Using AI for Therapy: What the Research Shows
Clinical trials show that some AI chatbots reduce depression and anxiety symptoms, but the results do not establish AI as a replacement for a licensed therapist. The tool, the people using it, and the safety measures all affect what a study proves. Using AI for therapy also brings risks, including harmful advice, excessive agreement, and privacy concerns. This guide explains the findings and their practical limits.
Key findings at a glance:
In an online sample of US adults, 35.2% used AI for mental health support at least weekly. (JMIR Mental Health, 2026).
Therabot users with depression reported an average 51% reduction in symptoms in a monitored trial. (Dartmouth, 2025).
A review of 39 trials found small average improvements in depression and anxiety. (npj Digital Medicine, 2026).
Licensed therapists gave appropriate responses 93% of the time in one safety study, outperforming the tested bots. (University of Minnesota research release, 2025).
The FDA reported zero authorized AI-enabled devices for mental health uses in its November meeting materials.
What Does “Using AI for Therapy” Mean?
Using AI for therapy means turning to a chatbot for help with thoughts, feelings, relationships, or mental health symptoms. That includes asking for coping exercises, discussing a difficult day, and seeking advice about anxiety or depression.
The phrase covers different products. Some follow planned exercises. Others create new replies using a large language model, or LLM, a system trained to generate language from patterns in data. Their evidence and safeguards differ.
General-Purpose Chatbots
General-purpose tools such as ChatGPT, Claude, and Gemini handle many tasks, including writing, explanations, and conversation. Asking it to act as a therapist changes its instructions. It does not create a professional license, a clinical assessment, or a system of patient care. A fluent answer also gives no assurance that the advice fits your history or symptoms.
Purpose-Built Mental Health Chatbots
Purpose-built tools include Therabot and structured programs such as Wysa. They focus on activities such as mood tracking, guided reflection, or coping skills. Some have clinical trials, and the evidence belongs to the exact program tested. Some use cognitive behavioral therapy, or CBT, which works on connections between thoughts, feelings, and behavior. Look for research on the actual product. A reference to CBT alone does not prove that a chatbot delivers it well.
AI Companion Apps
Character-based companion apps focus on conversation and a sense of ongoing connection. Keep that purpose separate from treatment. An app that remembers details and responds warmly has not, through those features alone, shown that it treats a mental health condition or handles a crisis safely.
How Many People Use AI for Therapy?
Surveys show that people already use chatbots for emotional support. They do not provide one reliable percentage for everyone using “AI therapy.” Researchers ask different questions, study different age groups, and measure different periods. A survey about asking for advice is also different from a trial that measures changes in symptoms.
Adults:
In an October 2025 adult survey of AI-based mental health support, 35.2% of 1,805 respondents ages 18–49 reported using AI for mental health support at least weekly. Another finding was that 5.5% met the researchers’ definition of heavy use. The study was published in JMIR Mental Health in 2026.
The researchers used a commercial online panel, so these findings are not a nationally representative estimate of all US adults. The survey measured reported behavior at one point in time. It does not show that AI use caused better or worse mental health. (Ueda et al., 2026).
The APA’s June 2026 survey adds a view from clinical practice: 33% of surveyed US psychologists providing patient care reported patients using AI to assist therapy or treatment. That percentage describes psychologists’ reports, rather than the share of all patients using AI.
Teens and Young Adults:
The RAND findings on youth chatbot use show an increase across two surveys of US adolescents and young adults ages 12–21:
13.1% reported using AI for mental health advice in the February–March 2025 survey, reported in 2025.
19.2% reported such use in the November 2025 survey, reported in 2026.
63% of users in the later survey had told no one about using AI for this purpose.
92% of users in that later survey considered the advice somewhat or very helpful.
These are separate surveys, rather than conflicting estimates from the same survey. Perceived helpfulness measures satisfaction, not clinical safety or recovery.
Why People Turn to AI for Support?
Common attractions include:
Immediate access without an appointment.
Free or inexpensive conversation.
Less worry about judgment or embarrassment.
A private-feeling place to put difficult feelings into words.
Access during gaps in professional care.
These explain the appeal. They do not establish the quality of care.
Does AI Therapy Work? What the Clinical Trials Show
Some chatbot programs have improved symptoms in clinical trials. The strongest conclusion is specific: certain tools helped certain groups under the conditions tested.
A randomized controlled trial, or RCT, assigns people by chance to an intervention or comparison group. A meta-analysis combines results from several studies. An effect size puts differences on a shared scale. Hedges’ g and Cohen’s d are effect-size measures, not percentages of people who recovered.
