How chatbots are enhancing, not replacing, clinician training

Clinical training in mental health has been limited by a chronic bottleneck: There are not enough supervision hours available for trainees to receive intensive, high-quality feedback. To close this training gap, psychologists are developing and exploring artificial intelligence tools to help both new and seasoned clinicians master and retain skills at any career stage and ultimately provide better care for patients.

AI training apps have the potential to offer a safe place to make mistakes and to deliver fine-grained feedback throughout a mock session, along with many more opportunities to practice important skills. And, like any AI tool, they run the risk of generating feedback that sounds authoritative but may be clinically questionable—or may be outside the scope of a trainee’s capabilities.

As demand for mental health care continues to surge, practitioners are paying attention to research findings that evaluate the effectiveness of these apps—and the early evidence hints that AI-assisted training could reshape clinical supervision. Some psychologists are experimenting with this promising yet complex use of technology by securing research funding to create their own AI training tools, testing these tools with supervisees, and advising others who are trying to incorporate the applications into training practices.

“I’ve been involved in clinical training for 50 years, and I’m very optimistic about the potential of AI because up until now, supervisors and clinical trainers have not done a very good job of introducing skills and providing precise, immediate feedback in training,” said Hanna Levenson, PhD, a professor emerita at the Wright Institute in Berkeley, California, who has worked with colleagues to develop an AI platform to help trainees learn short-term dynamic therapy. “Now is the time for psychologists to get involved in discovering how to leverage AI safely because ultimately this can improve care for our patients.”

Psychologists developing these tools emphasize that the chatbots are designed to supplement—not replace—human supervision and are working with tech teams to ensure the technology is safe, effective, and user-friendly for supervisors and trainees.

“We are also gathering feedback from clinicians about how they want to receive feedback from a chatbot, such as the frequency, the tone, and the content of the guidance they receive,” said Shannon Wiltsey Stirman, PhD, a professor of psychiatry and behavioral sciences and codirector of the Center for Responsible and Effective AI Technology Enhancement of PTSD Treatments (CREATE) at Stanford University. “We want to make sure the feedback is given in a way that clinicians want to hear.”

One persistent challenge in supervision is that psychologists often cannot directly observe how trainees engage with patients in each session because it is not feasible for a supervisor to review entire sessions regularly in addition to maintaining their own caseload. Trainees are often in the position of reporting their own work to supervisors. Torrey Creed, PhD, vividly remembers during training how she met with her supervisor once a week to review a dozen cases. “My recall was filtered through my perception of the sessions, and there was so little time to talk about an entire caseload,” said Creed, an associate professor of psychology in the Perelman School of Medicine at the University of Pennsylvania who partnered with a team that has developed an AI platform for training and quality improvement called Lyssn.

The supervision literature on therapist self-report suggests that therapists may not be the best judges of their own abilities, often under- or overestimating their skill level (Hogue, A., et al., Administration and Policy in Mental Health and Mental Health Services Research, Vol. 42, 2015opens in new window; McManus, F., et al., British Journal of Clinical Psychology, Vol. 51, No. 3, 2012opens in new window).

Researchers at Stanford University recently tackled this problem by developing an AI chatbot called TherapyTrainer that tracks what a trainee is saying in a session. Funded by an $11.5 million grant from the National Institute of Mental Health (NIMH), they created a simulated therapy room in which providers use text or audio to chat with a simulated patient to practice written exposure therapy (WET)—a treatment that focuses on writing about traumatic memories (Stade, E. C., et al., Cognitive and Behavioral Practice, 2025opens in new window). One of the AI patients is a veteran who experienced a traumatic event in the military related to hazing; the second virtual patient experienced intimate partner violence and is in remission from substance use disorder.

Therapists and trainees can practice all aspects of WET, such as providing psychoeducation about post-traumatic stress disorder (PTSD), explaining the treatment rationale for WET, and conducting the narrative writing session. The AI consultant operates as a supervisor would by providing feedback to the therapist about the interactions with the AI patient. The therapist can pause the session at any time and ask the AI consultant questions such as “How am I doing?” and “What should I do next?”

The researchers are launching two clinical trials in the summer of 2026 to study how TherapyTrainer compares with human consultation. “Consultation is one of the gold standards for training therapists in evidence-based treatments, but the problem is that most therapists who have learned a treatment do not have the opportunity to get consultation because there are not enough consultants and the cost is high,” said Elizabeth Stade, PhD, a computational clinical psychologist at Stanford’s Institute for Human-Centered AI and an associate director at CREATE. “And many therapists are not able to give up billable hours to attend consultation.”

Tim Overton, PsyD, who runs a private practice in Philadelphia, used TherapyTrainer to refresh his WET skills even though he was initially skeptical about the value of AI in training. “When providers are learning a new therapy technique, trainings include vignettes and possibly consultation on one or two cases, but this is not enough to become proficient,” he said. “The AI trainer helped me identify certain aspects of the treatment that I was not prioritizing.”

