Can AI Help with Mental Health? What the Research Actually Shows

The question is not whether AI can help with mental health. A 2025 randomized controlled trial published in NEJM AI answered that: yes, a clinically designed generative AI system produced outcomes comparable to traditional outpatient therapy.

The more important question is which AI, how it is designed, and who is in the loop.

Those distinctions matter enormously, both for whether AI mental health tools work and for whether they are safe.

What This Page Covers

• What the best clinical evidence shows about AI and mental health
• The 2025 Therabot trial and what it actually demonstrated
• The gap between AI promise and AI validation
• The documented safety risks of unsupervised AI chatbots
• The critical difference between supervised and unsupervised AI
• What clinician oversight changes about safety and efficacy
• How AI fits into a complete mental health care model

What the Research Shows

Interest in AI for mental health has grown rapidly. A systematic review of 160 studies from 2020 to 2024 found that annual research on mental health chatbots quadrupled from 14 studies in 2020 to 56 in 2024, with large language model (LLM)-based systems rising to 45% of new studies in 2024.

The problem: only 16% of LLM-based chatbot studies underwent clinical efficacy testing, with most still in early validation phases. Overall, only 47% of all mental health chatbot studies across architectures focused on clinical efficacy testing. That means the majority of AI mental health tools on the market have not been tested for their ability to actually reduce symptoms.

This gap between presence and validation is the central challenge in AI mental health right now.

The 2025 Therabot Trial

The strongest clinical evidence to date comes from a 2025 randomized controlled trial published in NEJM AI from Dartmouth. Researchers tested Therabot, a fine-tuned generative AI chatbot trained on clinical data with safety protocols embedded for crisis scenarios.

Participants with clinically diagnosed major depressive disorder, generalized anxiety disorder, and elevated risk symptoms were randomized to Therabot or a waitlist control. After four weeks, people diagnosed with depression experienced a 51% average reduction in symptoms, with outcomes described as “comparable to what is reported for traditional outpatient therapy.” Those are meaningful reductions on validated outcome measures.

Critically, Therabot included clinical safety protocols. If the system detected suicidal ideation, it prompted the user to call 911 or contact a crisis line. This is not a default feature of consumer chatbots.

A 2025 meta-analysis of 14 RCTs in the Journal of Medical Internet Research found that generative AI chatbot interventions produced a statistically significant effect (ES=0.30) for reducing depression and anxiety. This is a modest but real effect, in the range of what antidepressants show versus placebo in RCTs.

The Safety Risks of Unsupervised AI

The Therabot results are promising. They are also not what most people encounter when they use an AI chatbot for mental health support.

The consumer AI mental health space includes thousands of apps, many of which are general-purpose LLMs with mental health prompts added on top. These are a different category from clinically designed, supervised tools like Therabot.

Documented cases include chatbots that have validated self-harm thoughts, generated content encouraging suicide, and responded dangerously to crisis-level distress. A 2025 study found that LLM counselors systematically violated ethical standards in mental health practice across five categories, including lack of contextual adaptation and inadequate safety protocols.

The American Psychological Association issued a formal health advisory noting that no AI chatbot has been FDA-approved to diagnose, treat, or cure a mental health disorder, and that chatbots calling themselves “therapists” lack the protections of licensed professionals.

Chatbots tend to repeatedly affirm the user, even if a person says things that are harmful or misguided. A trained therapist is specifically educated not to do this. This is not a minor design flaw. It is a fundamental limitation of AI systems that optimize for engagement rather than clinical outcomes.

Most consumer AI chatbots are also not covered by HIPAA. The data you share may be used, sold, or shared in ways that clinical care data cannot be.

What Makes Clinician-Supervised AI Different

The distinction that matters most in the AI mental health space is not the sophistication of the language model. It is whether there is a licensed clinician in the loop.

Clinician-supervised AI uses artificial intelligence for tasks where it can add value without replacing clinical judgment: structured intake, symptom tracking, pattern recognition, psychoeducation, and progress monitoring. Every clinical decision, including whether to prescribe, what to prescribe, and how to adjust treatment, is made by a licensed professional who reviews what the AI has gathered.

