The meta-analyses agree that mental health chatbots reduce symptoms of depression and anxiety. The effect size is around 0.3, which in this literature is modest but real. The problem lies in what emerges when you read the trials one by one. In the trial that compared a purpose-built therapy app against plain ChatGPT, the app did not outperform the generic chatbot on any measure. In the one that followed 540 people for six weeks, the AI version multiplied usage frequency by 2.4 and time spent by 3.8, and symptoms improved just as much as with static workbooks. And in the meta-analysis with three-month follow-up, the effect had disappeared. This article walks through the five measurements.
The 30-second summary: the seven figures that matter
Each row names the study and the year. Always look at what the chatbot is being compared against.
| Finding | Result | Study and year |
|---|---|---|
| Effect on depression across 38 randomised trials | g = 0.31 (95% CI: 0.17-0.46), with 7,401 participants | Sohn et al., NPJ Digital Medicine (2026) |
| Effect on anxiety across 34 of those trials | g = 0.28 (95% CI: 0.05-0.51) | Sohn et al., NPJ Digital Medicine (2026) |
| Effect at three-month follow-up | No substantial effects on depression or anxiety | Zhong et al., J. Affect. Disord. (2024) |
| Therapy app against generic ChatGPT | No differences on any measure | Kuta et al., JMIR Mental Health (2025) |
| Use of an AI app against workbooks | 2.4 times the frequency and 3.8 times the duration | McFadyen et al., Comm. Medicine (2026) |
| Symptoms in that same comparison | Comparable: no added improvement in anxiety or depression | McFadyen et al., Comm. Medicine (2026) |
| Comparator in the most-cited trial | A waiting list with no access to any app | Heinz et al., NEJM AI (2025) |
What you will find here
• How much these systems reduce symptoms, according to the synthesis of 39 trials.
• The trial where the therapy chatbot drew level with plain ChatGPT.
• Why multiplying usage fourfold did not improve symptoms.
- What happened when measured three months after the conclusion.
1. The magnitude of the impact
The most recent and comprehensive synthesis was published by Jun-Seok Sohn and colleagues in NPJ Digital Medicine in 2026, searching five databases from 2017 with prior PROSPERO registration. Of 39 eligible randomised trials, 38 were analysed for depression, with 7,401 participants, and 34 for anxiety, with 7,621.
The pooled results: a statistically significant reduction in depressive symptoms, with g = 0.31 (95% CI: 0.17-0.46), and in anxiety symptoms, with g = 0.28 (95% CI: 0.05-0.51). On the usual scale, these are small effects.
There exists a helpful nuance within the subgroup analysis: the impact on depression proved greater among clinical and subclinical cohorts relative to non-clinical groups (p = 0.001). In other words, these interventions perform more effectively for individuals experiencing greater distress, which represents the preferred trajectory.
What that figure does not distinguish
An effect of 0.3 against a group receiving nothing is not the same as an effect of 0.3 against a reasonable alternative. Many of these trials compare the chatbot with a waiting list, that is, with doing nothing, and in mental health that systematically inflates the result through expectation, attention and the simple passage of time.
| Study | What it compared against | What it found | What it cannot support |
|---|---|---|---|
| Heinz et al., NEJM AI (2025) | Waiting list, with 210 adults over 4 weeks | The reference trial of an expert-fine-tuned chatbot | Without an active comparator, it cannot isolate the effect of receiving something |
| Kuta et al., JMIR Mental Health (2025) | Generic ChatGPT and a no-intervention group, 185 adults | Both reduce depression; between them, no difference | Small sample and a three-week intervention |
| McFadyen et al., Comm. Medicine (2026) | Digital cognitive behavioural therapy workbooks, 540 adults | Far more use, the same symptoms | Self-guided and unblinded |
| Sohn et al., NPJ Digital Medicine (2026) | Synthesis of 39 trials with any control | g = 0.31 on depression and 0.28 on anxiety | It pools very different comparators |
| Zhong et al., J. Affect. Disord. (2024) | 18 trials with 3,477 participants | Improvements at eight weeks and nothing at three months | Heterogeneous designs and therapeutic approaches |
2. The trial that compared against plain ChatGPT
This is the study that should circulate most and circulates least. Barbora Kuta and colleagues published in JMIR Mental Health in 2025 a trial with 185 adults split into three groups: a generative therapy app based on solution-focused therapy, general-purpose ChatGPT, and a no-intervention control group. Three weeks, nine planned sessions in each active group.
