top of page
Search

The Chatbot Followed You Into the Room: What 311 Psychiatrists Told Me About AI and Real Care

4 days ago
7 min read

Imagine, just for a moment, sitting in a therapist’s waiting area with your phone in your hand. You have already told a chatbot the whole story: what happened, what you are worried about, maybe even the parts you have not said out loud to anyone. Now you are wondering whether to mention it in the appointment.

I want to say this plainly: you do not need to hide it or apologize for it. If an AI conversation helped you find words, that can be useful information in therapy. The important question is not whether AI is good or bad. It is what happens when someone’s AI conversation becomes part of the care they receive from another person.

What the study actually did

In a study published on June 2, 2026, a research team in South Korea asked psychiatrists to describe their experiences with generative AI in psychiatric care. The paper, “Mapping Practice-Based Signals of Generative AI in Psychiatric Care: Qualitative Study of Korean Psychiatrists' Experiences, Interpretations, and Implementation Priorities,” appeared in the Journal of Medical Internet Research.

The researchers analyzed written answers to three open-ended questions: about memorable clinical experiences involving AI, how AI compares with human therapists, and what is needed to introduce AI safely. Invitations went out through the Korean Neuropsychiatric Association between October 27 and December 26, 2025. Of 408 people who took part in the wider survey, 311 provided at least one meaningful written answer and formed the qualitative sample.

These were psychiatrists’ accounts, not a review of actual chatbot conversations. I find that distinction important: the study helps us understand what clinicians were noticing in care, but it does not tell us what every patient’s AI use looks like.

Four ways AI showed up in care

The psychiatrists described AI entering care in several different ways. These are not four kinds of people; they are four kinds of moments.

Patients used AI as a first doorway. Some people turned to chatbots to organize feelings, try self-help, or check whether they might need support. One psychiatrist described a patient who asked a chatbot not to comfort them, but to help organize their situation and emotions. The patient found words for feelings that had been hard to process, and the psychiatrist observed that symptoms improved. Another patient sought care after an AI conversation advised them to do so.

Clinicians used AI for structured tasks. Psychiatrists described using AI to summarize records, prepare patient education materials, compare medications, or find quick information. They saw potential in repetitive, information-dense work, especially when a clinician reviewed the output. One psychiatrist cautioned that medically inaccurate statements could be noticeable and concluded that it was too early to put such systems in front of the public for counselling, diagnosis, or treatment without expert review.

A therapist and client talk together in a warm, comfortable room with pale sage and blush accents

AI became a relational object. Some patients described chatbots as companionship or a safer place to disclose. One psychiatrist said a socially isolated patient’s sense of isolation eased through chatbot use. Another described a patient who had rarely shared inner thoughts with anyone becoming more able to express feelings in treatment, and trying new interpersonal behavior with the chatbot’s help. A psychiatrist also noticed chatbot conversations increased when a patient relapsed and wondered whether those interactions could help signal a change. These observations are not proof that chatbot use predicts relapse; they show the kinds of questions clinicians were beginning to ask.

AI changed what happened in the appointment. Patients brought AI-generated summaries, discussed earlier chatbot conversations, or used AI to organize what they wanted to say. One psychiatrist found a patient’s summaries of AI conversations genuinely useful for making limited clinic time more efficient. But some patients also compared a clinician’s advice with chatbot answers, or asked for medication a chatbot had suggested.

The good it was doing deserves room

I do not want the risks to erase what felt helpful. A chatbot may be available when a person is not ready to call someone, may feel less exposing than asking for help face to face, and may help someone prepare to talk. For one patient in the study, AI was a step toward care. For another, it helped make difficult feelings easier to put into words.

As one psychiatrist put it, “AI chatbots have the advantage of immediacy and accessibility... they should be understood not as therapeutic replacements, but as entry points for those who cannot access treatment and as tools that support human care” (P33).

That is a generous and grounded way to think about it: a tool can help someone approach the door without being the whole room.

Where it got complicated

The clearest clash came when a patient described symptoms to a chatbot, received a diagnosis, and arrived angry that it differed from the psychiatrist’s assessment. The patient said they could no longer trust the psychiatrist. The psychiatrist’s diagnosis turned out to be correct; the chatbot had been given only the symptoms the patient recognized themselves.

The study also included serious safety concerns. One psychiatrist described a patient who asked a chatbot about medication, was told it was not a high dose, and later took an overdose for self-harm. In another account, a patient with suicidal thoughts received a crisis hotline number during chatbot use and was linked to services. I share these examples carefully: they are reports from clinicians, not a reason to assume that using a chatbot is dangerous.

The authors describe the central tension as an “access-protection tension”: AI can make it easier to reach support, but that access may not come with a reliable floor of protection.