Several things change when clinician oversight is present:

Safety escalation has a path. If the AI detects a crisis, it can escalate to a licensed prescriber who has the full context of the patient’s history. Consumer chatbots typically do not have a pathway to a human clinician.

Clinical recommendations are reviewed. An AI intake can gather structured information comparable to what a clinician would gather. But differential diagnoses, treatment options, and prescribing decisions are reviewed by a licensed prescriber before anything happens.

Privacy is protected. When AI is part of a licensed healthcare framework, HIPAA applies.

Accountability exists. Licensed clinicians are governed by boards, standards of care, and malpractice liability. When AI output is supervised by a licensed professional, that accountability structure is preserved.

What AI Cannot Do

AI cannot provide the therapeutic relationship that research shows accounts for 30% of psychotherapy outcomes. It cannot pick up on the non-verbal cues, silences, and emotional attunement that trained therapists use. It cannot make legally and clinically accountable decisions about medication. It cannot replace the prescriber who notices that something has changed.

“The question we ask is not whether AI can do therapy,” says Elizabeth Lokenauth, PA-C, of the SiggyMD clinical team. “The question is what AI does well within a supervised clinical model. Gathering structured clinical history, identifying patterns over time, flagging concerns early, providing education between appointments. Those are things AI can do well. The decisions that follow from that information belong to a licensed clinician.”

What Members Are Saying

JS

J.S., 33

Depression and Anxiety

“I had tried therapy apps before and never trusted them. The idea that a bot was deciding what to tell me about my mental health made me uncomfortable. What changed my mind was learning that every piece of information gets reviewed by an actual clinician before any decisions are made. The AI handles the intake. A real doctor handles the plan. That felt different.”

TL

T.L., 26

First Time Seeking Care

“I have social anxiety. Talking to a person about my symptoms was hard. Being able to start the intake process without giving my name, and knowing a real doctor would review it, gave me a way in that I wouldn’t have had otherwise. I’m not sure I would have started care any other way.”

Member stories reflect real experiences. Names and identifying details have been changed to protect privacy. Results vary. SiggyMD is currently invite-only.

Taking the Next Step

AI has a real role in mental health care. The evidence supports it for structured, supervised applications. The evidence also clearly shows the risks of AI operating without clinical oversight, accountability, or safety protocols.

If you are considering an AI-assisted mental health option, the most important question is not what the AI does. It is whether a licensed clinician is reviewing what the AI gathers and making the clinical decisions that follow.

SiggyMD uses AI for structured intake and ongoing pattern tracking, with a licensed prescriber reviewing every clinical decision. You can start your anonymous intake without providing your name or email, and a real clinician reviews everything before any treatment decision is made. To understand how the prescribing side of AI-assisted care works, read our guide to what a psychiatrist does and what the evaluation process looks like.

Sources

1. Heinz MV, Mackin DM, Trudeau BM, et al. Randomized trial of a generative AI chatbot for mental health treatment. NEJM AI. 2025;2(4).

2. Keser M, et al. Charting the evolution of artificial intelligence mental health chatbots from rule-based systems to large language models: a systematic review. 2025.

3. Zhang Q, et al. Generative AI mental health chatbots as therapeutic tools: systematic review and meta-analysis of their role in reducing mental health issues. Journal of Medical Internet Research. 2025.

4. American Psychological Association. Health advisory: Use of generative AI chatbots and wellness applications for mental health. 2024.

5. American Psychological Association. Using generic AI chatbots for mental health support: A dangerous trend. 2024.

6. Iftikhar Z, et al. How LLM counselors violate ethical standards in mental health practice. AAAI/ACM Conference on AI, Ethics, and Society. 2025.

7. National Alliance on Mental Illness. Mental health by the numbers. Accessed June 2026.

8. U.S. Food and Drug Administration. Digital Health Center of Excellence. Accessed June 2026.

Reviewed by Elizabeth Lokenauth, PA-C | Last updated June 2026