Evidence Scorecard:
Study | Design | Sample | Main finding | Key limitation |
Therabot versus a waiting list | 210 adults | Improved depression, anxiety, and eating-disorder risk measures | Clinician monitoring; no human therapy comparison arm | |
Meta-analysis | 18 RCTs; 3,477 participants | Small depression and anxiety effects, g = −0.26 and −0.19 | No significant effects at three-month follow-up | |
Sohn et al., npj Digital Medicine, 2026 | Meta-analysis | 39 eligible trials; outcome-specific samples | Depression g = 0.31; anxiety g = 0.28 | Mixed chatbot types; most trials had high risk of bias |
Feng et al., JMIR, 2025 | Review of youth trials | 15 trials | Depression improved; adjusted anxiety and stress findings were nonsignificant | Different tools and study methods; limited follow-up |
Kuta et al., JMIR Mental Health, 2026 | Structured bot versus ChatGPT versus control | 147 eligible participants completed baseline assessment | Both chatbot groups improved on one depression measure | Three weeks; substantial dropout; other outcomes nonsignificant |
AI platform versus group therapy versus waiting list | 995 university students | Better anxiety and well-being results than both comparators | Selected student sample; self-reports; industry ties disclosed | |
Ratings of written therapy responses | 830 raters; responses from 13 therapists | ChatGPT responses received higher ratings on common therapy factors | Response ratings, rather than treatment outcomes |
The signs on effect sizes depend on how researchers coded improvement. A negative value in one paper and a positive value in another do not automatically mean opposite findings.
The Therabot Trial:
51%, 31%, and 19% were the average reductions reported for depression symptoms, anxiety symptoms, and body image and weight concerns among relevant Therabot users. These were changes in symptom scores, not cure rates. (Dartmouth, 2025).
The study tested a purpose-built system with expert training and clinician monitoring. Participants received four weeks of access, with assessments through eight weeks. Its results support further study of that system. They do not establish the same outcomes for ordinary chatbot conversations.
What Meta-Analyses Find?
0.31 for depression and 0.28 for anxiety were the pooled effect sizes in the 2026 review. Both represent small average benefits. Its depression analysis included 7,401 participants; its anxiety analysis included 7,621. These are overlapping outcome groups, not separate totals to add together.
General Chatbots Versus Purpose-Built Tools
Three weeks was the intervention period in the direct comparison of a structured chatbot and ChatGPT. Both improved PHQ-9 depression scores against the control group, with effect sizes of −0.47 and −0.44. Neither significantly outperformed the other. This small feasibility study does not establish equal effectiveness or dependable anxiety treatment.
Do the Benefits Last?
Three-month findings differ across studies. The 2024 review found no significant follow-up benefits. A newer student trial reported persistence of several symptom-score benefits, but substantial follow-up dropout limited confidence. Long-term recovery across products and patient groups remains unestablished.
Can AI Match a Human Therapist?
AI performs well on some conversation ratings and selected outcomes. Those findings do not establish that it matches the full work of a licensed therapist.
The practical differences involve access, responsibility, and the quality of care:
Availability: Chatbots offer on-demand conversation. Therapists work through scheduled sessions and agreed support arrangements.
Cost: Chatbots have free and paid plans. Therapy costs depend on the provider, insurance, and service.
Accountability: A licensed therapist has professional and legal duties. A chatbot is not a licensed clinician.
Emergency support: Therapists follow risk-assessment and emergency procedures. An automated crisis message is different from clinical intervention.
In one 2025 safety evaluation, licensed therapists responded appropriately 93% of the time, compared with under 60% for the tested therapy bots. These figures describe that evaluation, rather than every product or therapist.
Where AI Scores Well?
In the PLOS study, people rated AI-written responses highly for qualities associated with therapy. The researchers selected responses for the comparison, and participants judged written scenarios rather than receiving ongoing treatment. The result shows that AI produces convincing supportive language. It does not show that readers received effective therapy.
The newer student trial also provides a real human-care comparison. Its results deserve attention, but one group-therapy format does not represent every type of psychotherapy.
Where AI Falls Short?
Safe care requires more than an answer that sounds kind. The Stanford-led research documented inappropriate replies to suicidal intent and delusional beliefs. Its findings concern the systems and situations tested in 2025. They establish specific failures without assigning the same performance to every later model.
What a Therapeutic Relationship Provides
Work with a licensed therapist includes:
Assessment of symptoms, history, and day-to-day functioning.
Shared goals and a treatment plan.
