This article contains affiliate links, meaning Modern MedEd may earn a commission at no cost to you should you decide to make a purchase through such a link. This helps support our free content. Thank you!
AI in Mental Health: LLMs and Therapy
Show Infographic
- Key Points
- LLMs are effectively acting as major mental health providers right now. They are attractive because they are free, available 24/7, and offer immediate responses without insurance hurdles.
- Usage is highest among young adults, males, and Black users, especially those belonging to groups that often face barriers to accessing traditional mental healthcare.
- Studies suggest purpose-built chatbots can reduce short-term depression and anxiety symptoms, but long-term safety data is needed.
- General-purpose LLMs can be dangerous when used for therapy because they may confabulate medical facts, reinforce delusions (AI psychosis), and lack protocols for handling active crises.
History of Chatbots in Psychiatry
Like it or not, your patients are using ChatGPT for therapy. And while the trend may feel like a 2025 thing, it really isn’t. The first Conversational User Interface (CUI) (ie, chatbot) for mental health was called ELIZA, and launched in the 1960s. What is new is the scale, sophistication, and 24/7 accessibility of today’s tools.
There are contemporary conversational bots that were built for mental health (Woebot, Wysa, and others), that even have a modicum of evidence supporting their use. These tend to be rule-based or ‘scripted’ systems (like ELIZA), which differ substantially from generative systems like large language models (LLMs).
Some studies (predating 2022, the year ChatGPT was released, followed by its many competitors) were able to show a reduction in depression and anxiety symptoms with conversational bots compared to passive controls. These findings established early proof of concept for digital conversational support, but fell short in anticipating the capabilities and risks of generative systems.
General-purpose LLMs are great at displaying conversational ability and producing what could reasonably pass as human-sounding dialogue. They are incredibly affordable, easily accessible to anyone with an internet connection, and available any time of day or night, all without a single copay charge.
Who Uses AI for Therapy?
A survey from Sentio University, a Marriage and Family Therapy training organization, asserts that LLMs like ChatGPT, Claude, or Gemini, may now be the largest mental health ‘providers’ in the United States. This statement is by no means definitive, and requires a lot of “if/then” conditions to be true. However, it is a claim that we cannot ignore.
Data from early 2025 suggest over half (52%) of US adults use AI LLMs on a regular basis. A Stanford survey of US adults reported that 24% of respondents used LLMs for mental health reasons. People who use LLMs for psychotherapy tend to be young, male, black, report poorer mental health at baseline, and have difficulty accessing traditional mental health treatment because of cost and insurance coverage issues.
Despite their popularity, important safety risks persist. Ethical and legal issues, privacy concerns, hallucinations, misinformation, inconsistent crisis management, dangerous levels of sycophancy potentially contributing to AI psychosis, and dubious accountability are just a few issues that need to be mitigated before clinicians can confidently recommend AI-assisted psychotherapy. For now, it’s good to talk to your patients about their use of LLMs for psychotherapy, but it’s dicey to recommend it for therapy today.
A Review of the Scientific Literature
An unblinded randomized controlled trial from 2017 showed that conversational agents (CA) could be a feasible, engaging, and effective way to deliver CBT. Seventy young adults were randomized to receive either two weeks of self-help CBT from the now defunct conversational agent (pre-scripted, rule-based) Woebot or directed to a National Institute of Mental Health (NIMH) educational e-book on depression.
Scores on PHQ-9, GAD-7, and Positive and Negative Affect Scale (PANAS) improved significantly in the chatbot group, with an effect size of 0.44. For comparison, SSRIs typically have an effect size of around 0.3 to 0.5 for depression and anxiety, respectively. Users were generally satisfied with the intervention.
Limitations include the short duration of the intervention, small sample size, and lack of follow up data evaluating sustained improvements. The second author on the study was the founder of the technology being studied, and the company at least partially funded the study.
A 2023 systematic review and meta-analysis examining generative AI based CAs concluded that these tools could significantly reduce short-term depressive symptoms. However, they did not improve broader psychological well-being, and safety data was severely lacking. The authors correctly identify risks such as privacy infringement, bias, and safety issues despite the very real potential for improving access and mental health symptoms.
There are also encouraging findings. A 2024 study of Spanish-speaking adolescents reported that a domain-specific LLM improved early recognition of mental health symptoms and appeared to motivate help-seeking behavior.
Overall, these tools appear to score well on user satisfaction ratings and provide some relief from mild psychological symptoms. The duration and magnitude of the effect, along with the safety profile, especially in populations with severe mental illness, have yet to be established.
Risks of AI Therapy
As we touched on above, using LLMs for therapy isn’t harmless. General purpose LLMs are not built for nor intended for this use case. Risks include hallucinations (of the AI), AI psychosis (of the patient), reinforcement of maladaptive beliefs and thought patterns, data privacy and regulatory mismatch, health equity, bias, and other ethical issues.
Hallucinations
LLMs give confident answers even when blatantly incorrect. This might involve false medical claims, misinterpretation of symptoms, and fabricated evidence. These things were built to maximize engagement rather than factual accuracy after all.
