AI & Therapy Archives - L'aura and Co

AI and therapy: Guide for practitioners (2026)

par Laura | Juin 4, 2026 | AI & Therapy

Your client arrives on Tuesday and mentions, almost as an aside, that they spent most of last week talking to ChatGPT or Claude about their anxiety. Actual Tuesday morning.

This is the clinical reality now. AI is inside the therapeutic frame whether clinicians invited it or not - and the training programmes, ethics codes, and supervision models haven't fully caught up. This guide maps what the research actually shows, where AI is being used in practice, and what your ethical obligations are under current APA and NASW guidance.

What the research shows

The empirical picture is forming quickly, and a few things are already clear.

A landmark review in Practice Innovations examined how AI-assisted tools interact with deliberate practice frameworks in clinical training. The finding: AI augments skill acquisition most effectively when it works alongside human supervisory feedback - not instead of it. AI tools are best understood as clinical extenders, not autonomous agents.

A parallel line of research looked directly at LLMs as simulated therapy providers. The results were sobering. Models like GPT-4 could produce syntactically competent empathic responses - but consistently failed to track clinical state changes across sessions, couldn't formulate dynamic case conceptualisations, and were unable to detect rupture-repair patterns in the therapeutic alliance. These aren't minor gaps. They're core competencies of effective psychotherapy.

On documentation, the picture is more favourable. A study in JMIR Mental Health found AI-assisted clinical note generation reduced documentation time by 35-45% without measurable loss in note quality - though the study was limited to structured outpatient settings and shouldn't be generalised uncritically.

The pattern across all of this: AI's current utility is in narrow, well-defined tasks. Documentation support. Psychoeducation delivery. Between-session mood tracking. Not in the relational and diagnostic work that defines clinical practice.

· · ·

Clinical applications

AI for Documentation and Admin

This is where AI delivers the clearest value with the lowest risk profile. Natural language processing tools embedded in EHR platforms - Nabla, Heidi Health, Freed - generate structured SOAP or DAP notes from session audio or transcripts. The clinician reviews, edits, signs.

The ethical floor here is informed consent. Clients must be told if any session audio or transcript is processed by a third-party AI system. This is not optional. APA Ethics Code Section 4.02 (Discussing the Limits of Confidentiality) applies, and several state licensing boards have begun issuing explicit guidance on AI-assisted documentation. Check your jurisdiction.

A practical protocol: a two-sentence disclosure in your informed consent document specifying which AI tool is used, what data it processes, and how long it is retained. Review your vendor's Business Associate Agreement if you are HIPAA-covered.

· · ·

AI-Powered Psychoeducation and Between-Session Support

Several platforms - Woebot, Wysa, Limbic - deploy AI-driven conversational agents to deliver structured psychoeducation between sessions, primarily CBT-based skills. The evidence for these tools in mild-to-moderate anxiety and depression is modest but real. A meta-analysis of digital mental health interventions found small-to-moderate effects on depression symptoms compared to waitlist controls, though heterogeneity across studies is high.

The clinical value is step-down support: reinforcing skills between sessions without replacing the therapeutic relationship. The risk is client over-reliance - particularly in attachment-disordered presentations where a chatbot's 24/7 availability can paradoxically intensify dependency rather than build autonomy. Assess accordingly.

The therapeutic frame matters here. If a client is using a between-session AI tool, that use is clinical data. What they bring to it, how they interpret its responses, whether they find it more comfortable than talking to you - all of it tells you something about how they relate to support, authority, and vulnerability.

· · ·

AI in Clinical training and supervision

The Practice Innovations work on AI-assisted deliberate practice shows real promise for specific skill domains - identifying cognitive distortions, delivering psychoeducation, structuring session agendas. The critical caveat: trainees need human feedback to calibrate their self-assessment. AI-simulated role-plays are not a substitute for live supervision. They are pre-work that makes supervision more efficient.

For supervisors: AI practice cases can increase the volume of deliberate practice reps a trainee gets between supervisory sessions. The skill transfer holds when trainees also receive direct feedback from a human with clinical experience. Alone, AI role-plays may produce competent-looking performance without the corrective relational experience that builds genuine clinical judgement.

· · ·

What AI cannot cDo

Clinicians often field questions from clients - and administrators - about whether AI could "replace" therapy. The honest clinical answer is no, and the reasons matter.

Psychotherapy efficacy is substantially mediated by the therapeutic alliance - a relational construct that cannot be reproduced by a system without subjectivity, continuity of self, or genuine stakes in the outcome. AI can simulate warmth. It cannot offer it.

Risk assessment - suicidality, homicidality, child protection concerns - requires human clinical judgement. No current AI system is validated for real-time clinical risk stratification. Using AI in crisis-adjacent work without human oversight is ethically indefensible and, in many jurisdictions, legally actionable.

A note on client use that often goes unaddressed: some clients will find it easier to disclose to an AI than to a human therapist. That preference is clinically meaningful, not a failure. It may reflect shame, social anxiety, fear of judgement, or previous experiences of not being heard. The therapeutic work isn't to redirect them away from AI - it's to understand what the preference reveals.

The Tuesday morning question - your client, ChatGPT, their anxiety - isn't going away. What clinicians do with it is still being written. The research gives us a frame. The ethics codes are catching up. The clinical judgement is yours.

· · ·

Key Takeaways

AI tools are best positioned as clinical extenders for documentation, psychoeducation, and training, Absolutly not as autonomous therapy providers
Informed consent for AI-assisted documentation is ethically and often legally required; review APA Ethics Code 4.02 and your state licensing board's guidance
The therapeutic alliance, crisis assessment, and dynamic case formulation remain exclusively human clinical competencies per current evidence
Between-session AI use is clinical data - what a client brings to an AI tool tells you something about how they relate to support and vulnerability
APA, NASW, and AAMFT are all in active guideline development cycles as of 2026 - stay current

· · ·