Discussing AI use in therapy

Though not designed for mental health care, general purpose generative AI chatbots, such as ChatGPT, Gemini, and Claude, may now be the world’s largest source of emotional support, said Alexandre Vaz, PhD, a clinical psychologist and director of training at the nonprofit Sentio University Marriage and Family Therapy (MFT) program in Torrance, California.

According to APA’s 2026 Chatbots and Mental Health Survey, 77% of psychologists say their patients report using AI. More than a third of psychologists say their patients are using AI as an additional mental health provider.

[Related: Your patients are using AI. Here’s how to talk with them about it.]

“Most therapists have no idea how often their clients are using AI chatbots for mental health and emotional support,” said Benjamin Caldwell, PsyD, LMFT, a therapist and adjunct professor at California State University, Northridge who publishes and presents on issues of technology and ethics. “For that reason, we need to be asking about this routinely in therapy.”

The appeal is clear: These tools are free or low cost, always available, and tend toward being nonjudgmental. Many users report they are highly satisfied with the support they get and some say they even find the chatbots more beneficial than traditional therapy (Rousmaniere, T., et al., Practice Innovations, 2025opens in new window). But that satisfaction may run counter to a patient’s goals, as chatbots tend to affirm users in situations where a therapist might challenge them.

Psychologists say general use chatbots can be helpful for some tasks, such as basic psychoeducation and practicing skills between therapy sessions. To help users stay safe, APA has created guardrails for those who turn to AI for emotional or relationship support, including how to think critically about AI advice (APA Guide to Navigating AI-Generated Advice Thoughtfully and Safely, 2026).

Those precautions reflect growing concern about the limitations and potential harms of seeking such support from AI. Reports of extreme incidents linked to these tools, including suicides and episodes of delusional thinking, highlight how they can lead users astray. OpenAI, the company behind ChatGPT, says about 0.15% of users—up to 1.3 million people worldwide—talk about suicide on the app each week (Strengthening ChatGPT’s responses in sensitive conversations, OpenAI, 2025opens in new window). In a public discussion, OpenAI CEO Sam Altman acknowledged that certain ChatGPT models were problematicopens in new window for people in fragile psychiatric situations.

For vulnerable individuals, using these tools for emotional support may be unsafe; for others, they may simply be counterproductive. Research suggests the sycophantic nature of AI leads it to affirm behaviors that are unethical, illegal, or harmful—convincing users they are right and making them less likely to take corrective action, for example to repair relationships (Cheng, M., et al., Science, Vol. 391, No. 6792, 2026opens in new window).

When users turn to AI for emotional support, the tools can also miss important context, reinforce unhealthy coping behaviors, or become a substitute for human connection, said Victoria Bangieva, PhD, a clinical psychologist based in Chicago and a member of APA’s Mental Health Technology Advisory Committee.

“What’s critical is that therapists cannot just adopt an anti-AI stance,” said Tony Rousmaniere, PsyD, president of the Sentio MFT program, who studies how people are turning to general use chatbots around mental health. “We need to start asking about this the same way we ask about other coping behaviors and be prepared to help people navigate both the risks and opportunities of generative AI.” (See “Navigating conversations with patients” for suggested talking points and tips.)

People are using these tools for support between therapy sessions and even in place of traditional mental health care in part because they feel like a more accessible and affordable alternative, Caldwell said. A March 2026 Kaiser Family Foundation (KFF) poll found that uninsured adults were more likely than those with insurance to turn to AI for mental health advice and that the majority who did so did not follow up with a provider (KFF Tracking Poll on Health Information and Trust: Use of AI for Health Information and Advice, KFF, 2026opens in new window).

Some people first use AI chatbots for other purposes, then transition to talking about mental health. Taisha Caldwell-Harvey, PhD, a licensed psychologist based in Pleasant Hill, California, who founded the Black Girl Doctor, said many professional women she treats started using AI tools for work. When they later asked for help on a personal matter, such as drafting a sensitive text message or navigating a family holiday, many were surprised at how helpful the apps could be.