Over-Validation and Reinforcement of Maladaptive Beliefs
Whenever I have an idea and talk to ChatGPT or Claude (or Gemini, or Deepseek, etc) about it, I feel like the smartest person in the room (probably because I’m the only person in the room). The point is, I’ve never been told I had a bad idea unless I specifically ask for criticism (and even then, it’s delivered very diplomatically). Providing excessive reassurance to someone with OCD or playing into the paranoia or delusions of a person with psychosis is far too easy a trap for LLMs to fall into.
AI Psychosis & Crisis Management
AI psychosis describes situations in which heavy engagement with LLMs intensifies or contributes to delusional thinking. There have been reports of vulnerable people being pushed into full-on delusions and psychosis from intense conversations with LLMs. Crises like these can be challenging even for clinicians to manage; for a bot with no way to collect collateral or contact emergency services, is asking for trouble.
Data Privacy (or lack thereof)
Commercial LLMs are not configured for nor covered under HIPAA. Users may share sensitive health or personal information that could be stored, logged, or accessed in ways inconsistent with medical privacy standards (completely legally, too). This presents a meaningful risk, including future discoverability of chat transcripts. That’s not paranoia – even ChatGPT says so.
Ethical and Legal Issues
See ‘Data Privacy’ above. Since your chats are, in a legal sense, discoverable, and not protected health information, it’s not hard to imagine the myriad ways this could go wrong. Statements made impulsively during emotional distress could be retrieved in legal, employment, or insurance contexts, with unpredictable consequences.
Equity and Bias
Models trained on text from the web may underperform or produce biased outputs for minoritized populations. This was seen in a 2024 study evaluating LLMs as a clinical decision support (CDS) tool for bipolar disorder. The authors reported that the LLM performed worse appraising vignettes where the patient was a black woman compared to a white man.
Modern MedEd Takeaway
Talk to your patients about how they use LLMs in a mental health context. Provide appropriate caution, but try not to sound like a luddite. Most people like their AI therapy chatbots (they may even help some people), and they aren’t about to give them up anytime soon. That said, make sure they (and you) have at least a basic understanding of these systems and what appropriate use looks like.
Additional Citations
Al-Shamsi, B., Al-Ghaithi, A., & Al-Adawi, S. (2024). Cultural bias in large language models for bipolar disorder: A vignette-based evaluation study. Translational Psychiatry, 14, 148. https://www.nature.com/articles/s41386-024-01841-2
Elon University. (2025, March 12). Survey: 52% of U.S. Adults Now Use AI Large Language Models Like ChatGPT. https://www.elon.edu/u/news/2025/03/12/survey-52-of-u-s-adults-now-use-ai-large-language-models-like-chatgpt
Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19. https://pmc.ncbi.nlm.nih.gov/articles/PMC5478797/
Glicksman, A. S. (2025). AI psychosis is real: Case studies show risk. Psychiatric News, 60(10), 5. https://psychiatryonline.org/doi/10.1176/appi.pn.2025.10.10.5
Guo, Z., Lai, A., Thygesen, J. H., Farrington, J., Keen, T., & Li, K. (2024). Large language models for mental health applications: Systematic review. JMIR Mental Health, 11(1), e57400. https://mental.jmir.org/2024/1/e57400
Miners, A., Coyle, D., & O'Hanlon, C. (2023). Generative artificial intelligence-based conversational agents for mental health: A systematic review and meta-analysis of randomized controlled trials. npj Digital Medicine, 6, 223. https://www.nature.com/articles/s41746-023-00979-5
OpenAI. (2024, July). Law enforcement response policy (Version 2024.07). https://cdn.openai.com/trust-and-transparency/openai-law-enforcement-policy-v2024.07.pdf
Sentio University. (n.d.). The New Largest Mental Health Provider in the US: AI Large Language Models. https://sentio.org/ai-research/ai-survey
Vianna, L. A., Zaid, T., & Miller, J. (2024). The mental health use of AI large language models among US adults. OSF Preprints. https://sciety-labs.elifesciences.org/articles/by?article_doi=10.31219/osf.io/ygx5q_v1
Vílchez-Conesa, P., Almagro-Ríos, P., & López-Sánchez, C. (2024). Impact of a domain-specific large language model on mental health recognition and help-seeking behavior in Spanish-speaking adolescents: A pilot study. Computers in Human Behavior, 158, 108493. https://doi.org/10.1080/10447318.2024.2344355
Wang, L., Bhanushali, T., Huang, Z., Yang, J., Badami, S., & Hightow-Weidman, L. (2025). Evaluating generative AI in mental health: Systematic review of capabilities and limitations. JMIR Mental Health, 12(1), e70014. https://doi.org/10.2196/70014
Weizenbaum, J. (1966). ELIZA—A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36–45. https://dl.acm.org/doi/10.1145/365153.365168
First Published: December 28, 2025
Further Reading
5 Reasons You Never Qualify for Paid Medical Market Research Surveys
Like many clinicians, you may have tried joining paid medical market research survey panels for extra income. Similarly, if you are frustrated because you never seem to pass the screener, allow me to shed some light on why.
Patients Recording Clinical Encounters
While there are conflicting viewpoints on the issue of patients recording of clinical encounters, today’s article will explore the topic from a medicolegal liability and risk management perspective.