Compared to the control cohort, both active arms showed reductions in depressive symptoms according to the PHQ-9 inventory: d = −0.47 (p = 0.01) within the AI therapy cohort alongside d = −0.44 (p = 0.02) inside the ChatGPT cohort. Regarding anxiety levels and overall wellbeing, neither arm achieved statistical significance.
And the conclusive outcome: the therapy app group failed to surpass the ChatGPT group across every single metric. Discrepancies between them proved negligible and statistically insignificant regarding depression, anxiety, and overall wellbeing.
A brief three-week test ought not to be given excessive weight. Even so, it prompts an unsettling dilemma for the entire industry: if an instrument built expressly for therapeutic work proves indistinguishable from a standard chatbot, the active ingredient might not stem from the clinical framework at all.
3. Far more use, the same symptoms
The trial led by Jessica McFadyen and her team, featured in Communications Medicine in 2026, stands as the most extensive among the separate investigations: 540 adults across the United States dealing with heightened levels of anxiety or depression, whose study design was preregistered, were divided evenly between a cognitive behavioural therapy application driven by generative AI and digital workbooks based on the identical therapy, undertaken independently across a six-week duration.
| What was measured | AI app | Digital workbooks |
|---|---|---|
| Usage frequency | 2.4 times higher | Baseline |
| Usage duration | 3.8 times higher | Baseline |
| Anxiety symptoms (GAD-7) | Comparable | Comparable |
| Depression symptoms (PHQ-9) | Comparable | Comparable |
| Adverse events | No differences | No differences |
The researchers arrive at a candid deduction that resonates industry-wide: during independent application, personalized generative-AI-powered therapy securely boosts user participation significantly past traditional static content, while failing to demonstrate a generalized improvement in alleviating anxiety or depression symptoms.
Put another way: AI solves the problem of people abandoning these tools, which is real and important, and not the problem of making them work better. An exploratory analysis in the same work suggests that those who used the clinical personalisation features did see greater anxiety reduction, but that is a comparison between people who chose to use them, not a random assignment.
4. What remains after three months
Wei Zhong, Jiang-Hua Luo and Hong Zhang published a meta-analysis in the Journal of Affective Disorders in 2024 covering 18 randomised trials with 3,477 participants, searching five databases.
Their immediate outcomes align with the others: gains in depression (g = −0.26; 95% CI: −0.34 to −0.17) and anxiety (g = −0.19; 95% CI: −0.29 to −0.09), reaching peak therapeutic advantage following eight weeks of care.
And the figure that almost never makes the headlines: at three-month follow-up, no substantial effects were detected for either condition. The benefit exists while use continues and does not appear to leave anything behind.
That constraint links directly to a nuance within the domain’s foundational study, conducted by Michael Heinz and co-authors and published in NEJM AI in 2025, which evaluated an expert-adapted generative chatbot involving 210 adults with clinically significant symptoms of major depression, generalised anxiety or high risk for an eating disorder. The control cohort relied on a waiting list with no access to any app throughout the four-week investigation. Such a setup invariably maximizes apparent outcomes, and this very methodology drives a significant share of the broader public discourse surrounding the topic.
Practical ways to handle this
Ranked by the robustness of its supporting data:
• Ask what it was compared against before believing a figure. A waiting list is not an active comparator, and many of these trials use one.
• Do not expect the effect to last. The three-month follow-up meta-analysis uncovered no meaningful impact regarding anxiety or depression.
- When retention is the primary objective, artificial intelligence certainly delivers results. Compared to conventional static content, it boosted engagement frequency by 2.4 times and extended session lengths by a factor of 3.8.
