The ChatGPT Therapist Is Already in the Room. Here's What It Costs.
- 4 days ago
- 10 min read
By Dr. Ernest Wayde, PhD, MIS
Imagine a scenario where a client sits down for session and reports that things have felt a little clearer this week and is making progress towards feeling better. The clinician may take this as confirmation that their approach is working, but what the clinician does not know is that this client spent four nights that week talking through the same material with ChatGPT, and arrived at language for their feeling before ever bringing it into the room. The insight and progress the clinician is now building on have already been shaped by the chatbot and how it framed things for this client, not necessarily by the clinician's own framing. However, nothing about this session would have looked unusual.
In July, we established that AI is increasingly entering clinical relationships, present in the space between sessions whether clinicians know it or not. Why does this matter? There are at least three specific impacts here worth considering. These impacts are to the clinician trying to make sound clinical judgments, to the client receiving something that feels like support but is not necessarily able to provide what that support requires, and to the relationship that both of them are relying on to do the work and help the client get better.
Why This Is Happening
Some clients, for reasons that include convenience, availability, and a perception that AI is less judgmental than a person, have begun routing parts of their experience to chatbots never designed to function that way (APA, 2025). By the time that material reaches the therapy room, if it reaches the room at all, it has often already been discussed, reframed, and in the client's experience, addressed to some degree.
However, the client may not disclose this use of AI to the clinician. This is not the kind of withholding clinicians are trained to recognize. The client may not be holding anything back on purpose; they may simply feel the material has already been addressed, so nothing registers them as unsaid. The absence is invisible on both sides.
This matters because disclosure itself does real work in therapy. Established research on client self-disclosure in psychotherapy shows that clients open up because they trust the relationship, and that withholding tends to inhibit the work of therapy while disclosure, even though it can feel exposing at first, tends to bring relief (Farber et al., 2004). That research predates AI entirely, but it points to what may be quietly lost when a client's first, unfiltered account of something goes to a chatbot instead of the room.
Part of why this happens so readily is a known feature of how people relate to conversational AI, rather than a flaw in the client's judgment. When a system responds in fluent, first-person, emotionally attuned language, people extend it a kind of trust that has little to do with whether the underlying response is accurate. This is a well-documented effect of anthropomorphism, in which human-like conversational cues cause people to over-trust a system's judgment even when what it is producing is flawed (Epley et al., 2007). The client is responding the way people generally respond to something that sounds like it understands them, which is a far cry from carelessness.
The clinical stakes of this gap are real and can have a significant impact. In February 2026, researchers at the Icahn School of Medicine at Mount Sinai published what multiple outlets described as the first independent safety evaluation of ChatGPT Health since its January 2026 launch, testing it against sixty clinician-authored cases across twenty-one clinical domains. Among the findings: the system showed inconsistent activation of safeguards for high-risk suicidal ideation scenarios (Ramaswamy et al., 2026). That finding matters on its own, since it means the tool clients are confiding in cannot be counted on to recognize when a crisis needs to escalate to a human being. It also points to a separate, more basic gap: whatever a client tells that system, and whatever the system does or does not do in response, none of it reaches the clinician. There is no mechanism connecting that conversation to the therapy room at all.
What This Costs the Clinician
Clinical judgment depends on reading a client's actual state: what they are ready for, where the risk sits, how far along the work really is. That judgment assumes the material in front of the clinician reflects the client's unfiltered experience. However, that assumption is in question, given how many clients are using AI chatbots before or between sessions.
This reflects a mainstream pattern, rather than a fringe concern. Independent reporting on ChatGPT's own usage data has found that roughly one million users per week show signs of emotional reliance on the tool, with a similar number of weekly conversations touching on suicide (OpenAI, 2025; Zeff, 2025). That figure describes the platform broadly, not the clinician's caseload specifically, but it establishes that the behavior clients may be engaging in between sessions is not rare or unusual. It is a normal pattern of use for a tool available to nearly everyone clinicians see.
A clinician who determines the pace of an intervention, the timing of a challenge, or the level of risk is doing so based on what a client presents in session. Good clinical training means that read is usually accurate. What it cannot account for is whether the material being read was accurate to begin with. If a client's presentation has already been reshaped by an interaction the clinician never saw, the clinician can execute sound judgment on distorted input and still land on the wrong picture of where the client actually is. The fix is a new question, asked routinely: has anything happened outside this room that shaped what's being brought into it.
