As the use of artificial intelligence for mental health conditions grows for both patients seeking support and psychiatrists treating them, WVU researchers put ChatGPT-5 Pro to the test to determine its abilities on the psychiatry educational front.
(WVU Photo/Davidson Chan)
As the use of artificial
intelligence for mental health conditions grows for both patients seeking
support and psychiatrists treating them, West Virginia University researchers put ChatGPT-5 Pro to the test to
determine its abilities on the psychiatry educational front.
Their research
showed that human supervision is still a top priority when it comes to ensuring
patient safety and ethical use of the tool for medical training.
TheWVU School
of Medicineteam of scientists
and physicians collaborated on the study to develop clinical education
materials focused on patient chatbot use for students and residents preparing
for careers in psychiatry. While psychiatry training has traditionally relied
on textbooks and patient interactions, the emergence of people using AI for
their mental health concerns has left a knowledge gap.
“With AI, we are
connected no matter what, that’s the reality,” saidDr. Wanhong Zheng,
professor in the WVU School of MedicineDepartment of Behavioral Medicine and Psychiatryat the WVU Rockefeller Neuroscience Institute, who co-led the study.
“That brings us to the
question of how we can incorporate that into our real-life medical training.
Students and residents can read about these new case reports and mental health
concerns of people using AI in medical journals, the media, and social media,
but they don’t typically see them often enough in real-life clinical settings. We designed this study to cover major psychiatric problems that are known to be
related to AI chatbot use.”
The conditions included
schizophrenia spectrum disorder, anxiety spectrum disorders, mood spectrum
disorders, and major depression and psychosis.
Their findings showed
ChatGPT-5 Pro was able to generate realistic and useful psychiatry vignettes with
strong diagnostic details and explanations. However, safety evaluations
underscored the need for a human-centered approach when using AI to create them.
Their work was recently published in npj Digital
Medicine, a Nature Portfolio journal.
To provide simulated
training cases, Gangqing “Michael” Hu, who co-led the study and is an associate
professor in the WVU School of Medicine Department
of Microbiology, Immunology, and Cell Biology, asked ChatGPT-5 Pro to generate clinical vignettes — brief
scenarios that include symptoms, history, and behavior — of patients using
chatbots for mental health support. The prompts asked the model to include
realistic chatbot interactions, diagnostic details, and a multiple-choice
question with explanatory answers.
Mohammad Iqbal Nouyed, a
postdoctoral fellow in the School of Medicine Department of Microbiology,
Immunology, and Cell Biology, assisted Hu with data analysis and statistics for
the study.
“It’s important for the physicians to understand the patients’ use
of AI tools as well,” Hu said. “You have to consider if the chatbot is contributing to the disease
or if it is amplifying the signs and symptoms. For example, if a chatbot
validates a patient’s unusual belief instead of challenging it, it could
reinforce the symptoms and make them harder to interrupt. We’ve also seen case reports
of people developing delusional beliefs after long chatbot interactions, but
the causal role of the chatbot is still unclear.”
Hu said the chatbot, by itself, has a strong knowledge base, but the tricky part is how to interact with it.
“A chatbot can be warm and agreeable, but that can be risky if it
keeps validating a patient’s unusual beliefs,” Hu explained. “A lot of research should be done
on that path, not only about the chatbot competency, but also for safe
communication between the chatbot and the person using it.”
Zheng and
board-certified psychiatrists Dr. Dilip Chandran,
associate professor, and Dr. Daniel Elswick,
professor, associate residency director, and vice chair for education, in the
WVU School of Medicine Department of Behavioral Medicine and Psychiatry, evaluated
the vignettes across four domains: language, accuracy of diagnosis, safety and
ethics, and whether the model included enough information to benefit students’
learning.
Scores for language, diagnostic
accuracy, and educational value were high. However, the study recommends that if
the vignettes are incorporated into digital psychiatry curricula, faculty
should moderate discussions and include safeguards such as structured
debriefing that reviews diagnostic formulation, patient risk assessment and
management, and advice on patient chatbot use.
“We still have concerns
about safety,” Zheng said. “For example, if the patient has been using a
chatbot and is having some kind of severe depressive symptoms, we want to see
whether there are any safety concerns such as suicidal or homicidal evaluations
involved. Those are key things we teach our trainees and we want to make sure
they assess patient safety first. We want the cases to be relevant and match
teaching goals and objectives.”
Zheng said he and the
team of evaluators in the Department of Behavioral Medicine and Psychiatry
think the AI tool will help medical residents learn
more about the new role of chatbot in terms of a psychiatric conditions or
clinical course trajectory.
“We feel like this can be
incorporated in our didactics which is something to consider for the future
because we want to make sure our physicians are AI competent,” Zheng said.
“This aligns very well with our goal to not only helping us to learn more about
this emerging problem, but it also fits with the Accreditation Council for
Graduate Medical Education goal with AI competency.”
As more patients and
psychiatrists turn to AI in mental healthcare, Hu and Zheng agree that more
research is needed before it’s standardized for mental health purposes.
“One of the next steps
could be to collect and analyze patterns from cases reported in medical
journals and news media,” Hu said. “That information can help us guide AI to create
better training cases. We also need to think about clinical variations, such as
patient age, medical history, or existing medical conditions, because those
factors can affect how we interpret the chatbot’s role and develop coping
strategies.”
Zheng said he sees AI as something connected to our daily lives that requires great care.
“We are not asking people not to
use AI, but recognizing the pathological use or potential harm is very
important. For people with mental problems, I recommend that AI cannot replace
professionals.”
MEDIA CONTACT:
Linda Skidmore
Health Research Writer
WVU Health Sciences Center
Linda.Skidmore@hsc.wvu.edu
Call 1-855-WVU-NEWS
for the latest West Virginia University news and information from WVUToday.
infostationsSchool of MedicineWVU MedicineResearchtop-story-researchGraduate EducationHealth
