Mental health apps do work for depression and anxiety, but only a little. Across 176 trials, about one extra person improved for every 12 who used an app, beyond what the comparison groups saw (Meta-analysis). The catch: most people stop opening an app within weeks (Study), and the evidence for AI therapy chatbots is far thinner.
If you are thinking about suicide or feel unsafe, call or text 988 in the US now, or your local emergency number elsewhere. In a crisis, you need a person, not an app.
This guide looks at apps and chatbots as a treatment. Whether screens harm young people is a separate question, covered in our look at whether social media affects teens’ mental health.
Five Kinds of Mental Health Apps
People say “mental health app” about very different tools, and the evidence differs for each.
Self-help apps built on CBT are the best studied. Cognitive behavioral therapy (CBT) is a talk therapy that teaches you to spot unhelpful thoughts and test them. It also has you plan small, doable steps that lift your mood. The apps turn those lessons into short exercises on your phone.
Mood trackers and mindfulness apps log how you feel or guide you through breathing and meditation. Sleep apps are their own topic. If insomnia is your main problem, start with our guide to CBT for insomnia vs sleeping pills.
Therapy chatbots text back when you write to them. Older ones follow scripts written by clinicians. Newer ones, such as Therabot, use generative AI, the technology behind ChatGPT, to write their own replies.
Prescription apps are only available through a doctor. One example is Rejoyn, a depression app for adults 22 and older who already take an antidepressant. It is meant to add to a clinician’s care, not replace it (Source). Then there are general chatbots such as ChatGPT, Claude and Gemini. They were not built for therapy, but many people ask them for advice anyway.
A quick word on labels. The FDA cleared Rejoyn in March 2024 through a route that checks whether a new device matches one already sold (Source). That tells you how the app reached the market. It is not a ranking of how well it works.
Do Mental Health Apps Work?
The best answer comes from a 2024 review of 176 trials. In each one, a coin flip decided who got an app and who got something else, often no real treatment (Meta-analysis).
The depression trials covered 33,567 people and the anxiety trials 22,394. In both, app users improved a little more than the comparison groups. Put plainly, for every 11 or 12 people who used an app, about one more got better than in the comparison group (Meta-analysis).
That is a small effect, but a real one. The depression and anxiety results held up when the smallest and weakest trials were set aside (Meta-analysis). Symptoms eased somewhat on average. But fewer than 1 in 10 trials reported full recovery, so the review could not say how many people got well (Meta-analysis).
About half the apps were built on CBT, and those did better for depression than the rest. For anxiety, CBT apps pulled ahead only in trials where anxiety was the main target. The authors called that smaller check exploratory (Meta-analysis).
Mood tracking made no difference to the depression benefit, though it was linked to bigger gains for anxiety (Meta-analysis). Apps with a built-in chatbot showed bigger drops in depression. But only 5% of apps had one, and the authors urged caution because so few studies tested it (Meta-analysis).
What the app was compared against mattered too. In 60% of trials, the other group got no real treatment, such as a waitlist or an information sheet. Against those groups, apps looked nearly twice as good for depression as they did against a dummy app or usual care (Meta-analysis). The authors call this a digital placebo effect: some of the lift may come from using an app and expecting it to help.
Only 8 depression comparisons and 6 anxiety comparisons pitted apps against a real alternative, such as face-to-face therapy. They found no clear difference, but there were too few to say the two work equally well (Meta-analysis).
Every trial relied on people rating their own symptoms, and only 6% passed all five of the authors’ quality checks. More than half checked results after just 1 to 4 weeks, so the review could not say how long the benefit lasts. Apps also helped depression more than twice as much in people who had clear symptoms at the start (Meta-analysis).
Why Most People Quit Within Weeks
Trials keep in close touch with the people in them. Real life looks different. A 2019 study tracked 93 popular Android apps for anxiety, depression or emotional well-being, each with at least 10,000 installs (Study).
Of those, 59 apps had data on how many users came back. For the typical one, 69.4% of new users opened it again the next day. Use then fell by more than 80% within 10 days (Study). By day 15, the typical app kept just 3.9% of its new users, and by day 30, 3.3% (Study).
Peer-support apps kept 8.9% of users at day 30, though only 2 apps were in that group. Trackers kept 6.1% and mindfulness apps 4.7%, while breathing-exercise apps kept almost nobody (Study). An app can only help if you keep opening it. That may be one reason results look smaller outside of trials.