It is also a tool he wishes had existed when he started working with veterans early in his career, when he lacked a military background or Veterans Affairs (VA) training. “This type of tool can give budding clinicians access to practice with a new population to prepare them to implement evidence-based treatments,” Overton said.

While many AI training applications focus on techniques associated with cognitive behavioral therapy (CBT) interventions, these tools can also help trainees learn depth-oriented approaches related to unconscious mental processes. Levenson and colleague, clinical psychologist Josh Krieger, PsyD, developed an AI platform called LearnTLDPopens in new window that gives trainees immediate, actionable feedback on their time-limited dynamic psychotherapy skills as they work with virtual patients. “They also have the opportunity to question or even challenge an AI supervisor regarding their feedback without fear of academic reprisals, which promotes a true depth of understanding,” said Levenson.

Human supervision is extraordinarily time-intensive, and AI tools are giving psychologists the ability to evaluate trainees much more efficiently. “The traditional process of coding to evaluate the quality of therapy sessions requires specialized training and hours of manual coding per session,” said Milena Esherick, PsyD, who consults with companies that are developing AI training tools and is the associate director of the Master of Science in Counseling Program at Kaiser Permanente School of Allied Health Sciences in California. “AI can quickly analyze every session and has the capacity to evaluate how trainees are doing on the clinical skills most tied to client outcomes, such as the therapeutic alliance, empathy, goal consensus, and treatment fidelity.”

Although clinicians may feel anxious about AI evaluating their clinical skills, the technology is being designed to give feedback the same way good supervisors do. “It should be digestible, calibrated, and delivered in a way that promotes growth rather than overwhelm,” Esherick said.

Krista Kaur, a clinical mental health counseling graduate student at Palo Alto University in California, interned at a clinic that uses an AI tool to record sessions with patients, create a redacted transcript, and produce session notes. “It felt like a personal assistant that was helping me grow as a clinician,” she said. She could click a link to get feedback from the AI tool about things she did well in the session and things to consider for next time. “Before going into a session, it would give me a few reminders about where we had left off last time, which was really helpful,” Kaur said. Occasionally the AI would record and transcribe words incorrectly from a session; Kaur would disregard those aspects of the case summary.

Psychologist Philip Held, PhD, an associate professor in the Department of Psychiatry and Behavioral Sciences at Rush University Medical Center in Chicago, recently developed an AI tool called Socrates Coachopens in new window that helps therapists practice Socratic dialogue skills—open-ended questions that encourage patients to explore and challenge beliefs associated with distress and maladaptive thoughts. The tool gave therapists feedback about their overall strengths and areas for growth with this skill, a part of many cognitive therapy approaches that is closely linked to outcomes but difficult to master (Vittorio, L. N., et al., Behaviour Research and Therapy, Vol. 150, 2022opens in new window).

Held originally developed an AI chatbot that patients could use between therapy sessions to engage in Socratic dialogue instead of completing standard therapy worksheets. Clinical colleagues started asking if he could adapt the tool to help them develop and refine their Socratic dialogue skills.

His team built prototypes for Socrates Coach, and practitioners rated the quality of the responses from the AI coach. Based on the human ratings, the team used prompt optimization—refining the input prompts to generate more accurate responses from large language models (LLMs)—to improve the feedback from the AI supervisors (Journal of Traumatic Stress, Vol. 38, No. 5, 2025opens in new window). The AI feedback includes alternative questions clinicians could ask the AI patient, possible directions for the conversation, and information about how well they are using the technique.

Even though AI trainers can be scaled quickly, psychologists caution that this technology cannot assess whether a clinician is ready to apply these skills to human patients. “These tools are an additional way to practice, but they are not a replacement for human supervisors,” said Held. Human therapists should be helping trainees with therapy decisions that depend on context, such as complex family dynamics, subtle risk signals that suggest abuse or suicidality, and other types of clinical intuition based on years of expertise.

AI tools may also miss details about patients that are communicated nonverbally. “Therapists are seeing facial affect, body language, and other nuances that may influence how a therapist talks to a patient or the treatment recommendations,” said Eric Kuhn, PhD, an associate professor at the Stanford University School of Medicine and a member of APA’s Mental Health Technology Advisory Committee.

Ideally, these tools can provide valuable insights that supervisors can read and discuss with trainees. Donna Sheperis, PhD, a professor in the Department of Counseling at Palo Alto University, worked with one company that gave her access to a tool that recorded therapy sessions—with permission from patients—and provided content analysis of the sessions to help her track her trainees’ progress. The analysis showed which interventions were used, what was discussed, how much time each person spoke, and more. In one case, she noticed that the trainee was talking nonstop for 4 minutes, which is generally considered to be too long for most patients to listen, she said. When Sheperis discussed this observation with the trainee during supervision, the student explained that she was saving up her comments to share with the patient at one time because she did not want to interrupt. “We talked about the difference between interrupting and interjecting,” said Sheperis, who directs the eClinic at Palo Alto University.