- Do not assume a specialised tool is better. In the only trial that compared them, it did not outperform generic ChatGPT on any measure.
- Consider it an adjunct for mild or moderate situations. The combined effect, g = 0.31, is modest yet significant, displaying greater magnitude among clinical groups than within the broader community.
Frequently asked questions
Do therapy chatbots work?
They exert a genuine, modest impact. A meta-analysis incorporating 39 randomized trials released in NPJ Digital Medicine throughout 2026 calculates a decrease regarding depressive symptoms of g = 0.31 (95% CI: 0.17-0.46) across 7,401 subjects, alongside anxiety symptom drops amounting to g = 0.28 (95% CI: 0.05-0.51), demonstrating more pronounced outcomes among clinical and subclinical cohorts relative to the broader public.
Does the improvement hold over time?
The available evidence says no. The meta-analysis of 18 trials with 3,477 participants found improvements in depression (g = −0.26) and anxiety (g = −0.19), peaking at eight weeks, but detected no substantial effects at three-month follow-up.
Is a therapy app better than using ChatGPT directly?
In the solitary trial that evaluated them, negative. Among 185 adults divided between a generative therapy application, standard ChatGPT and a control cohort, both active arms decreased depression relative to the control (d = −0.47 alongside d = −0.44), while the dedicated app failed to surpass ChatGPT across any metric. It represents a modest, three-week investigation.
What is AI actually for in these apps?
Above all, for getting people to use them. In the trial with 540 participants, the generative-AI version multiplied usage frequency by 2.4 and duration by 3.8 compared with digital workbooks of the same therapy, with comparable results on anxiety and depression and no differences in adverse events. Dropout is a real problem for digital therapies, and that is where it does help.
Why is the comparison group of such vital importance?
Because it changes the reading of the result entirely. The field’s most-cited trial, published in NEJM AI in 2025 with 210 adults with clinically significant symptoms, compared the chatbot with a waiting list that had no access to any app. Against receiving nothing, almost any intervention improves measured symptoms.
What we do not know
• How they compare with human therapy. None of these trials contrasts the chatbot with equivalent face-to-face psychotherapy.
- What happens in a crisis. Trials usually exclude people at acute risk, who are precisely the greatest concern.
• Whether the effect comes from the content or from the conversation. The draw with a generic chatbot suggests the therapeutic design may not be the determining factor.
• Virtually nothing concerning teenagers and older individuals. The 2026 assessment explicitly highlights that investigations within those two demographic cohorts are entirely lacking.
References
1. Sohn, J.-S., Ha, B.-G. et al. Systematic review and meta analysis of chatbots in the management of depressive and anxiety symptoms. NPJ Digital Medicine, 2026. DOI: 10.1038/s41746-026-02566-w
2. Kuta, B., Novak, L. et al. Effectiveness of a Fully Automated Mobile Therapeutic Versus a General Chatbot in Reducing Depression and Anxiety and Improving Well-Being. JMIR Mental Health, 2025. DOI: 10.2196/82642
3. McFadyen, J., Habicht, J. et al. Increasing engagement with cognitive-behavioral therapy (CBT) using generative AI: a randomized controlled trial. Communications Medicine, 2026. DOI: 10.1038/s43856-025-01321-8
4. Zhong, W., Luo, J.-H. and Zhang, H. The therapeutic effectiveness of artificial intelligence-based chatbots in alleviation of depressive and anxiety symptoms in short-course treatments. Journal of Affective Disorders, 2024. DOI: 10.1016/j.jad.2024.04.057
5. Heinz, M. V., Mackin, D. M. et al. Randomized Trial of a Generative AI Chatbot for Mental Health Treatment. NEJM AI, 2025. DOI: 10.1056/aioa2400802
This material is purely informational and does not serve as medical or psychological guidance, nor does it substitute for treatment provided by a mental health expert. Should you find yourself in a dangerous situation, please reach out to your nearest emergency services or crisis hotline.