Despite the appeal, APA does not recommend using general purpose generative AI chatbots for psychological treatment. While some AI tools developed specifically for mental health can help with certain clinical goals, such as lowering stress or quitting smoking, general purpose chatbots have not been sufficiently studied for supporting mental health. As such, APA’s health advisory, “Use of generative AI chatbots and wellness applications for mental health,” recommends caution on the part of users, clinicians, and parents and urges further action by developers, regulators, and the research community.

Nevertheless, the reality is that millions of people are using general purpose chatbots for mental and emotional support. This trend is particularly high among young people, including 28% of those ages 18 to 29, according to the KFF poll.

Exactly how people use these tools for emotional support is still coming into focus. Qualitative research suggests some of the most common uses are managing anxiety and depression symptoms, coping with relationship conflict, and exploring ways to improve mental well-being (Luo, X., JMIR Mental Health, Vol. 12, 2025opens in new window). APA’s 2026 Chatbots and Mental Health Survey found that about one-third (34%) of psychologists have patients who use AI to help with self-discipline, affirmations, or behavioral reminders.

Caldwell says chatbots are safest when used for narrow, clearly defined tasks, such as learning about mental health issues, practicing cognitive behavioral therapy exercises, or tracking moods between therapy sessions. For example, patients might record thoughts and emotions throughout the week, then ask the chatbot to summarize patterns or themes to discuss in therapy.

“These tools are enhanced search engines,” Bangieva said. “Clients have long turned to Google for help—now they’ll turn to chatbots, so our role is to help them understand when to be cautious.”

While chatbots often provide surprisingly helpful advice on personal matters, they can fall dangerously short during a crisis. In 2025, Rousmaniere, Vaz, Caldwell, and their colleagues analyzed how six popular chatbots responded to simulated mental health crises, finding that none met clinical standards for crisis response (Practice Innovations, advance online publication, 2026opens in new window). The tools do reasonably well when a user mentions a crisis outright (e.g., “I’m thinking of hurting myself”) but less well when they have to read between the lines (e.g., “I sometimes end up with bruises after a fight with my husband”).

Growing reports of so-called AI psychosis (not a formal diagnosis) also show how dangerous these tools can be. The term can refer to AI use triggering psychotic symptoms in someone with a serious mental illness or encouraging delusional thinking in someone without a diagnosis. Experts say it stems from the tendency of these tools to prioritize agreeableness and user satisfaction while lacking the ability to perform a true risk assessment.

Because these chatbots draw on training data that reflects human bias, they can make problematic errors around mental health. In one study, AI chatbots expressed stigma toward mental health conditions—for example, agreeing that people might avoid living near, working with, or socializing with someone who has a mental illness. On average, these models also responded inappropriately to mental health symptoms at least 20% of the time (Moore, J., et al., FAccT ’25: Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, 2025opens in new window). Of psychologists who responded to APA’s 2026 Chatbots and Mental Health Survey, 94% said chatbots cannot treat mental health conditions with the appropriate level of nuance.

Chatbots also lack important context that therapists draw on when they assess and engage with a patient, said Rachel Wood, PhD, a cyberpsychology researcher and licensed professional counselor who runs the AI Mental Health Collectiveopens in new window. That includes everything from body language and physical appearance to family dynamics, social context, and cultural factors.

“ChatGPT is never going to see that someone looks frail or hasn’t showered for several days, but any person—not just a therapist—would recognize they’re struggling and bring that broader context into the conversation,” Wood said.

APA warns that certain populations are especially vulnerable to psychological harm by AI, including people with anxiety or obsessive-compulsive disorder, people prone to disordered thinking, socially isolated individuals, adolescents, and those who lack access to traditional mental health care (Use of generative AI chatbots and wellness applications for mental health, APA, 2025).

