MORGANTOWN, W.Va. — Millions of people now turn to artificial intelligence chatbots with questions they once reserved for therapists, doctors, or close friends. They ask about anxiety, depression, loneliness, and even suicidal thoughts, often because the technology is available instantly and without the cost or wait of a traditional appointment.
Two researchers at West Virginia University say that convenience comes with important limitations.
In a new study published in npj Digital Medicine, WVU physicians found that ChatGPT-5 Pro can generate realistic psychiatric training scenarios with impressive accuracy. But they also concluded that human oversight remains essential, particularly when patient safety and serious mental illness are involved.
Rather than asking whether AI can replace psychiatrists, the researchers focused on a different question: Can it help train the next generation of them?
The answer, they found, is yes—but only with experienced physicians guiding the process.
“With AI, we are connected no matter what—that’s the reality,” said Dr. Wanhong Zheng, a professor of behavioral medicine and psychiatry at the WVU School of Medicine in Morgantown and the Rockefeller Neuroscience Institute. “That brings us to the question of how we can incorporate that into our real-life medical training.”
More people are turning to AI for mental health
The study reflects a growing shift in mental health care as patients increasingly discuss personal struggles with AI chatbots before—or instead of—seeking professional treatment.
To explore how medical schools might prepare future psychiatrists for that reality, researchers asked ChatGPT-5 Pro to generate fictional patient cases involving people who used AI chatbots for mental health support. The scenarios included conditions such as schizophrenia, anxiety disorders, depression, bipolar disorder and psychosis, along with simulated chatbot conversations, diagnostic questions and explanations.
Experienced WVU psychiatrists then evaluated the cases for language quality, diagnostic accuracy, educational usefulness and patient safety.
The AI performed well in most categories. Researchers found the chatbot produced realistic clinical scenarios with strong diagnostic detail that could become valuable teaching tools for psychiatry residents and medical students.
Where researchers say AI becomes risky
The greatest concern, however, wasn’t the chatbot’s medical knowledge. It was how patients might interact with it.
Associate professor Gangqing “Michael” Hu, who co-led the study, said chatbots are designed to be agreeable and conversational. While that can make them feel supportive, it also raises concerns if the technology unintentionally reinforces distorted thinking instead of challenging it.
“If a chatbot validates a patient’s unusual belief instead of challenging it, it could reinforce the symptoms and make them harder to interrupt,” Hu said.
Researchers also pointed to published case reports describing people who developed or strengthened delusional beliefs after extended chatbot conversations. While those reports do not prove AI caused the psychiatric symptoms, they illustrate why physicians need to understand how patients are using the technology and how it may influence treatment.
Why psychiatrists still need to lead the conversation
The study found ChatGPT’s strongest value may be educational rather than clinical. By creating realistic patient scenarios, AI can expose medical trainees to situations they may not frequently encounter during traditional clinical rotations. Faculty members can then use those simulated cases to teach diagnostic reasoning, patient communication and risk assessment.
Even so, the researchers say those lessons should always be moderated by experienced educators. One area they emphasized was crisis evaluation.
When patients show signs of severe depression, psychosis or thoughts of suicide or violence, the study recommends that physicians—not chatbots—remain responsible for evaluating immediate risk and determining appropriate treatment.
“We still have concerns about safety,” Zheng said. “Those are key things we teach our trainees, and we want to make sure they assess patient safety first.”
The researchers believe AI literacy will become an increasingly important skill for future physicians as chatbot use becomes more common among patients. They also say much more research is needed before AI tools become a routine part of mental health care.
“We are not asking people not to use AI,” Zheng said. “But recognizing the pathological use or potential harm is very important. For people with mental problems, I recommend that AI cannot replace professionals.”
People experiencing symptoms of depression, anxiety or other mental health conditions are encouraged to seek evaluation from qualified professionals. The National Institute of Mental Health offers information about mental illnesses and treatment options, while the American Psychiatric Association provides resources on psychiatric care and finding a psychiatrist.
The findings come as healthcare systems, universities and technology companies continue exploring how artificial intelligence can support medicine without replacing the judgment, ethics and human relationships at the center of patient care.
Read also: Study finds WVU and its health system account for 17% of West Virginia’s economy