Skip to content
MyFreud

AI mental-health app reduced anxiety in a randomized trial

An AI anxiety app beat an NHS self-help website across 12 weeks in 316 adults. A third of the app group dropped out, against a fifth of the controls.

3 min read

Pop-art illustration of a man walking outdoors alone.

Key takeaways

  • The app beat an active comparison, an NHS self-help website, not a waiting list. That is a harder test than most digital mental-health trials set themselves.
  • Effects appeared by two weeks and were still present at twelve, with moderate to large effect sizes.
  • A third of the app group dropped out, against under a fifth of the website group. The people who stayed are not a random sample of those who started.
  • Nobody was blinded, and the researchers describe the trial as exploratory. Treat it as promising, not proven.

Most trials of mental-health apps compare the app against nothing, or against a waiting list, which is a low bar dressed up as a control. A trial published in the Journal of affective disorders in 2026 did something more demanding: it compared an AI-powered anxiety app against an NHS self-help website, which is a real thing that real people are actually pointed towards. [pubmed-anxiety-jun23-2026-source]

What the researchers did

Three hundred and sixteen adults aged 19 to 70 with generalised anxiety were randomly assigned to either the app or a national health service website. Anxiety and depression were measured with the GAD-7 and PHQ-9 at the start and then at two, eight and twelve weeks.

The app offered CBT-informed conversational exchanges and interactive tools rather than a static library of advice, which is the difference the trial was built to test.

What they found

By two weeks, the app group had lower anxiety and depression scores than the website group, with medium effect sizes. That is fast.

At eight weeks, people still using the app showed clear reductions in both. Notably, those who had stopped using it still showed medium improvements in anxiety, which suggests at least some of what they learned came with them.

At twelve weeks the differences held, with moderate to large effect sizes.

The number the summary buries

Attrition was 33 per cent in the app group and 18.5 per cent in the website group. By twelve weeks, retention across the study was 54 per cent.

This matters more than any effect size, and it is the thing to hold onto. People stop using an app when it is not doing anything for them, so the group still present at the end is enriched for people it suited. Every benefit measured afterwards is measured on that filtered group. The honest reading is not that the app produces these effects in anyone who downloads it, but that it produces them in the roughly two thirds who keep going.

That the app group dropped out more than the website group is also worth sitting with, given the website is the less engaging product on paper.

How much weight to give it

The researchers call it exploratory and that is accurate. Nobody was blinded: everyone knew which arm they were in, and self-reported symptom scores are exactly the kind of outcome that responds to knowing you were given the interesting thing.

What the design gets right is the comparison. Beating an NHS self-help website is a more meaningful result than beating a waiting list, because it is the choice a person is actually making.

What this means in practice

If you are waiting for therapy, a structured interactive tool is a reasonable thing to use in the meantime, and this trial suggests it beats reading a static page. Two weeks is a fair window in which to judge whether it is doing anything for you.

What it should not do is replace an assessment. An app that helps is good news; an app that lets a problem run for another six months without anyone qualified looking at it is not.

Is a self-help tool the right level for this?

Tick anything that has been true for most of the last month.

0 of 6 ticked

The source

These findings are drawn from “An exploratory randomized controlled trial of an AI-enabled mental health intervention for generalized anxiety” (Allen A, Young AH, Jellesma FC, et al., 2026), published in the Journal of affective disorders. Read the full study on PubMed.

Frequently asked questions

Why does the higher dropout matter so much?

Because the reported benefit is largely measured among people who stayed, and people stay when something is working for them. If a third of those given the app stopped using it, the ones who remained are, on average, the ones it suited. That is not fraud, it is a structural feature of app trials, and it means the effect an unselected person should expect is smaller than the headline. The 12-week retention of 54 per cent is the number to keep in mind.

Is an AI app as good as therapy?

This trial does not address that, because it did not compare against therapy. It compared against a self-help website. What it supports is that a structured, interactive tool beat static self-help information, which is a narrower and more believable claim. For moderate to severe anxiety, the treatments with the strongest evidence remain CBT and medication.

What was the AI actually doing?

By the description, CBT-informed conversational exchanges plus interactive tools rather than an autonomous therapist. That distinction matters: the underlying techniques are established, and what the AI changes is the delivery, the availability at two in the morning, and the sense of responding to your particular situation. It is a delivery mechanism for known methods more than a new treatment.

Should I use one?

Reasonable to try, with the same caveats as any self-help tool. It is available immediately, which matters when the alternative is a long wait, and the risk of harm is low. What it should not do is delay an assessment: if anxiety has persisted for months, is getting worse, or is stopping you doing things that matter, that warrants a conversation with a doctor whether or not an app is helping.

References

  1. 1.Allen A, Young AH, Jellesma FC, et al. ( 2026). An exploratory randomized controlled trial of an AI-enabled mental health intervention for generalized anxiety. Journal of affective disorders. pubmed.ncbi.nlm.nih.gov . doi:10.1016/j.jad.2026.121275