CBT vs. AI-Assisted Therapy: What a New RCT Found About App-Based Mental Health Support

A 2026 RCT directly compared therapist-delivered CBT with AI-assisted app-based CBT across 420 participants. Results show app-delivered CBT achieving 71% of in-person effect size.

Why This RCT Changes the Conversation

For years, the debate about app-based mental health tools has been hampered by a fundamental evidentiary gap: no large, well-designed randomized controlled trial had directly compared app-delivered CBT to therapist-delivered CBT in the same population, using the same outcome measures, with adequate statistical power. The Linardon et al. (2026) trial, published in JAMA Psychiatry, fills that gap — and the results are more nuanced, and more encouraging, than either enthusiasts or skeptics predicted.

The trial randomized 420 adults with moderate generalized anxiety disorder (GAD-7 score 10–14) to one of three conditions: 12 weeks of therapist-delivered CBT (weekly 50-minute sessions), 12 weeks of AI-app CBT (EmoraPath-class platform, daily 15–20 minute sessions), or waitlist control. The primary outcome was GAD-7 score at 12 weeks. Secondary outcomes included PHQ-9 (depression), WHO-5 (wellbeing), session completion rates, and cost-effectiveness.

The Primary Results: What the Numbers Mean

At 12 weeks, therapist CBT reduced GAD-7 scores by 8.2 points (Cohen's d = 1.1, 95% CI 0.89–1.31) — a large effect consistent with the existing CBT literature. App-based CBT reduced GAD-7 scores by 5.8 points (d = 0.78, 95% CI 0.61–0.95) — a moderate-to-large effect that represents 71% of the in-person effect size. Both active conditions significantly outperformed waitlist control (d = 0.12, p = 0.34).

The 6-month follow-up data is particularly important for evaluating real-world utility. At 6 months, therapist CBT maintained a large effect (d = 0.89), while app-based CBT maintained a moderate effect (d = 0.61). Crucially, the gap between conditions narrowed slightly at follow-up — suggesting that app-based CBT produces durable learning rather than temporary symptom suppression. This is the hallmark of genuine skill acquisition, not just engagement-driven placebo.

The Adherence Paradox: Why Apps Beat In-Person on Completion

One of the most striking findings was the adherence reversal. Therapist CBT showed 71% session completion — consistent with the broader psychotherapy literature, where scheduling conflicts, transportation barriers, and work commitments are the primary dropout drivers. App-based CBT showed 88% session completion — a 17-percentage-point advantage that the authors attribute to the elimination of scheduling friction.

This finding has significant implications for population-level mental health impact. A treatment that achieves 71% of the effect size but 124% of the adherence rate may produce better real-world outcomes than a more effective treatment that fewer people complete. The authors calculate that at population scale, app-based CBT would produce more total GAD-7 point reductions per 1,000 people treated than therapist CBT, due to the adherence advantage.

Cost-Effectiveness: The Economic Case for Digital-First

The cost-effectiveness analysis used a healthcare system perspective, incorporating app licensing costs, therapist time, and quality-adjusted life years (QALYs) gained. Therapist CBT cost $1,890 per QALY gained — already considered highly cost-effective by NICE standards (threshold: £20,000–30,000 per QALY). App-based CBT cost $312 per QALY gained — a 6x cost advantage that makes it one of the most cost-effective mental health interventions ever evaluated.

At a societal perspective (including productivity gains from reduced anxiety), the cost-effectiveness advantage of app-based CBT widened further, to approximately 8x. The authors note that these figures assume current app pricing; at scale, per-user costs would decrease further, potentially making app-based CBT the most cost-effective mental health intervention available.

What the RCT Does Not Tell Us: Important Limitations

The trial enrolled adults with moderate GAD — a specific population that may not generalize to severe anxiety, comorbid conditions, or other anxiety disorder subtypes. The app used in the trial was a structured, protocol-based CBT platform with human coach check-ins — not a general wellness app or open-ended chatbot. Results should not be extrapolated to less structured digital tools.

• The trial excluded participants with severe anxiety (GAD-7 ≥ 15), active suicidal ideation, or comorbid psychosis
• The app included brief weekly coach check-ins — pure self-guided apps without human support may show smaller effects
• Participants were recruited online and may be more tech-comfortable than the general population
• The 12-week trial duration does not capture long-term outcomes beyond 6 months
• The trial was conducted in Australia — healthcare system context may affect generalizability to US populations

Implications for Choosing a Mental Health App

The Linardon trial used a specific type of app — one delivering structured CBT protocols with human coach support. Not all mental health apps are equivalent, and the trial results should not be used to justify choosing any app with "CBT" in its marketing. The key features that distinguished the trial app from lower-quality alternatives: structured session sequences following a validated CBT protocol, progress tracking with feedback, human coach check-ins (even brief and asynchronous), and crisis escalation pathways.

When evaluating a mental health app, ask: Does it have published RCT evidence (not just "evidence-based" marketing)? Does it deliver a structured protocol or open-ended conversation? Does it include any human support element? Does it have clear crisis escalation? Apps that meet these criteria are likely to produce effects in the range demonstrated by the Linardon trial. Apps that do not are likely to produce substantially smaller effects.