Discussion of patterns that reassurance alone does not change.
Attention to tone, behavior, and other clinical information.
Professional responsibility for risk assessment and care decisions.
An AI conversation does not establish that professional relationship.
What Are the Risks of Using AI as a Therapist?
The risks involve the advice itself, the relationship a person develops with the tool, and the handling of sensitive information. Positive trial results do not remove these concerns. A tool also needs reliable safeguards outside the narrow topics and time periods covered by a study.
Unsafe Responses in a Crisis:
Under 60% was the appropriate-response rate reported for therapy bots in the 2025 safety evaluation. In one test, bots answered a question signaling suicidal intent with factual information rather than appropriate support. A chatbot’s friendly tone therefore provides no assurance of safe crisis handling.
During a mental health emergency, seek human support directly. Do not depend on the chatbot to recognize the situation, assess your safety, or arrange help.
Agreeing Too Much: Sycophancy and Delusions
15 ethical risks were identified in Brown University’s research on chatbot ethics in 2025. These included poor crisis handling and excessive validation. Giving a chatbot instructions to use therapy methods did not prevent the failures.
Sycophancy means agreeing to please the user. In mental health conversations, the distinction matters: acknowledging fear is different from confirming that a feared event is real. Support should not turn an unsupported belief into an accepted fact.
Stigma and Bias
Five therapy chatbots were examined in the Stanford-led study of chatbot stigma and safety. The researchers found greater stigma toward schizophrenia and alcohol dependence than toward depression. Larger and newer models in the tests did not consistently solve this problem.
That contradicts the idea that a chatbot is automatically free of judgment. A response that shames someone or stereotypes a diagnosis is a reason to reject that advice, rather than accept its judgment about the person.
Loneliness and Emotional Dependence
981 adults participated in a four-week study of extended chatbot use by MIT and OpenAI researchers. Heavier voluntary use was associated with worse psychosocial outcomes, including dependence and problematic use. The randomized conversation conditions did not show significant effects. (Fang et al., 2025 preprint).
This is an association, not proof that extra use caused the problems. For daily use, watch what the chat replaces: sleep, contact with friends, responsibilities, and professional appointments all matter more than time spent receiving reassurance.
Privacy: Where Your Conversations Go
HIPAA does not automatically cover personal consumer apps. The Health Insurance Portability and Accountability Act protects information handled by covered health organizations and their business associates. Entering health details into an unrelated app does not give those details the same protection.
Before sharing personal information:
Read the policy on storage, deletion, and model training.
Check whether staff or service providers review chats.
Remove names, addresses, and identifying details.
Keep medical records and other people’s private information out of general chat tools.
Who Should Not Rely on AI for Mental Health Support?
People in crisis, people with severe symptoms, and children need particular care. General chatbot use does not provide the assessment and protection required in these situations. The following boundaries are practical guidance, rather than a claim that a diagnosis makes every digital tool unsuitable.
People in Crisis or Having Suicidal Thoughts
Contact a crisis counselor, emergency service, or qualified clinician directly.
Tell a trusted person what is happening and ask for support.
Do not treat an automated safety message as a completed risk assessment.
Seek urgent medical help for immediate danger or a medical emergency.
The tested crisis-response failures explain why a chatbot should not be the only source of help.
People With Psychosis, Mania, or Eating Disorders
Use qualified professional care for assessment and treatment decisions.
Do not ask a chatbot to confirm unusual beliefs or interpret them as established facts.
Do not use chatbot-generated food rules, weight targets, or medication changes as a treatment plan.
Discuss worsening symptoms and AI conversations with the treating clinician.
Research on a monitored tool for eating-disorder risk does not establish safe treatment for every eating disorder.
Children and Teens
Keep trusted adults and qualified professionals involved in mental health care.
Ask about chatbot use without punishment or embarrassment.
Explain that a convincing reply is not a diagnosis or proof of safety.
Pay attention to secretive use that accompanies withdrawal or worsening distress.
Open discussion matters because many young users keep this use to themselves.
Is AI Therapy Legal? How It Is Regulated
There is no single legal rule covering every AI mental health product across the United States. State laws address different activities, including professional treatment, advertising, disclosure, and companion safety. The selected measures below were checked on November 6, 2026; they are not a complete list of state requirements
.
State Laws
Illinois: The state’s restrictions on AI use in professional therapy took effect on August 1, 2025. Therapy must be conducted by licensed professionals. The law restricts independent AI therapeutic decisions and direct AI therapeutic communication.