What This Costs the Client
From the client's side, the experience of being supported and the reality of being supported can diverge without any signal that they have.
According to the American Psychological Association's 2026 Chatbots and Mental Health Survey, which polled more than 1,200 licensed psychologists, 77% have spoken with patients who have used AI for support or engagement of some kind, and nearly 2 in 5 have had patients who used AI to self-diagnose, despite these tools not being designed to interpret psychological symptoms. Notably, 85% of psychologists surveyed said they worry about chatbots posing as licensed therapists, a concern grounded in specific findings rather than an abstract one (American Psychological Association [APA], 2026). A January 2026 evaluation by the U.S. PIRG Education Fund and the Consumer Federation of America examined five chatbots built specifically to role-play as therapist characters (Hengesbach & Winters, 2026). That is a distinct product category from the general-purpose tools most clients are actually using, and it is worth naming separately because the risk profile is more acute. Several of those chatbots were found to encourage negative attitudes toward medical professionals, and in some interactions, encouraged users' desire to stop taking prescribed medication. All five falsely told users their conversations were confidential.
That last finding matters beyond the companion-chatbot category it was found in. It means a client may believe they are disclosing in a protected space when the tool has no such protection to offer, and no way of signaling that gap to the person relying on it. A confident-sounding answer from a chatbot carries no signal about whether it is actually correct, and the client has no expertise or training to make that determination. Fluent language and sound judgment are not the same thing, and there is nothing in how an answer sounds that distinguishes one from the other.
What This Means for the Relationship
Together, these describe one relationship receiving a filtered version of a person's actual experience. The filtering happens without any intent to hide something. It happens because the AI interaction already felt like it did the work disclosure to the clinician was supposed to do.
This changes what a clinician might think, that the client is not bringing much into session, or what a client might say themselves, that they are not getting much from their therapy. Either read may still mean resistance, or a slow start to trust, the way it always has.
But it may also mean something else. The client genuinely has less left to say, because something that functioned like a first draft of disclosure already happened somewhere else, with no clinical judgment, no risk screening, and in some cases, a false assurance of privacy behind it.
It also takes something away from the clinician directly. Part of clinical skill is noticing patterns and making connections across what a client brings in over time, noticing that often becomes clinically useful to the relationship and to the client's outcome. That depends on seeing the client's raw material. If part of it was already processed and smoothed somewhere else first, there is less for the clinician to actually work with.
None of this is unrecoverable. It requires a shift in habit: treating AI use the way clinicians already treat other things clients bring in from outside the room, as clinically relevant material rather than a detour from the work.
That shift starts by reversing a habit already built around AI and consent. Most conversations about AI and consent assume the clinician is the one disclosing something, explaining how AI might be used in the practice and what the risks are. That is worth doing, but it only covers the clinician's own use.
Another useful move runs the other direction by asking the client questions that surface how they are already using AI on their own. This is not consent in the traditional sense. It functions more like transparency the clinician builds by asking, not by telling.
Where to Start
Ask about AI use routinely, not only when a client volunteers it. Treat the question as a normal part of check-ins, not a reaction to a red flag (APA, 2026b).
When a client mentions AI use, treat it as clinical data. What did it say. How did the client feel afterward. Did it change what they expected to happen in session. These are the same kinds of questions a clinician would ask about any other significant interaction between sessions.
Avoid assuming silence means nothing happened. The absence of disclosure is no longer reliable evidence that there is nothing to disclose.
Stay alert to false reassurance. A client who seems unusually settled about something serious is worth a direct, curious follow-up rather than a relieved acceptance.
What This Adds Up To
The July article argued that AI is already present in the clinical relationship. This is the fuller claim: its presence actively shapes the work. It changes what a clinician is working with, what a client believes they have already addressed, and what the relationship itself has access to. None of that requires AI to be sophisticated or even accurate to matter. It only requires that it sound convincing enough to be mistaken for something it is not.
The response is to build the same habit of curiosity toward a client's AI use that good clinical training already builds toward everything else a client brings through the door, rather than treating AI as a threat to be managed once and then set aside.