Creed, founder and director of the Penn Collaborative for CBT and Implementation Science, started looking into AI after she spent years training providers in underfunded community mental health systems around the world. “The human hours needed to train clinicians and score their competence in new skills at scale were impossible,” said Creed. “I knew there had to be a better way.”

Creed was eager to explore whether Lyssn—an AI platform developed by psychologists and computer scientists in 2017 to analyze therapy sessions and give feedback to therapists and trainees—could support community mental health agencies (BMC Health Services Research, Vol. 22, No. 1177, 2022opens in new window). Using LLMs trained with peer-reviewed research publications and expert-evaluated behavioral health conversations, the platform provides training for motivational interviewing, CBT, suicide prevention, and more. The tool gives clinicians feedback on whether they are demonstrating the skill and points to areas of strength and improvement. Supervisors can also see aggregated data on their teams, which can inform decisions about targeted training that may be needed.

Creed recently finished collecting data for a 5-year study funded by the NIMH to investigate how Lyssn training affects the skill level of clinicians and patient outcomes. “I had one clinician who was near retirement say that the AI feedback has shifted how she does her clinical work,” said Creed. “If she had this support all along, the care she would have delivered would have looked really different.”

Psychologists responsible for supervision should also be prepared to talk with trainees about how they may be using general-purpose AI tools such as ChatGPT, Claude, and Gemini. “Being fearful of AI or prohibiting it is a reactive way to deal with something that is very much a part of our lives,” said Levenson. “We need to ask trainees if they are using tools like ChatGPT, and the worst thing we could say is ‘Just don’t use it.’” She has little doubt that many trainees are presently using AI for help with their clinical cases but are afraid to disclose this to their supervisors for fear of offending them or being criticized. “It is really up to the supervisor to inquire about a trainee’s clinical use of AI in a way that opens up opportunities for discussion and learning.”

If used responsibly, these tools can be useful brainstorming partners to help therapists with case conceptualization, treatment planning, and identifying strategies to overcome challenges with patients. Supervisors can start by asking trainees how they are using AI, what they are learning from it, what is working well, and if they have concerns. “It’s also important to help supervisees recognize when they are getting good feedback from AI and when errors are being made,” said Levenson.

Supervisors can help guide trainees to identify AI-related red flags, such as bias, confidentiality risks, and hallucinations—when an AI tool generates output that sounds authoritative or plausible but is factually wrong. To help students learn to safely use publicly accessible AI, Tony Rousmaniere, PsyD, program director of the Sentio University Marriage and Family Therapy Program in California, started offering online courses for trainees, licensed therapists, and clinical supervisors on this topic.

Rousmaniere teaches students to protect patient confidentiality by avoiding identifying information such as name, age, gender, and job details. In cases when a clinical question requires identifying information to answer effectively, “that question does not belong in an AI tool,” said Rousmaniere. “It belongs with a supervisor. AI is for brainstorming around a case, not for getting clinical answers that hinge on who the specific person is.”

He teaches students how to provide AI with useful demographic framing that does not identify someone. For example, a client can be described as a white middle-aged male who works in the technology industry—a profile that fits thousands of people. A trainee might ask the AI to brainstorm workplace stressors common in the technology industry that could contribute to a midcareer depression presentation, or CBT homework variants that tend to work well with analytical and high-control occupational identities.

Rousmaniere also encourages trainees to double-check the treatment protocols AI is suggesting to ensure that the ideas are valid and a good fit for the patient. AI may also suggest a therapy protocol that the trainee is not competent to use, he said, so trainees should perform a self-assessment to determine if they can provide a treatment.

“We have found that roughly 90% of the guidance from general-purpose AI is accurate and helpful,” said Rousmaniere. “Our focus is to help trainees identify the 10% that is not accurate.”

Supervisors, teachers, and trainers in the field of mental health can be most effective in helping trainees navigate these tools by becoming AI literate themselves, added Levenson. Instructors can experiment with using AI to practice interacting with different mock patients and ask for feedback from the AI. This, in turn, will equip them to guide supervisees as they use the technology to help them master new therapy skills.

By exploring how to use AI in clinical training, psychologists can talk to trainees about both benefits and risks of this technology. “There are tremendous challenges in terms of human resources to scale up a mental health workforce to effectively use evidence-based treatments,” said Kuhn. “AI tools are offering a safe environment for unlimited practice for skill-building, and psychologists have to be involved in creating these products to ensure they adhere to approaches that result in the best patient outcomes.”