Nevada: AB 406 restricts professional AI mental health care, including systems marketed or programmed to provide professional mental or behavioral care. Approved on June 5, 2025, the law distinguishes self-help products from professional treatment.
Utah: HB 452 established mental health chatbot disclosure and information protections in 2025. These cover the chatbot’s AI identity and the handling of personal information.
New York: The original companion-AI safeguard requirements took effect on November 5, 2025. Covered operators must provide safety protocols addressing suicidal thoughts and self-harm, along with required notices.
These different approaches do not amount to a nationwide ban on talking to AI about feelings.
FDA Status:
FDA authorization belongs to a specific device and intended use. An app-store listing, research paper, or claim to use CBT does not establish it. In November 2025, the FDA reported no AI-enabled devices authorized for mental health uses. Its August 2026 discussion paper on generative AI medical-device regulation considers evaluation and monitoring approaches. The paper requests input; it does not provide product authorization or final guidance.
Professional Guidance
The APA advises against relying on general-purpose chatbots and wellness apps as replacements for qualified care. Its 2026 guidance on chatbots and mental health supports limited tasks such as preparing appointment questions and practicing homework already discussed in treatment.
How to Use AI for Mental Health Support More Safely
These boundaries reduce avoidable risks. They do not turn a general chatbot into a clinically validated treatment.
Choose a narrow task before opening the chat. Ask for journaling questions, help organizing notes, or a plain-language explanation. Keep diagnosis and treatment decisions with a qualified clinician.
Check evidence on the exact tool. Look for a published study, its comparison group, the people studied, and its safety procedures. Treat a therapy-related name as marketing until the evidence supports the claim.
Ask for alternative views. Request other explanations and questions worth discussing with a therapist. An instruction to disagree does not guarantee safe or accurate advice.
Limit sensitive information. Remove names and identifying details. Read the data policy before entering health information, and check the available privacy controls.
Check health claims with a reliable source. Use a qualified clinician or an established health organization. Open cited links and confirm that the source supports the answer.
Protect time for real relationships. Set a clear stopping point. Keep appointments, sleep routines, and contact with people who know you.
Stop using the chat when distress increases. Bring the concern to a trusted person or clinician. Suicidal thoughts, severe confusion, or immediate danger require direct human support.
Lower-Risk Ways to Use AI
Generate questions for a journal entry.
Turn your own notes into a list for a therapy appointment.
Prepare questions about a therapist’s training and approach.
Practice wording for an everyday conversation.
Review homework assigned by your therapist without changing its instructions.
These uses keep the task limited and leave care decisions with a person.
Red Flags That Mean Stop and Get Human Help
The bot supports self-harm, restrictive eating, or other harmful behavior.
It treats an unsupported frightening belief as proven fact.
It tells you to abandon care or ignore a clinician.
Chatting replaces sleep, responsibilities, or human contact.
You feel increasingly dependent on its reassurance.
Your symptoms worsen despite feeling briefly comforted by the replies.
What the Research Does Not Tell Us Yet
Long-term recovery: Short-term score changes do not establish lasting recovery or protection against relapse.
Replacement of individual therapy: Results from a waiting-list comparison or one group-therapy program do not settle this question.
Safety across products: Evidence from one monitored tool does not establish safe use of every consumer chatbot.
Outcomes for excluded groups: Trials that exclude severe illness and active crisis do not establish effectiveness for those groups.
Performance after updates: Findings about a tested model version do not automatically transfer to a changed product.
Safety reporting also needs improvement. In the 2026 review, 23 of 39 studies did not report systematic safety monitoring or adverse-event data. Missing reports do not prove that no harm occurred.
Talk to a Licensed Therapist:
For support based on your history, symptoms, and goals, talk with Dr. Lisa Palmer at The Renew Center of Florida about therapy with a licensed therapist. Start with a complimentary discovery call to discuss the care you are seeking.
For urgent support in the United States: Call or text 988. For immediate danger or a medical emergency, contact emergency services.
Frequently Asked Questions:
Can I Use ChatGPT as a Therapist?
ChatGPT is not a licensed therapist, and using it for emotional support does not create a professional care relationship. Limited research has tested general chatbot conversations, including a short trial with improved depression scores. That result does not establish dependable diagnosis, crisis care, or long-term treatment. Use it for bounded tasks such as preparing appointment questions, and keep clinical decisions with a qualified professional.
Is It Safe to Use AI for Therapy?
