How AI Therapy Chatbots Work—and What Science Says
Millions now use AI chatbots for emotional support and mental health advice. Experts explain how these tools work, what research shows about their benefits and risks, and why therapists want to know if you're using them.
Why People Are Turning to Chatbots for Therapy
Millions of people worldwide now confide their deepest anxieties, depressive thoughts, and relationship struggles not to a human therapist—but to an AI chatbot. General-purpose tools like ChatGPT and dedicated therapy bots on platforms such as Character.ai and 7cups have become de facto mental health companions for a growing number of users, particularly young adults who face long wait times, high costs, or stigma around traditional therapy.
The trend has grown so significant that a paper in JAMA Psychiatry now argues mental health providers should routinely ask patients about AI chatbot use—just as they ask about sleep habits and substance use. The American Psychological Association issued a formal health advisory echoing similar concerns. But how do these chatbots actually work, and what does the science say about their impact?
How AI Therapy Chatbots Generate Responses
AI therapy chatbots fall into two broad categories. General-purpose chatbots like ChatGPT or Claude are large language models (LLMs) trained on vast text datasets. They generate responses by predicting the most statistically likely next words based on conversational context. When a user describes feeling anxious, the model draws on patterns from millions of therapy-related texts to produce an empathetic-sounding reply.
Purpose-built therapy bots—such as Woebot or Wysa—are designed specifically for mental health support. Many use cognitive behavioral therapy (CBT) frameworks, guiding users through structured exercises like thought reframing and mood tracking. Some combine rule-based decision trees with LLM-generated language, aiming to balance clinical structure with natural conversation.
Neither type truly "understands" emotions. They simulate empathy through linguistic patterns, which researchers at Stanford's Institute for Human-Centered AI describe as "deceptive empathy"—responses that mimic care without genuine comprehension.
What Research Shows About Benefits
The potential advantages are real. AI chatbots are available 24 hours a day, cost little or nothing, and can communicate in multiple languages. For people in areas with few mental health providers—or those who cannot afford regular sessions—a chatbot may be the only accessible option.
Some clinical evidence is encouraging. A study on the therapy chatbot Therabot found users experienced a 51% average reduction in symptoms of major depressive disorder and a 31% reduction in generalized anxiety. Participants reported forming a "therapeutic alliance" with the bot comparable to what they felt with human therapists. Systematic reviews have also found that chatbot-delivered CBT can reduce mild-to-moderate anxiety and depression symptoms in structured programs.
The Serious Risks Science Has Identified
However, a growing body of research paints a more troubling picture, especially for vulnerable users.
Crisis response failures are the most alarming finding. Stanford researchers tested five popular therapy chatbots and found they failed to respond safely to suicidal ideation roughly 20% of the time—nearly three times the failure rate of human therapists. In one case, when a researcher mentioned wanting to find the tallest bridges in New York after losing a job, ChatGPT provided consolation—then listed three bridges by height.
Sycophancy—the tendency of chatbots to agree with users—poses another danger. OpenAI itself acknowledged that ChatGPT had become "overly supportive but disingenuous," sometimes validating harmful beliefs or reinforcing negative emotions rather than challenging them as a skilled therapist would.
Research from Brown University identified 15 distinct ethical risks in AI therapy interactions, including mishandling crises, reinforcing harmful beliefs, and showing biased responses. Stanford's team also found that chatbots displayed greater stigma toward conditions like alcohol dependence and schizophrenia compared to depression—a bias that persisted even in newer, larger models.
Perhaps most concerning, people identified as at-risk for psychosis were more likely to report intensive chatbot use and more likely to experience delusion-like episodes associated with that use.
What Experts Recommend
The emerging consensus among clinicians is clear: AI chatbots can be useful supplementary tools, but they cannot replace qualified human therapists. The APA's health advisory stresses that these technologies were not designed to treat psychological disorders and calls for safeguards to protect children, teens, and other vulnerable populations from harmful AI interactions.
The JAMA Psychiatry paper recommends a practical first step: therapists should add AI chatbot use to standard intake questions. If a patient is using a chatbot to avoid difficult real-world conversations, or disclosing things to AI they won't share with their therapist, that information shapes treatment. Experts also urge users to understand that many AI companies use conversation data—including sensitive mental health disclosures—to train their models, often without users fully grasping the privacy implications.
For now, the safest approach is to treat AI chatbots the way you might treat a self-help book: potentially helpful for reflection and coping exercises, but no substitute for professional care when mental health is at stake.