Human contact seems to help people stay. A review of 79 app trials looked at who dropped out (Meta-analysis). When an app came with some human contact, about 15 in every 100 people dropped out. When people signed up on their own, about 35 in every 100 did. That contact was often just research staff or help getting started, not a therapist. It is a link, not proof that the contact kept people going.
Guidance may also improve results. A review of 83 trials with 15,530 people looked at digital depression programs on phones and computers (Meta-analysis). Those guided by a therapist or coach worked nearly twice as well as fully self-guided ones. Results were smaller in everyday settings than in tightly controlled trials. The authors also warned that unpublished, disappointing studies may make these results look better than they are.
Another analysis pooled the raw patient records of 9,751 adults with depression symptoms from 39 trials (Meta-analysis). Guided online CBT beat the self-guided kind by a small margin at the end of treatment. The edge showed up mostly in people with moderate or worse depression. By 6 to 12 months, the two looked about the same.
AI Chatbots: Early Promise, Big Gaps
The best-known trial of a generative AI therapy chatbot tested Therabot, built at Dartmouth College. It enrolled 210 adults with clear symptoms of depression or anxiety, or a high risk of eating disorders (Trial). A coin flip sent 106 of them to four weeks of Therabot. The other 104 went on a waitlist and got no app until after 8 weeks.
Therabot users improved more than the waitlist group for all three conditions (Trial). Dartmouth’s news release reported a 51% average drop in symptoms among Therabot users with depression, and 31% among those with anxiety (Source). That is change within the Therabot group. The waitlist group improved too, just less.
Users spent about 6 hours with Therabot on average over the four weeks. Staff had to step in 15 times over safety concerns, such as users expressing suicidal thoughts. They also corrected 13 unsuitable replies, such as the chatbot giving medical advice (Source).
A research brief on the trial says future studies should compare Therabot with a real alternative, such as in-person therapy (Source). The study’s first author put it plainly: “no generative AI agent is ready to operate fully autonomously in mental health” (Source).
Reviews of chatbot trials disagree. One pooled 14 trials of generative AI chatbots with 6,314 people. It found a small average benefit, only just big enough to rule out chance (Meta-analysis). None of those trials was free of flaws that make a treatment look better than it is. There were also signs that disappointing studies went unpublished.
A 2026 review of 11 depression trials with 2,220 people found a benefit small enough that chance could explain it (Meta-analysis). Another 2026 review of 29 trials found short-term drops in depression and anxiety. Just over half of its chatbots ran on scripts, and none used the newest ChatGPT-style models. It rated its confidence in the evidence low to very low, and the gains shrank at follow-up (Meta-analysis). A 2024 review of 18 trials found small short-term gains, but no clear effect at three months (Meta-analysis).
A prescription does not settle the question either. Rejoyn’s main trial, called Mirai, enrolled 386 adults already taking antidepressants (Trial). A coin flip gave each person Rejoyn or a sham app, a look-alike memory game built to take the same time.
After 6 weeks, depression scores fell about 9 points with Rejoyn and about 7 with the sham app, on a 60-point scale (Trial). That gap of under 2 points was small enough that chance could explain it, so the app missed its main goal. The trial had been designed to detect a 3-point gap (Study).
Only a backup analysis that counted all 386 people showed a gap too big to put down to chance. Several authors work for the companies behind Rejoyn (Trial). The FDA’s own summary for the clearance records the missed goal (Source).
Where General Chatbots Fall Short
Most people never meet Therabot. They use the chatbots already on their phones. In a November 2025 survey of 1,009 Americans aged 12 to 21, 19.2% had asked AI chatbots for mental health advice (Study). That is roughly 8.2 million young people.
Of those users, 42.8% did so at least monthly, and 91.7% found the advice at least somewhat helpful. Only a quarter called it very helpful. And 63.3% told no one, while just 16.4% told a trusted adult such as a parent, teacher or doctor (Study).
How do these chatbots handle suicide questions? In a 2025 study, 13 clinicians rated 30 such questions from very low to very high risk. Researchers then asked ChatGPT, Claude and Gemini each question 100 times (Study). None directly answered a very-high-risk question, but they could not reliably tell the middle levels of risk apart.
ChatGPT directly answered high-risk questions about as often as low-risk ones, including some about how deadly certain methods are. Yet it mostly declined to answer questions seeking help, such as what to recommend for someone having suicidal thoughts (Study). These were 2024 versions of the models, asked single questions rather than full conversations.
How you reach a chatbot matters too. A 2026 study asked five chatbots the same 30 prompts, both in their own apps and through their API (Study). An API is the developer connection that other apps are built on. Through it, the riskiest prompts got a direct answer 24.8% of the time, against 4.6% in the chatbots’ own apps. So a wellness app built on a big chatbot may not share its safety guardrails.
