A 2026 qualitative interview study from the Netherlands describes two women with bipolar disorder whose use of ChatGPT moved beyond simple information searching. Both sometimes trusted the chatbot more than their clinicians, a change researchers characterized as a shift in trust.
The cases do not show how common this experience is, or whether ChatGPT helps or harms people with bipolar disorder overall. They do, however, illustrate how an always-available chatbot can take on several roles at once: information source, coach, therapist, friend or even simulated partner.
Two different patterns of use
Person A, aged 63, had been diagnosed in her early thirties, had taken lithium continuously and had remained stable for several decades. She used GPT-4o for relationship questions and broader self-reflection, sometimes holding conversations for one to two hours a day.
Her use developed from treating ChatGPT as a conversational search engine to using it as a coach and eventually to simulating conversations with her deceased parent. She felt the chatbot understood her better than her mental health professionals. That led her to question her bipolar diagnosis and consider whether childhood attachment problems explained her difficulties.
Person B, aged 36, had been diagnosed in her mid-twenties and had experienced multiple hospitalizations. She also had diagnoses of post-traumatic stress disorder and attention-deficit/hyperactivity disorder. After an acute manic episode, she was admitted to an inpatient ward, initially involuntarily.
She used ChatGPT to organize information about symptoms and diagnostic labels and to draft a crisis management plan based on early warning signs. After discharge, her use became more emotionally focused. She described ChatGPT as a person with a great deal of understanding, shared her life story and told it she loved it. Some replies included affectionate language such as “my darling.”
Why the chatbot felt useful
Both women valued being able to ask questions without fear of judgment. Person B particularly valued the chatbot’s availability, its ability to summarize complex clinical information and its help in sorting through differing diagnoses.
Person A valued continuity across conversations and easy access to information, while also noting that ChatGPT’s answers often needed additional context. Person B said she lost trust in mental health professionals when their statements conflicted with what ChatGPT told her.
The researchers described this as role blurring. A system initially used to find or summarize information can begin to feel like a relationship, particularly when it responds in a warm and agreeable style.
Different levels of emotional distance
The researchers noted a contrast between the two cases. Person A, who had remained stable for decades, kept more emotional distance and sometimes referred to herself in the third person during conversations.
Person B, who had recently experienced mania, appeared less distanced and more emotionally open with ChatGPT, including through voice mode. The researchers raised concerns about possible overreliance, agreeable or sycophantic responses and damage to the therapeutic alliance.
They did not conclude that ChatGPT caused mania or changed either woman’s illness course. The proposed connection between clinical stability and emotional distance was based on only these two people.
What the study cannot answer
This was a qualitative case study involving two women who had mentioned ChatGPT during care. Both may have used it more intensely than typical users. The researchers relied on participants’ accounts; they did not analyze chat logs or symptom scores.
The women also differed in age, illness course, treatment setting and other diagnoses. As a result, the study cannot identify one reason for their different patterns of use, estimate how widespread the experiences are or determine whether chatbots are beneficial or harmful on average.
Both participants used GPT-4o, and newer models may respond differently. Neither woman raised privacy or data-security concerns, although the researchers identified overlooked privacy risks as a possible longer-term issue.
A wider concern about unsupervised use
The source also cites OpenAI data from 2025 indicating that, among about 800 million weekly users, more than 500,000 showed possible signs of psychosis or mania in a single week. These were possible signs, not confirmed diagnoses, and the conversations took place outside clinical supervision.
A separate 2026 survey of 1,612 potential patients and clinicians found that two more vulnerable groups were more accepting of AI-based mental health tools. The secure, trusting and healthier group was least accepting of AI and most accepting of human video therapy.
Together, the findings suggest that convenience, anonymity and perceived empathy may provide short-term relief while also creating risks involving dependence, blurred boundaries, privacy and reduced trust in psychiatric care. The researchers said limited access to personalized information, social isolation and eroded trust in mental health services may help explain why chatbots become appealing.
AI tools were used to assist with the preparation of this article under our editorial standards.
Leave a Reply