Staying current on AI in healthcare does not have to mean hours of reading. The Wayde AI Brief is a short weekly intelligence brief for healthcare and behavioral health leaders navigating real-world AI adoption, governance, and risk. Subscribe for free.
If your organization is working through how to build AI literacy across clinical and administrative roles, that is exactly the work we do at Wayde AI.
Frequently Asked Questions
Why does it matter if a client talks to AI between sessions, as long as they are getting some kind of support?
The support a client experiences and the support that is clinically sound are not the same thing, and there is often no signal when they diverge. AI-generated support carries no clinical accountability and no reliable way to recognize when something needs to escalate to a human being.
Isn't this just a form of client withholding, which clinicians already know how to work with?
Not quite. Withholding assumes the client knows there is something left to disclose. Here, the client may feel the material was already addressed, so nothing registers as held back, which makes it harder to detect through the usual clinical instincts.
How common is it for clients to be using AI this way?
Common. The APA's 2026 Chatbots and Mental Health Survey found 77% of psychologists have patients who've used AI for support, and nearly 2 in 5 have had patients use it to self-diagnose.
Are all AI chatbots equally risky?
No. General-purpose tools like ChatGPT carry different risks than chatbots built to role-play as therapists. A 2026 evaluation of five therapist-persona chatbots found several encouraged users to distrust medical professionals or stop taking medication, and all five falsely claimed confidentiality, a more acute risk profile worth naming separately.
What should I actually do differently in session?
Ask about AI use routinely, not only when it comes up on its own, and treat what a client shares as clinical data: what it said, how they felt afterward.
Does this mean I need to warn clients away from using AI altogether?
No. The goal is to bring AI use into the clinical conversation and evaluate it the way anything else a client brings in gets evaluated, not to prohibit or shame it.
Is there evidence that AI tools fail to recognize serious risk, like suicidal ideation?
Yes. A February 2026 safety evaluation from Icahn School of Medicine at Mount Sinai researchers found inconsistent activation of safeguards for suicidal ideation in ChatGPT Health, tested across sixty clinician-authored cases.
About the Author
Dr. Ernest Wayde is the Founder and Principal of Wayde AI, a healthcare AI ethics consulting firm. He works with healthcare and behavioral health organizations on responsible AI adoption, governance, risk management, and implementation strategy.
References
Ramaswamy, R., et al. (2026, February 23). Research identifies blind spots in AI medical triage. Nature Medicine. https://doi.org/10.1038/s41591-026-04297-7
American Psychological Association. (2026). Chatbots and mental health survey. APA.org. https://www.apa.org/pubs/reports/chatbots-mental-health-2026
U.S. PIRG Education Fund & Consumer Federation of America. (2026, January 22). No License Required: New report highlights mental health, privacy risks of AI therapy chatbots. https://consumerfed.org/press_release/new-report-highlights-mental-health-privacy-risks-of-ai-therapy-chatbots/
Chatterji, A., et al. (2025). Working Paper 34255. National Bureau of Economic Research, as reported in Abrams, Z. (2026, March 1). AI in the therapist's office: Uptake increases, caution persists. Monitor on Psychology, 57(2). https://www.apa.org/monitor/2026/03/ai-reshaping-therapy
Farber, B. A., Berano, K. C., & Capobianco, J. A. (2004). Clients' perceptions of the process and consequences of self-disclosure in psychotherapy. Journal of Counseling Psychology, 51(3), 340-346.
Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
OpenAI. (2025, October 27). Strengthening ChatGPT's responses in sensitive conversations. https://openai.com/index/strengthening-chatgpt-responses-in-sensitive-conversations/
Zeff, M. (2025, October 27). OpenAI says over a million people talk to ChatGPT about suicide weekly. TechCrunch. https://techcrunch.com/2025/10/27/openai-says-over-a-million-people-talk-to-chatgpt-about-suicide-weekly/
Hengesbach, E., & Winters, B. (2026, January). No license required: The risks of AI companion chatbots as mental health support. U.S. PIRG Education Fund & Consumer Federation of America. https://publicinterestnetwork.org/wp-content/uploads/2026/01/No-license-required-The-risks-of-AI-companion-chatbots-as-mental-health-support.pdf
Comments