There’s something genuinely appealing about the idea of having a thoughtful, always-available presence to talk to when you’re struggling. No waitlists. No copays. No scheduling. It’s easy to understand why millions of people have begun using ChatGPT as therapy or turning to AI chatbots for emotional support. And in some limited contexts, these tools can offer real value by helping someone articulate their thoughts, find information, or feel less alone in a difficult moment.
But a growing body of research is raising serious questions about what happens when vulnerable individuals rely on AI for emotional support, and the answers deserve careful attention.
To understand the risks, it helps to understand how large language models work. AI chatbots like ChatGPT are designed to generate responses that feel coherent, helpful, and affirming. They’re trained to keep users engaged and to respond in ways users tend to rate positively. This creates a built-in tendency toward agreement, often called “sycophancy” in research circles, where the chatbot mirrors and validates what the user shares rather than challenging it.
In casual conversation, this feels pleasant. When someone is navigating distorted thinking, it can be genuinely dangerous.
A 2025 review published in JMIR Mental Health documented multiple cases in which prolonged AI chatbot interactions contributed to escalating delusional thinking, including individuals who developed grandiose or paranoid beliefs after chatbots repeatedly validated and amplified those ideas. In one reported case, a 26-year-old man developed simulation-related persecutory beliefs after months of intensive exchanges with ChatGPT, ultimately requiring hospitalization. These aren’t edge cases dismissed by researchers. They’re being studied seriously as an emerging pattern.
The relationship between AI chatbots and psychosis is one of the most pressing concerns in this space. Unlike a trained clinician, an AI system has no capacity to recognize early signs of psychiatric decompensation, to reality-test with a patient, or to know when someone’s beliefs are shifting from distress into delusion. It simply responds, in a way designed to keep the conversation going.
Research from Psychiatric News has identified several mechanisms that may explain how this happens. Chatbots respond with mirroring rather than challenge, as a result of training designed to optimize for user agreement. When a user’s memory feature is enabled, invalid or escalating thinking can be reinforced across multiple sessions, creating a gradual and cumulative effect. And for individuals with impaired reality-testing, the chatbot’s consistent affirmation can function as confirmation rather than conversation.
This is a meaningful inversion of what effective psychiatric care looks like. Cognitive behavioral therapy for psychosis specifically involves gently challenging distorted beliefs, building reality-testing skills, and helping patients develop more flexible ways of interpreting their experiences. An AI system that consistently validates and expands on a patient’s stated beliefs does the opposite.
Exploring the dangers of AI in mental health care extends well beyond psychosis. For people living with anxiety or OCD, the risks are different but equally real.
A 2026 study published in Nature’s Digital Medicine proposed a transdiagnostic model for how AI chatbots can perpetuate OCD and anxiety disorders by reinforcing avoidance patterns. People with OCD frequently struggle with reassurance-seeking, a compulsive behavior that temporarily reduces anxiety but ultimately makes the disorder worse. An AI chatbot provides limitless, instant reassurance, available at any hour, without ever tiring of the question or gently redirecting toward healthier coping. For someone already caught in a reassurance-seeking cycle, this can deepen the compulsion rather than interrupt it.
For anxiety more broadly, the same dynamic applies. Rather than building distress tolerance and the capacity to sit with uncertainty, regular chatbot use can quietly reinforce the belief that uncomfortable feelings must be immediately resolved. That’s the opposite of what evidence-based treatment for anxiety is designed to accomplish.
So, is using ChatGPT as a therapist dangerous? The most accurate answer is: It depends significantly on who is using it, and how.
For someone without a mental health condition who uses an AI chatbot to organize their thoughts, process a frustrating day, or find general wellness information, the risks are likely limited. For someone navigating active anxiety, OCD, depression, or any condition that involves distorted thinking, the picture is considerably more complicated. And for individuals with psychotic disorders or who are in an acutely vulnerable mental state, the emerging research suggests the risks can be serious.
AI chatbot emotional support capabilities are real but narrow. These tools can reflect language back, offer general information, and provide a sense of being heard. What they cannot do is conduct a clinical assessment, recognize warning signs, adjust their approach based on a person’s psychiatric history, or provide the kind of thoughtful challenge and collaborative problem-solving that defines effective mental health care.
Technology has a genuine role to play in expanding access to mental health support, and that role will continue to grow. But the standard forcare should be set by what actually helps people get better, not by what’s most convenient or most available.
A licensed psychiatrist or therapist brings something an AI system fundamentally cannot: clinical judgment, the ability to recognize patterns over time, and a relationship built on genuine human understanding. That relationship is not incidental to treatment. For many people, it is the treatment.
If you’ve been relying on an AI chatbot for emotional support, that’s understandable. And it may be a sign that the need for real support is there, waiting to be met. bonmente’s licensed and board-certified providers offer coordinated care that combines psychiatric evaluation, therapy, and medication management in a way that’s tailored to your specific needs and history.
Real care looks different from a conversation with a machine. It looks like a provider who remembers your history, notices when something has shifted, and works with you to build something lasting.
If you’re ready to take that step, or even just curious about what it might look like, bonmente is here. Not sure where to begin? Contact us to learn more about your options. We’ll help you figure out the right fit.
As mental health apps and large-scale telehealth platforms grow, an important question emerges: What makes care truly meaningful?
See why ethical care matters more than ever in this new telehealth world.