Research has yet to establish that AI chatbots can safely replace therapy. Tests have documented harmful responses, and safeguards vary across products. How you use a chatbot also matters. Organizing your notes is a lower-risk task than asking for advice during a crisis, when even a supportive response can reflect poor clinical judgment. If you use a consumer chatbot, stick to simple tasks and avoid sharing sensitive personal information. If symptoms are affecting your daily life, seek an assessment from a qualified mental health professional.
Is AI Therapy Effective for Anxiety and Depression?
Some chatbot interventions have reduced depression and anxiety scores in trials. Average benefits are generally small. The evidence combines different tools and comparison groups, and symptom-score changes do not establish full recovery. Look for evidence on the specific product and condition. Claims that every AI chatbot treats anxiety or depression go beyond these results.
Will AI Replace Human Therapists?
Current evidence does not establish that AI replaces the full role of a human therapist. Some AI tools have performed well on selected outcomes, including in a comparison with group therapy. Those results support research into defined uses. They do not settle whether a chatbot provides equivalent individual care across conditions, ages, and levels of risk.
Is AI Therapy Legal?
The rules depend on the state and the service being offered. Professional treatment, general self-help, and companionship are different legal categories. Illinois restricts AI use in professional therapy, and Nevada restricts systems marketed or programmed to deliver professional mental or behavioral care. A product being available online does not establish compliance. Check the rules governing the service in your state rather than assuming one national rule covers every tool.
Is AI Therapy FDA Approved?
“AI therapy” is not a single FDA-approved treatment category. Authorization must be checked for the exact product and intended use. The FDA’s November 2025 materials reported no AI-enabled devices authorized for mental health uses. Its August 2026 discussion paper explores regulatory questions and explicitly states that it is not draft or final guidance. Clinical research, a wellness label, and availability for download are not substitutes for a product-specific authorization.
Is It Bad to Vent to an AI Chatbot?
Venting to a chatbot does not, by itself, establish harm. Pay attention to the content of its advice and the effect on everyday life. A conversation that helps you organize feelings is different from repeated reassurance that replaces relationships or treatment. Track whether you leave with useful questions and next steps, or return for the same reassurance while withdrawing from people. Heavier use and dependence were associated in one study, without proving cause.
Are My Conversations With an AI Chatbot Private?
Chatbot conversations do not automatically receive the confidentiality protections of a therapy session. Consumer-app privacy depends on the provider, its policies, applicable laws, and its relationship to health care organizations. Read the rules for retention, human review, deletion, and training. Avoid sharing identifying health records in a general chatbot. A chat window that feels private is not proof that only you have access to its contents.
What Should I Do After a Chatbot Gives Harmful Advice?
Stop following the advice and discuss it with a qualified professional or trusted person. Do not let a chatbot response override a clinician’s treatment instructions. Keep a copy of the concerning reply for discussion and report it through the platform’s safety tools. Seek urgent human support for advice involving self-harm, severe confusion, or immediate danger. Continuing the same conversation is not a dependable way to check whether the original advice was safe.
Can AI Help Me Find a Real Therapist?
AI can help prepare your search and the questions you want to ask. Use it to organize preferences such as location, appointment format, cost, and experience with a particular concern. Verify names, licenses, availability, and fees directly through the provider and the relevant licensing board. Treat generated recommendations as leads to check rather than verified referrals. Ask prospective therapists about their training, treatment approach, and experience with the support you need.
Additional Sources
The following documents support the article alongside the sources linked in the text. Sources from websites already linked above are listed here without repeating those website links.
APA, 2025. Health advisory: Use of generative AI chatbots and wellness applications for mental health. Supports the distinction between general chatbots, wellness tools, and professional care.
Feng et al., 2025. Effectiveness of AI-Driven Conversational Agents in Improving Mental Health Among Young People: Systematic Review and Meta-Analysis. Journal of Medical Internet Research, 27, e69639. DOI: 10.2196/69639. Supports the youth evidence-scorecard findings.
Kuta et al., 2026. Effectiveness of a Fully Automated Mobile Therapeutic Versus a General Chatbot in Reducing Depression and Anxiety and Improving Well-Being: Feasibility Randomized Controlled Trial. JMIR Mental Health, 13, e82642. DOI: 10.2196/82642. Supports the three-week structured-chatbot versus ChatGPT comparison.
RAND, 2025. Use of Generative AI for Mental Health Advice Among US Adolescents and Young Adults. Research published in JAMA Network Open. Supports the earlier youth survey figure and survey period.
Moore et al., 2025. Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency, pages 599–627. DOI: 10.1145/3715275.3732039. Primary paper underlying the Stanford and research-team safety summaries linked above.
