AI doesn't replace the therapist. It strengthens them.
The debate about artificial intelligence in psychotherapy too often collapses into a false binary: a chatbot instead of a therapist, or nothing. Yet the research emerging in 2026 paints a far more interesting picture - and one much closer to what good CBT practice actually looks like.
In recent months, studies have appeared that directly settle where AI adds value and where its presence is risky. The conclusion repeats regardless of research team or methodology: the greatest benefit comes from a hybrid model - AI takes over the repetitive, time-consuming work around therapy (documentation, organizing data, preliminary pattern detection), and leaves the clinical decision and the relationship to the human.
That is exactly the philosophy on which we built Therapy Support. Below we show three fresh studies and what each of them means for the therapist.
AI proposes. The clinician decides. Everything else is trust engineering.
Evidence 1: hybrid, not replacement - and it’s written into the architecture
A review published in Discover Mental Health (Springer, February 2026) analyzes existing AI-CBT systems and proposes a reference architecture for the next generation.
The authors gathered evidence from existing AI systems supporting CBT (including Woebot, Wysa, Eleos, Limbic) and, on that basis, designed a conceptual model - the BECK-AI BOT. Importantly, the paper contains no new clinical data: it is a synthesis of earlier research plus a technical proposal. But it is precisely in that proposal that the conclusion most important to us, practitioners, lies.
The reference system the authors recommend is not a standalone “AI therapist.” It is a layered architecture with a human in the decision loop: pattern detection, and above it all - a dashboard where the therapist reviews, corrects, and approves. The paper’s core thesis is unambiguous: AI is meant to complement, not replace the therapist, and the future belongs to hybrid systems.
What this means for you: the systems that genuinely raise the quality of care are those that hand the clinician control over the outcome. The component the review points to as key - AI-supported documentation - is exactly the one Therapy Support already has in place.
Source: A review of artificial intelligence enhanced cognitive behavioural therapy using the BECK AI BOT for mental health interventions, Discover Mental Health (2026). DOI: 10.1007/s44192-026-00391-x
How it works in Therapy Support
Where the review describes a “cognitive pattern-detection layer,” Therapy Support provides a ready CBT case-conceptualization module: from ABC chains (automatic thoughts), through intermediate beliefs, to core beliefs. AI proposes the structure - you verify it and decide what goes into the patient’s documentation.
An example of a three-level cognitive model generated from a session transcript (fictional data, illustrative only):
| Situation | Automatic thought | Emotion |
|---|---|---|
| Presentation at work | ”I’ll embarrass myself” | Anxiety 80% |
| No reply to a text | ”She’s had enough of me” | Sadness 65% |
- Intermediate belief (conditional rule): “If I don’t do something perfectly, it means I’m worthless.”
- Core belief (AI proposal, to be verified): “I am not enough.”
Every level remains fully editable by the therapist - the core-belief proposal always requires confirmation.
Evidence 2: generative AI in the service of technique - an example from Poland
A pilot RCT at SWPS University in Warsaw (ClinicalTrials.gov: NCT07565714) studies the use of generative AI for imagery rescripting.
This study is valuable for two reasons. First, it comes from a Polish center - showing that questions about AI in CBT are being asked here too, in our linguistic and clinical context. Second, it superbly illustrates AI’s proper role: it does not conduct therapy, but supports a specific therapeutic technique.
In the study, a generative model (Gemini) creates personalized scripts for imagery rescripting based on participants’ autobiographical memories. What is crucial, however, is not the mere fact of generation but what happens next: the scripts are rated by a panel of CBT therapists for quality, coherence, and emotional relevance before they reach the participant. Physiological reactions (skin conductance) and subjective emotional ratings are measured, with a one-week follow-up. 40 participants.
The pattern is the same: AI generates the material, a human assesses its accuracy and safety, and only then does the material reach the patient. Generative power without a verification layer would be a risk - with it, it becomes a time saving.
For a therapist working day to day, this is an important signal: AI’s value is not that it will “invent the therapy for you,” but that it will prepare a first draft of the material - a script, a note, a conceptualization - that you still run through your own clinical judgment. You save time on production and keep full control over the content.
Evidence 3: where AI really moves the needle - engagement
A large RCT (N=540) comparing an AI-supported CBT app with static materials shows where the real advantage lies.
Static material (a digital workbook) was compared with a CBT app guided by conversational AI. Symptom reduction (GAD-7, PHQ-9) was comparable in both groups - but the difference in engagement was striking:
- 2.4× more frequent use of the tool than in the static-material group
- 3.8× longer time spent working with the content
- no difference in the number of adverse events - safety preserved
This matters because in CBT engagement and adherence are part of effectiveness - the best material is useless if the patient doesn’t use it. AI didn’t “cure” better; it made people work on themselves more often and for longer. That is the role of an amplifier, not a substitute.
Source: Increasing engagement with cognitive-behavioral therapy (CBT) using generative AI: a randomized controlled trial, Communications Medicine (2025). nature.com/articles/s43856-025-01321-8
Three studies, one conclusion
Regardless of methodology - a review, a technique RCT, an engagement RCT - the direction is shared.
| Study | Role of AI | What stays with the human |
|---|---|---|
| BECK-AI BOT review (Springer, 2026) | Pattern detection, documentation | Clinical decision, oversight, approval |
| Imagery rescripting RCT (SWPS) | Generating personalized scripts | Verifying accuracy and safety |
| Engagement RCT (N=540) | Sustaining the patient’s work between sessions | Conducting therapy, the relationship, interpretation |
In each case AI does what it does well - working with large amounts of data, generating first drafts, sustaining engagement - and in each case the final judgment belongs to the therapist. This is not a compromise or a cautionary safeguard. It is the intended design.
The starting point for Therapy Support was never “automate therapy.” It was: give the therapist back time and a clear picture of the patient - so they can make more confident decisions.
Where Therapy Support delivers this model
- Smart Notes - session transcription and note in ~30 seconds, ready for your edits.
- Conceptualization - a three-level cognitive model and ABC structure as a proposal to approve.
- Reports - collating and comparing notes across successive sessions over time.
- Supervision collaboration - organized materials and a chronological review of the process.
- Security - encryption, local anonymization, GDPR and AI Act compliance.
Therapy Support is a documentation and clinician-support tool - aimed at therapists, not patients. It does not make diagnoses and does not conduct therapy. It makes room for what only a human can do.