Do CBT Apps Actually Work?

Two findings from the digital CBT literature should not be able to coexist.

In head-to-head trials, guided CBT delivered over the internet performs about as well as sitting in a therapist's office. Meanwhile, out in the real world, the median mental health app keeps fewer than 4 in 100 users past two weeks.

Both are true. So when someone asks "do CBT apps work," the honest answer starts with a counter-question: which part are you asking about? The therapy inside the app, or the app wrapped around the therapy?

This piece walks through what the research shows: the trial evidence, the place where apps genuinely fall short, and the design choices that separate a tool you still use in March from an icon you archive in February.

Do CBT Apps Work? Yes, With a Catch

CBT apps work, with conditions. In controlled trials, digital CBT reliably reduces stress, anxiety, and low mood, and guided programs can match face-to-face therapy. But those results depend on people finishing the program, and most real-world users quit within days. The method is well proven; sustained use is the weak point.

That catch is not a footnote. An effect measured in a trial belongs to the people who completed the program. If you download a CBT app and stop opening it on day four, the trial results no longer describe you.

This matters because nearly every app in the category markets itself with some version of "clinically proven." The phrase usually means the technique inside the app has trial support, which is true and worth something. It does not mean the app, as shipped, produced those results for typical users on their own phones.

Most people comparing options weigh feature lists and testimonials. The research suggests a narrower question: what separates the apps that get used from the ones that get archived? The literature has a specific answer, and it points at three design decisions.

First, though, the good news, because there is more of it than skeptics expect.

The Method Survives the Screen

CBT ports to software better than nearly any other form of therapy, and the trial record backs that up.

The reason is mechanical. Cognitive behavioral therapy is a protocol: notice the thought driving the spike, test it against what you know, replace it with something more accurate. It teaches skills in a sequence.

Anything that can be written down as a sequence can be encoded as software. Open-ended talk therapy depends on the relationship in the room; a protocol travels. A well-built digital program carries the same working parts a therapist would use: a plain explanation of how thoughts drive stress responses, structured exercises for catching and testing those thoughts, and practice that builds from one session to the next.

The trial data agrees. A 2018 meta-analysis pooled 20 trials that ran guided internet CBT head-to-head against in-person CBT for a range of conditions, and the pooled difference between formats was near zero. Guided digital CBT performed about as well as face-to-face therapy.

The authors are careful to say the per-condition samples remain small. But the direction has held across a decade of replication.

Phone apps specifically clear a lower but real bar. A 2024 meta-analysis of 176 randomized trials found small but statistically reliable improvements in depression and generalized anxiety symptoms, with effect sizes a bit under 0.3. Effect size is the researcher's yardstick for how far a treatment moves symptoms: 0.2 counts as small, 0.5 medium, 0.8 large.

Two honest caveats belong here. Many app trials compare against a waiting list, which flatters the app; effects shrink against active comparisons. And reviews focused on everyday stress in working adults, rather than diagnosed conditions, show the same pattern on a thinner evidence base: real reductions, delivered digitally.

So nothing about a screen breaks the core loop of catch, test, replace. If the technique were the problem, the trials would say so. They say something else.

Guided and Unguided Are Different Products

The biggest split in the digital CBT literature is whether a human is attached to the program.

Guided programs pair the software with light therapist or coach contact, often a few minutes per week, usually as short messages reviewing your work and nudging the next session. Unguided programs are the software alone. Almost every consumer CBT app you can download tonight is the second kind.

The best data on this split comes from a 2021 JAMA Psychiatry analysis that pooled individual patient data from 39 trials, roughly 9,700 people. Unguided digital CBT beat control conditions with a small but real effect, about 0.27. For people with mild symptoms, unguided and guided performed about the same.

Once symptoms passed mild, guided programs did meaningfully better.

The mechanism matters more than the number. The human doesn't add content; the program already contains the content. The human adds an appointment: someone who notices whether you showed up and follows up when you don't.

Unguided programs also bleed more people along the way. Dropout runs higher when nobody is positioned to catch the drop-off, which compounds the gap: weaker results, delivered to fewer finishers.

Why did consumer apps drop the human? Economics. A few minutes of therapist time per user per week is workable in a healthcare system and impossible at app-store scale, so the version you can download traded the human away. Understandable, but the evidence cost is what the JAMA analysis measured.

That reframes what "guidance" means for an app. It doesn't have to be a licensed therapist messaging you on Tuesdays. It has to be something that performs the same function: an active reach that arrives on schedule, whether or not you remembered the app exists.

When you compare CBT apps, make this the first spec you look up. What, if anything, plays the role of the human?

Do CBT Apps Work in the Wild? Here the Numbers Turn

Outside of trials, the usage data falls off a cliff.

