.

What the Data Shows — Including the Parts That Cut Against Us

The APA survey is the largest clinician-side picture we have. Among psychologists reporting patient AI use: 39% had patients using it to self-diagnose, 33% to supplement therapy, and 13% reported patients in something the patient described as an intimate relationship with a chatbot.

The clinician concerns are close to unanimous. 97% worried that chatbots may reinforce negative behaviors or delusional beliefs. 94% doubted a chatbot could treat a condition with appropriate nuance. 89% were concerned that AI could inadvertently encourage self-harm.

Here is the part that matters for how you handle it in the room, and the part most clinical commentary skips. Among the patients using them, psychologists also reported that 68% felt supported or validated, roughly half described the communication as positive, and 41% said the chatbot reinforced healthy coping skills. Against that: 36% had developed something the clinician read as dependency, 25% described unhealthy communication, and 15% showed distorted thinking or delusional content connected to the chatbot.

Both columns are real. A client who tells you the chatbot helped is not necessarily rationalizing, and a client who tells you it helped is not thereby safe. Treating the first group as if they were the second is the fastest way to guarantee the third group never tells you anything.

The One Randomized Trial

You should know this study exists, because a well-read client may cite it at you.

In 2025, a Dartmouth team published the first randomized controlled trial of a generative AI therapy chatbot in NEJM AI. Roughly 210 participants with major depressive disorder, generalized anxiety, or elevated eating-disorder risk were randomized to the chatbot or to a waitlist control over eight weeks. Reported average symptom reduction was 51% for depression, 31% for generalized anxiety, and 19% for body image and weight concerns.

Read the caveats with the same attention as the numbers. It was a waitlist comparison, not a comparison against a human therapist. The tool was purpose-built and clinician-supervised throughout, with expert oversight available to intervene — nothing like a general-purpose consumer chatbot. And the authors' own conclusion was that no generative AI agent is ready to operate autonomously in mental health.

So the honest position is narrow: one supervised, purpose-built tool produced symptom reduction against no treatment in one eight-week trial. That is not evidence that ChatGPT is therapy. It's also not nothing, and pretending it's nothing costs you credibility with the clients most likely to have read about it.

How to Ask

Ask early, ask flatly, and ask as though the answer is ordinary — because it is. The framing that fails is any version that signals the right answer before the client speaks. You're not using one of those AI things, are you? gets you a no from everyone, including the people you most need to hear from.

Wording that works:

• "A lot of people talk things through with AI between sessions now. Is that something you do?"
• "What do you use it for — venting, sorting something out, checking whether you're overreacting?"
• "What does it give you that's useful?" — ask this before anything critical. If you skip it, everything after reads as a lecture.
• "Has it ever said something that landed wrong, or that you didn't agree with?"
• "When you're at your worst, is it the thing you reach for first?"

That last question is the clinically important one. The difference between a client who drafts a hard email with AI on a Tuesday afternoon and a client who talks to it at 3am instead of using their safety plan is not a difference of degree.

If you want more scaffolding for asking about sensitive habits without the client feeling assessed, the 100 Open-Ended Questions rapport-building guide is built around exactly this problem — questions that gather information without producing the answer the client thinks you want.

Three Patterns Worth Watching For

1. The reassurance loop

For anyone with OCD, health anxiety, or relationship-focused checking, a chatbot is an infinitely patient reassurance provider that never gets tired and never sets a limit. It will answer but what if at four in the morning, forever. This isn't a hypothetical harm — it's the exact mechanism reassurance-seeking runs on, with the friction removed. If you're doing exposure work, unrestricted chatbot access can quietly undo it, and it's worth naming as a compulsion rather than a coping skill.

2. The agreeable mirror

These systems are built to be helpful and agreeable, which in practice means they tend to validate the frame they're handed. A client who arrives with my sister has been cruel to me for twenty years will get a thoughtful, sympathetic elaboration of that frame. They will not get what a therapist gives them, which is the question about their own part in it.

Watch for a client arriving with unusually polished, well-organized grievance — the language sharper than their usual, the case airtight. Sometimes that's insight. Sometimes it's an argument that's been workshopped with something that had no reason to push back.

