Digital CBT for insomnia works. Whether people can live with it is a separate question.

The lack of access to cognitive behavioural therapy for insomnia is widely recognised, and digital delivery is the proposed solution to it [s2]. The most ambitious version of that solution removes the therapist from the loop entirely — a programme that runs unguided, at any scale, at near-zero marginal cost.

Two reviews published this spring address the two halves of whether that works. The efficacy half is settled more firmly than most people assume. The other half has barely been studied.

The efficacy result

A systematic review and random-effects meta-analysis in Sleep and Breathing restricted itself to fully automated digital CBT for insomnia in general adult populations — no therapist guidance, no human contact [s1]. Fifteen randomised controlled trials met criteria, totalling 3,507 participants [s1].

Pooled across those trials, fully automated delivery was associated with substantial reductions in insomnia severity, with a standardised mean difference of −0.82 [s1]. Improvements in sleep initiation and sleep efficiency were consistent, with smaller but still significant improvements in sleep quality [s1].

An SMD of −0.82 is a large effect by conventional interpretation, and it is being produced by software with no clinician attached. That is the finding worth sitting with.

The authors also tested something the field argues about: whether particular delivery platforms outperform others. In exploratory multivariable meta-regression, the association between delivery platform and treatment effect was attenuated after adjustment for study-level characteristics, including publication year [s1]. In other words, apparent platform differences may reflect when and how trials were run rather than the platforms themselves.

The half nobody measured

A mixed-methods systematic review in Sleep Medicine Reviews asked the question efficacy trials skip: is digital CBT-I acceptable to the people receiving it [s2]?

The search yielded 2,347 unique citations, from which 68 studies were included [s2]. Thirty-three of those reported only proxy measures of acceptability — completion rates, adherence figures, dropout — and the results across them were extremely varied [s2].

That leaves 35 studies with actual acceptability data, covering 2,497 participants, from which six interrelated themes emerged: general acceptability, perceived helpfulness, individualised needs, congruence with personal life, functionality and design [s2].

The direction was broadly favourable. But two themes carried consistent negative signal: the interventions were not always congruent with recipients' personal experiences, and did not always allow tailoring to individual needs [s2].

The review's closing note is the important one. Methodological problems in how acceptability is defined and measured across this literature mean improvements in study design are needed before these findings can be trusted [s2]. Half a field's worth of evidence on whether patients will use a treatment consists of counting who did not quit.

Why congruence matters more here than elsewhere

Two of the six acceptability themes are about fit rather than quality: congruence with personal life, and individualised needs [s2]. Those are the themes on which the review found consistent negative signal — the programmes were not always congruent with recipients' personal experiences, and did not always allow tailoring [s2].

That is a specific failure mode, not a general complaint about usability. A programme whose instructions cannot bend to a person's circumstances is a programme that person cannot follow, and the evidence base has no way of counting who that applies to.

The efficacy meta-analysis measures what happens to people who complete the protocol. The acceptability review is about who never gets that far.

A case where automation held up

One recent replication is worth noting because it tested a fully self-guided programme in a population with a competing behaviour.

A trial published in Experimental and Clinical Psychopharmacology randomised 113 heavy drinkers with insomnia to an internet-delivered insomnia intervention (n = 61) or a sleep patient education programme (n = 52), with assessments after the nine-week intervention period and again at three and six months [s3]. The digital arm reported significantly greater improvements in sleep and greater reductions in drinking frequency than patient education, plus greater reductions in depression symptoms and sleep medication use [s3].

The authors list the limitations plainly: relatively high attrition, reliance on self-report measures, and demographic homogeneity in the sample [s3]. High attrition in a trial designed to demonstrate an intervention is exactly the acceptability problem the Sleep Medicine Reviews authors are asking the field to measure properly.

What to watch

The efficacy question for automated CBT-I is closer to answered than the access debate implies. The open questions have moved downstream: whether people with the circumstances that cause insomnia can complete a rigid protocol, whether tailoring can be automated without losing the effect, and whether any of it reaches populations outside the educated, digitally connected samples that trials keep recruiting.

This article is informational and is not medical advice.