A 2019 study pulled real-world usage data for 93 mental health apps, measured from devices rather than self-report. The median app had a 15-day retention rate of 3.9 percent. By day 30, 3.3 percent. On any given day, about 4 percent of an app's users opened it at all.

Those are medians across the whole category, not a verdict on any single app. But the shape of the curve is consistent across app types, from breathing exercises to mood logs to CBT modules, and it is steep.

Set that against the trial literature and the contradiction from the top of this article resolves. Trials don't test software alone; they wrap it in scaffolding. Participants are enrolled, reminded, sometimes paid, and followed up by research staff. A trial functions like a guided program even when the app being tested is unguided.

Then the same software ships into the real world, where the scaffolding is a home-screen icon and a notification permission many people decline. The therapy didn't change. The delivery system around it disappeared.

Researchers who build these programs describe dropout as the field's central unsolved problem, and it should drive your buying decision more than any outcome study. An app you stop opening has an effect size of zero, whatever its trials say.

A practical tell when you're comparing: vendors who publish real completion or repeat-use numbers are rare, and the silence is informative. If the retention story were good, it would be on the landing page next to the trial citations.

So ask the week-three question: what does this app do after the novelty fades, when opening it is no longer interesting? Most have no answer. The ones that do share three design choices.

The Problem Is Timing, Not Technique

Apps lose people because they wait to be opened, and stress does not wait.

Consider when a CBT skill is worth the most. It's mid-spike: 2:14 p.m., eleven minutes before the board call, when "this deck isn't ready" is quietly climbing toward "I'm going to lose this company." That is the moment to catch the thought and test it, while it's live.

Now consider when an app gets opened: 9:40 p.m., on the couch, after the day is over and the thought has gone cold. Working the skill hours later is better than nothing. But it's rehearsal, and a skill rehearsed on the couch is hard to summon under load the next afternoon.

Scheduled reminders don't close this gap, and most people have already run that experiment. A reminder that fires at a random calm moment teaches you the buzz means nothing. After a week of that training, it gets swiped away, then switched off.

The engagement research keeps converging on the same fixes, and none of them are breakthroughs in psychology. They're delivery engineering:

• Timing: the intervention reaches you at the moment of stress, instead of hoping you remember it exists. This is the same function trial staff and human guides serve, moved from a weekly cadence to the minute it's needed.

• Brevity: sessions short enough to fit the gap between meetings. A 20-minute module has no slot in a working day, so it gets deferred to the evening and then skipped. A 3-minute one has a slot almost anywhere.

• Guidance: the session walks you through each step, live. A worksheet library asks you to be your own therapist at the worst possible moment, which is precisely the skill you haven't built yet.

Each fix targets a different failure: timing beats the forgotten icon, brevity beats the skipped session, guidance beats the blank page.

None of this requires new therapy. It requires the software to know when you're stressed, and until recently, software had no way to know.

What a CBT Session Built for the Moment Looks Like

The sensor half of this problem got solved on your wrist, which makes the spec above buildable today.

An Apple Watch reads heart rate and HRV, heart rate variability, the tiny fluctuations between beats that shift when your nervous system ramps up. That gives software the input it never had: a live physiological read on rising stress, instead of waiting for you to notice and self-report. If you want the mechanics, here's how Apple Watch stress detection works.

We built Momomoon's Currents to this spec. When detection sees your nervous system spike, Momomoon reaches you in the moment, and a Current takes it from there.

A Current is a 3-minute guided session grounded in the same cognitive techniques the trial literature tested. It names what's happening in your body, walks you through catching the thought driving the spike, tests that thought, and reframes it. Step by step, with you the whole way, no blank worksheet.

Three minutes is a deliberate spec, because the gap between meetings is the only slot a stressed workday reliably has. The session ends before your next block starts, which is what makes the tenth use as plausible as the first.

A Current is also distinct from Momomoon's 90-second haptic reset on the watch. The reset settles your body when a spike hits. A Current goes after the thinking that produced the spike, which is the part CBT was built for.

To be straight about the evidence, since this whole article is about evidence: the trials above tested other programs. Momomoon has not run its own randomized trial, and I won't borrow the literature's numbers by association.

The claim here is a design claim. The research is unambiguous about where digital CBT breaks, and every one of those break points is an engineering decision: when the session reaches you, how long it asks for, whether it walks you through. Those three gaps were the first three lines of the spec.

So, do CBT apps work? The technique cleared its bar decades before smartphones existed. Delivery is the open question, and unlike a trial you can't rerun, delivery is inspectable before you commit.

Put three questions ahead of any effect size. When does it reach me? How long does it ask for? Does it walk me through, or leave me alone with a library?

Those answers will predict your outcome better than the studies will.