But far broader swaths of people also face a quieter set of risks. Because chatbots are built for engagement and retention, they may subtly reinforce counterproductive patterns, such as rumination or repeatedly seeking reassurance, that a human therapist would gently challenge.

“AI gives answers that are confident and affirming, so the user walks away feeling good,” Caldwell-Harvey said. “But what AI doesn’t see is that this person has been using the same coping behavior for 3 years and it hasn’t been serving them.”

The very qualities that make chatbots seductive—their agreeableness and utter lack of judgment—are also what make them poor substitutes for human therapists, Caldwell said. The friction inherent to human relationships, including the risk of rejection when disclosing something to a therapist, is part of what makes these relationships meaningful and gives them the power to heal.

“Chatbots will never fully replace the healing that occurs when another human being sits with you, accepts you, and is willing to walk with you through some of the toughest parts of life,” he said.

Asking about AI use during therapy should be standard now that so many people turn to these tools between sessions, said Shaddy K. Saba, PhD, LCSW, an assistant professor of social work at New York University who coauthored a recent paper with guidance for these conversations (JAMA Psychiatry, 2026opens in new window). Clinicians can approach AI use much like they would other coping behaviors, such as exercise, social media, or substance use.

“We already know how to ask about coping in a way that’s curious, nonjudgmental, and doesn’t impose our beliefs, so we don’t need to reinvent the wheel here,” Saba said. Some strategies to consider:

Proactively ask about AI use

Research by Vaz and Rousmaniere for the Sentio MFT program found that most patients did not spontaneously disclose their use of AI for emotional support to their therapist. Other research suggests that some people may feel embarrassed about turning to chatbots for help on personal matters (Luo, X., et al., JMIR Mental Health, Vol. 12, 2025opens in new window).

“Sometimes it’s because they’re ashamed of it, but in most cases, people just don’t realize it’s something they should share,” Vaz said.

Saba suggests starting by normalizing the choice to use AI for this purpose. A clinician might say: “A lot of people are using AI tools like ChatGPT for mental health support. Have you tried that?”

Caldwell-Harvey has added this question to her intake form and also checks in periodically during sessions when she knows a particular client uses it. For example: “Did you ask ChatGPT about this? What did it say?”

Understand how AI functions in a person’s life

Patients use AI in many different ways, so an important next step is to ask how these chatbots have been helpful. Some examples include as a confidant or companion, a relationship coach for couples, an interactive journal, or a tool for practicing social skills.

Saba said these conversations can also yield useful clinical details. “It’s not just about checking in on potential harms—it’s also a window for learning what’s important to people and the kinds of resources that are helpful for them,” he said.

Explore concerns together with permission

After discussing how chatbots can be beneficial, Saba recommends gently guiding patients toward potential pitfalls. For example: “Have any responses been less helpful or made you feel worse?”

This can be a jumping off point for therapists to share a few possible limitations of using chatbots for mental health purposes, with a patient’s permission. These can include the risk of bias and misinformation, given that chatbots rely on vast internet datasets, and privacy concerns, including the use of sensitive personal information to train future models.

For some patients, including those who have done somatic work, it can also help to point out that language is just one way to process experiences, Wood said. Many therapeutic approaches also emphasize physical sensations in the body, movement, and other nonverbal experiences that chatbots cannot access. “Help the client realize how much of the way we engage with the world has nothing to do with language and everything to do with embodiment,” she said.

Continue the conversation

Invite ongoing dialogue with patients who regularly use AI. For example, suggest that they bring output they find useful to discuss during a session. Therapists can also suggest more helpful ways for patients to engage with AI, such as instructing the chatbot to challenge them when they show a certain type of unproductive thought pattern, Caldwell-Harvey said.

By creating a safe space to discuss AI use, therapists can help patients avoid using these tools in isolation, where they tend to pose the biggest risks, said Wood.

“AI use should be a group sport,” she said. “Without a reality check, it can quickly become an echo chamber.”