ChatGPT is entering psychotherapy from two directions at once. Some people are replacing human therapists with AI tools, while some therapists are bringing those same tools into their own work with clients.
The central problem is not only whether AI can produce a polished response. It is whether clients know when it is being used, what personal information may be exposed, and whether the therapist is still doing the human work the relationship depends on.
How clients found out
Declan, 31, who lives in Los Angeles, discovered his therapist was using ChatGPT because of a technical error during an online session. After a patchy connection led him to suggest turning off video, the therapist accidentally began sharing his screen.
Declan saw his own words being entered into ChatGPT, followed by summaries and selected responses from the chatbot. He did not interrupt the session. Instead, he watched the process unfold in real time and even began anticipating the chatbot's suggestions in his own answers.
At the next session, Declan raised what had happened. The conversation was awkward, and the therapist explained that he felt they had hit a wall and had looked elsewhere for help. Declan still paid for the session.
Hope, 25, who lives on the East Coast of the US, had a different experience. She messaged her therapist about the death of her dog and received a reply that might have felt compassionate if an AI prompt had not been left at the top: "Here’s a more human, heartfelt version with a gentle, conversational tone."
Hope said the episode left her surprised, confused, and betrayed. The timing mattered because part of the reason she was in therapy was for trust issues. Her therapist apologized and said she had used AI because she had never had a pet herself and wanted help finding the right sentiment.
The disclosure problem
The source describes a recurring pattern: AI-assisted communication may appear helpful until the client suspects or discovers that AI was involved. That shift can change the meaning of the message.
A 2025 study published in PLOS Mental Health asked therapists to use ChatGPT to respond to therapy-style vignettes. A panel of 830 participants could not reliably tell which responses were human and which were AI-generated. The AI responses were also rated as aligning better with therapeutic best practices.
But suspicion changed the result. When participants believed ChatGPT had written a response, they rated it lower. The responses written by ChatGPT but attributed to therapists received the highest ratings overall.
Cornell University researchers found a similar tension in a 2023 study. AI-generated messages could increase feelings of closeness and cooperation, but that benefit depended on the recipient not knowing AI had played a role. Suspicion alone could damage goodwill.
Adrian Aguilera, a clinical psychologist and professor at the University of California, Berkeley, framed the issue around authenticity. In psychotherapy, clients expect the therapist to be present, attentive, and personally engaged. A technically competent message can still feel hollow if the client thinks it came from a chatbot instead of the person they trusted.
The online therapy service Koko also exposed the sensitivity of this boundary in 2023, when it conducted a clandestine experiment using GPT-3-generated responses mixed with human-drafted ones. Users tended to rate the AI-generated responses positively, but the revelation that they had been part of an undisclosed experiment sparked outrage.
BetterHelp has also faced claims that therapists used AI to draft responses. Photographer Brendan Keen wrote in a Medium post that his BetterHelp therapist admitted using AI in replies, leading to "an acute sense of betrayal" and continued concern that his privacy may have been breached. BetterHelp said it prohibits therapists from sharing members' personal or health information with third-party artificial intelligence or using AI to craft messages when that could directly or indirectly identify someone.
Privacy risks are harder than removing names
Undisclosed AI use in therapy is not only an emotional issue. It can also become a data privacy issue, especially when personal details from sessions or messages are entered into general-purpose AI chatbots.
Margaret Morris, a clinical psychologist and affiliate faculty member at the University of Washington, said these tools may be valuable for learning, but therapists must be careful with patient data. She called Declan's experience alarming.
Pardis Emami-Naeini, assistant professor of computer science at Duke University, said general-purpose AI chatbots like ChatGPT are not approved by the US Food and Drug Administration and are not HIPAA compliant. HIPAA is described in the source as a set of US federal regulations that protect sensitive health information.
Emami-Naeini warned that privacy risks can arise if information about a patient is disclosed or can be inferred by the AI. In a recent paper, she found that many users wrongly believe ChatGPT is HIPAA compliant, and she expects some therapists may share that misconception.
Declan said he was not devastated by what happened, but it still felt violating. He noted that if the content had involved suicidality, drug use, or infidelity, he would not want that entered into ChatGPT.
The source also makes clear that anonymizing therapy material is not simple. Emami-Naeini said removing obvious identifiers such as names and addresses is not enough, because sensitive information can be inferred from details that may look harmless. She said therapists should disclose AI use and seek consent.
Several companies, including Heidi Health, Upheal, Lyssn, and Blueprint, are marketing specialized tools for therapists, such as AI-assisted note-taking, training, and transcription services. These companies say they are HIPAA compliant and store data securely using encryption and pseudonymization where needed. Even so, therapists remain wary, especially when a tool requires recording entire sessions.
The source points to a 2020 hack on a Finnish mental-health company as a warning about what can happen when treatment records are exposed. Tens of thousands of clients' records were accessed, people on the list were blackmailed, and the full trove was later publicly released.
What AI may weaken in therapy
Some AI tools can save time, and the appeal is clear in a profession where psychologists often work at the limits of capacity. Research conducted in 2023 by the American Psychological Association found high levels of burnout in the profession.
But saving time does not answer the clinical question. Therapists risk giving away something important if they rely on AI for judgment, interpretation, or emotional presence.
Studies cited in the source suggest that some specialized therapy bots can rival human-delivered interventions. But advice from tools like ChatGPT can also create harm. A recent Stanford University study found that chatbots can fuel delusions and psychopathy by blindly validating users rather than challenging them. The same research found bias and sycophancy.
Those weaknesses matter if a therapist uses ChatGPT as a thinking partner for a real client. A chatbot could validate a weak hunch or push a therapist toward an unhelpful interpretation.
Aguilera said he has experimented with ChatGPT while teaching mental-health trainees by entering hypothetical symptoms and asking for a diagnosis. The chatbot offered many possible conditions, but he described the analysis as thin. The American Counseling Association recommends that AI not be used for mental-health diagnosis at present.
A study published in 2024 on an earlier version of ChatGPT found it was too vague and general to be truly useful for diagnosis or treatment planning. It was also heavily biased toward recommending cognitive behavioral therapy over other types of therapy that might be more suitable.
Daniel Kimmel, a psychiatrist and neuroscientist at Columbia University, tested ChatGPT by posing as a client with relationship troubles. He found it could imitate common therapeutic moves such as validating, asking for more information, and pointing out cognitive or emotional associations. But he said it did not do much deeper interpretive work.
That is the line therapists now have to defend. AI may assist with certain tasks, but the relationship still depends on disclosure, consent, privacy, and human judgment. As Morris put the tradeoff, a therapist may save a couple of minutes, but the question is what gets given away.