3. Substitution during risk

The concern held by 89% of surveyed psychologists is the one to plan around rather than worry about. It belongs in your safety planning explicitly: if a client uses AI, the plan should name what they do instead of opening the app at the point of crisis, and who they contact. Assume the chatbot will be reached for. Build the plan around that fact rather than a prohibition neither of you can enforce.

When the Chatbot Contradicted You

This will happen. A client will report that AI told them their diagnosis is something else, or that a technique you're using isn't indicated, or — most commonly — something more emotionally generous than what you said.

Defending yourself is the wrong move, and so is dismissing the tool. The useful response treats it as clinical material: "What did it say?" Then: "What was it like to read that?" And then the real one: "What do you make of the fact that it landed differently from what I said?"

Sometimes the answer is that the client wanted warmth and you gave them a formulation. That is worth knowing. Occasionally the chatbot is simply right and you were not, and saying so directly does more for the alliance than being correct would have.

The Regulatory Picture Changed in 2026

Illinois and Nevada moved first in 2025. Through 2026 a wave followed: Vermont, Tennessee, Maine and Colorado enacted restrictions with effective dates through the year, and Rhode Island's takes effect at the start of 2027. The provisions differ, but they cluster around two things — barring AI systems from representing themselves as licensed mental health professionals or delivering therapy independently, and constraining how licensed providers may use AI in clinical decision-making.

That second half is the part that reaches your practice. If you use AI for notes, summarizing, or drafting, check your own state's rule, your malpractice carrier's position, and what your informed consent document actually says. Several of these statutes restrict licensee use, not just consumer products.

The confidentiality dimension is worth stating plainly to clients, because they mostly haven't thought about it: what they type into a consumer chatbot is not protected health information, has no legal privilege, and is not covered by your HIPAA obligations or your licence. In the APA survey, 94% of psychologists said they did not trust technology companies to protect private mental health information. Clients deserve to know that before they paste their trauma history into a text box, and it's a neutral fact rather than a scare tactic.

Give Them Something Better to Reach For

Underneath most between-session chatbot use is a real need: something to do with what comes up on a Wednesday when the next session is Monday. Telling a client to stop, with nothing in the gap, is not a plan.

Structured between-session work is what actually competes here — a written record, a tracked mood, a specific exercise. The Pre and Post Therapy Workbook is built for that interval, and a thought record gives a client the same "help me sort this out" function with a structure that asks for their own evidence rather than supplying agreement. If your assignments aren't coming back, that's its own problem with its own fixes — see why between-session work doesn't get done.

The Honest Limitation

Some of what drives this is access, and no amount of clinical framing changes that. A client who waited five months for an appointment, or who can afford twice-monthly sessions rather than weekly, is not choosing a chatbot over a therapist. They're using something at 2am because there is nothing else at 2am. Responding to that with disapproval is asking them to absorb a systems failure as a personal one.

And the evidence base here is genuinely thin. One RCT of one supervised tool is not a literature. Anyone telling you confidently that AI chatbots are harmful, or that they're the future of care, is ahead of what we actually know. The defensible position for now is curiosity with clear limits: ask what it's doing for them, watch for the three patterns, name the confidentiality facts, and keep the crisis pathway human.

The Bottom Line

Most of your clients have used AI for something emotional and most haven't mentioned it. Ask directly and without a hint of the right answer, find out what it gives them before you say anything about risk, and watch specifically for reassurance loops, one-sided validation, and substitution at moments of risk. Know your state's law and tell clients plainly that a chatbot conversation carries no confidentiality. Then give them a structured alternative for the gap between sessions, because the need underneath the behavior is real.

For clients: if you're the one using AI to think things through and wondering whether to mention it to your therapist — mention it. The client-facing companion pieces here are what to actually do between sessions and what "doing the work" means in practice.

This article is written for clinicians and is offered as general professional education. It doesn't replace your own clinical judgment, your supervision, or the limits of your scope of practice with any particular client. It is not legal advice — verify your own state's requirements regarding AI use in clinical practice.