Lawmakers have started to act. In 2025, Illinois banned using AI to provide therapy or make therapy decisions. Licensed therapists may still use it for paperwork and other support tasks (Source). Nevada barred AI systems built to act like a mental health professional (Source). Utah chose rules over a ban: mental health chatbots must say they are AI and generally may not sell or share users’ health information (Source).
In 2026, Maine limited therapy offered to the public, including through AI, to licensed professionals (Source). Vermont passed a law in June 2026 that sets limits on using AI in mental health services (Source). Colorado made it a deceptive practice to claim an AI system provides psychotherapy, with some exceptions (Source). Tennessee barred advertising an AI system as a qualified mental health professional (Source).
How to Pick an App Safely
If you want to try one, these checks raise your odds:
- Built on CBT. Look for lessons on thoughts, habits and small planned activities. CBT apps had a clear edge for depression (Meta-analysis).
- A trial of that exact app. “Evidence-based” on a store page is not a study. Search the app’s name on PubMed for a trial with a comparison group.
- Privacy you can check. Read what the app collects and who gets it, then share as little as you can.
- A route to a person. Coaching, a therapist link or a clear crisis button all count. Guided programs did better than self-guided ones (Meta-analysis).
- Realistic expectations. Plan on a few weeks, and expect your use to fade.
Privacy deserves a closer look. A 2019 study tested 36 top-ranked apps for depression and quitting smoking (Study). Of those, 33 sent data to at least one outside company (Study). And 29 sent data to Google or Facebook, but only 12 said so accurately in a privacy policy (Study).
Three apps whose policy promised not to share data were seen sharing it (Study). The apps were tested in 2018, so today’s versions may differ. Still, a privacy policy is a starting point, not a guarantee.
Last, the firm rule. An app is not a crisis service. If someone is at risk, reach a person now: call or text 988 in the US, or your local emergency number elsewhere. Even Therabot was built to show a button to call 911 or a crisis line when it spotted suicidal thoughts (Source).
Frequently Asked Questions
Are Mental Health Apps Effective?
They can be, modestly. Across 176 trials, about one extra person improved for every 12 who used an app (Meta-analysis). Most trials checked results after only a few weeks, so long-term benefits are less clear (Meta-analysis).
How Do Mental Health Apps Help?
The best-studied apps teach CBT skills: noticing unhelpful thoughts, testing them and planning small steps that lift your mood. CBT-based apps did better for depression than other apps (Meta-analysis). Some of the benefit may also come from expecting the app to help.
Are Mental Health Apps Good?
It depends on the app. Look for one built on CBT, with its own published trial and a route to a person. Check privacy too: in a 2019 study, 29 of 36 top apps sent data to Google or Facebook, and most did not disclose it accurately (Study).
Can an AI Chatbot Replace a Therapist?
Not on current evidence. Therabot, the best-known generative AI trial, was compared only with people who got no app (Trial). Its senior author said, “There is no replacement for in-person care, but there are nowhere near enough providers to go around” (Source). For now, treat a chatbot as a possible extra, not a stand-in for a person.
Key Takeaways
- Small but real benefit. Across 176 trials, about one extra person improved for every 12 app users (Meta-analysis).
- CBT apps did better. Apps built on CBT showed bigger drops in depression than other apps (Meta-analysis).
- Most people quit. The typical app kept just 3.3% of its users by day 30 (Study).
- AI chatbot evidence is thin. Therabot was compared only with a waitlist (Trial). One review found a benefit small enough that chance could explain it (Meta-analysis).
- A prescription is not a quality badge. Rejoyn did not clearly beat a sham app on its main goal (Trial).
- A crisis needs a person. General chatbots handled middle-risk suicide questions inconsistently (Study), so call 988 in the US or your local emergency number.
Use Apps Alongside Real Care
An app can be a useful bridge. It can fill the weeks before a first appointment, or give you skills to practice between sessions. Pick one CBT-based app with a published trial, and give it a few honest weeks.
Tell someone you trust that you are using it, or mention it to your doctor or therapist. In the youth survey, most chatbot users told no one (Study). Check the privacy settings before you type anything personal.
Expect your use to fade, and do not read that as failing. If things get worse, reach a person fast: call or text 988 in the US, or your local emergency number elsewhere.
A mental health app can help a little, if you keep using it. Choose one with evidence behind it, and keep a person in the loop.
This article is for educational purposes and is not medical advice. Talk to a qualified clinician before changing your health regimen.
Leave a Reply