Standardized and tireless, but relationally thin

Psychiatrists valued AI’s consistency, neutrality, and availability. They also noted that a chatbot cannot read nonverbal cues, respond to emotional nuance as a person does, or participate in the mutual human exchange that can matter in therapy.

One psychiatrist summarized the balance this way: “The strengths of AI chatbots are that they are standardized, evidence-based, and less prone to error. Their limitations are the lack of individualized responses that respect each person's uniqueness and the limited ability to convey empathy” (P15).

The same quality can help in one situation and complicate another. A nonjudgmental response may make it easier to disclose something shameful. But, as another psychiatrist cautioned, “Its advantages are that it is always available and allows disclosure of shameful content. Its risks are that users may hear only what they want to hear, while delusional thinking may be reinforced and social isolation may increase” (P229). The concern is not support itself. It is support that never includes a question, a limit, or a check on reality.

A phone rests face up on a pale linen sofa beside a notebook and a cup of tea

Vulnerability amplification, gently understood

The authors call one pattern “vulnerability amplification.” In plain language, a response that feels containing in a lower-risk moment may land differently when someone is dealing with impaired reality testing, compulsive reassurance seeking, severe mood dysregulation, or suicidal intent.

This is not a claim that AI causes psychosis, or that using a chatbot is inherently unsafe. It is a reminder that an endlessly available, agreeable listener may affect people differently depending on what they are carrying. That is why human oversight matters most when the situation is especially serious.

If you use both, here are a few ways to make it easier

You deserve to bring your whole experience into care, including the parts that happened on a screen.

  • Tell your therapist you have been using a chatbot. It is information, not a confession.

  • Bring a summary if it helps you say the hard thing. You can also bring the question or feeling that stayed with you.

  • Notice whether the chatbot is the only place you say what is true for you. That may be a gentle invitation to find another safe person or space.

  • Use it to organize your thoughts, not to diagnose you. A chatbot cannot know the full context of your life or assess you in the way a clinician can.

  • Keep a human in the loop for anything crisis-shaped. If you may be in immediate danger, call or text 988, or call 911.

What this means for us as therapists

The authors suggest psychiatrists may increasingly act not only as providers of care, but also as interpreters, boundary-setters, and safety supervisors around patients’ AI interactions. They also note that how clinicians talk about using AI matters, because disclosing a physician’s AI use can affect how competent, trustworthy, and empathetic the public perceives that physician to be.

For me, the practical starting point is curiosity. I would rather ask what your phone has been helping you do than police whether you used it. We can look together at what felt useful, what felt unsettling, and what you want from care that a chatbot could not provide. That conversation can be respectful, collaborative, and centered on your goals.

Two adults in easy conversation on a light sofa in a sunlit room with soft curtains

When it may help to talk

You might want support if an AI conversation left you more confused or distressed, if it is becoming difficult to trust your own judgment, or if you would like help making sense of what you have been carrying. You do not have to wait until things feel unmanageable. Reaching out can be one small way to move toward support that fits you.

A practical next step

I offer a free 15-minute consultation so you can get to know a therapist before deciding what comes next. You can also book online. Therapy is available in many different languages, and insurance is accepted, including UHC, Aetna, and Cigna.

Call 404-369-3838 to get in touch. Care is provided only in the State of Georgia. For online sessions, you must be physically located in Georgia at the time of your session. This is not a crisis service. For immediate danger, call or text 988 or call 911.

The study’s message is not that you need to choose between a tool that helped you find words and a person who can listen. We can make space for both. And when care feels difficult to navigate, the relationship we build together can be the thing that holds.

Source: Kim M, An Y, Jeon M, Lee Y, Lahcine OF, Kim H, Lee S, Lee S, Yang JC, Jeon SW, Jung D, Cho CH (2026), “Mapping Practice-Based Signals of Generative AI in Psychiatric Care: Qualitative Study of Korean Psychiatrists' Experiences, Interpretations, and Implementation Priorities,” Journal of Medical Internet Research, vol 28, e96556, published June 2, 2026. https://doi.org/10.2196/96556. Open access under CC BY.

Limits: The data were short written survey answers rather than interviews, so many answers were brief and light on context. Patterns across questions are descriptive and come from a single point in time, so they do not show cause and effect. The study recorded psychiatrists’ accounts of patients’ AI use and did not verify the actual chats, prompts, or platforms. Only psychiatrists were surveyed, so patients’ own views are missing entirely. Respondents were psychiatrists in South Korea, a country with very high digital uptake and its own particular pressures around psychiatric records and employment, so these findings are not estimates for Georgia or for any particular person. What travels is the pattern, not the numbers applied to your family.

 
 
 

Comments


bottom of page