This essay is part of the At The Threshold series. Read the overview and full reading order on the At The Threshold page.
Featured image: Rorschach-style AI inkblot created by the author using Canva AI, illustrating the meeting point between machine intelligence and everyday mental health.
Mental health care has always been constrained by scarcity: too few practitioners, too much need, and too many people who cannot access support in time or at all.
The World Health Organization (2022) estimated that the global treatment gap — the proportion of people with mental health conditions who receive no treatment — exceeds 70 per cent in low- and middle-income countries and remains substantial in high-income countries, where cost, stigma, and waiting times constitute persistent barriers to care.
Against this backdrop, the emergence of AI-powered mental health tools — conversational agents, digital therapeutics, mood-tracking systems — has attracted significant attention, significant investment, and significant controversy. This article examines what the evidence shows about where AI can legitimately assist in mental health support, where it cannot, and what the distinction requires us to understand about the nature of therapeutic care.
Introduction
Conversational AI tools designed for mental health support represent a specific application of the broader shift in AI accessibility described in the preceding article in this series. Rather than general-purpose language systems, tools such as Woebot and Wysa are purpose-built to deliver evidence-based psychological techniques — primarily Cognitive Behavioural Therapy (CBT) — in a structured, conversational format.
The first randomised controlled trial of such a system, published by Fitzpatrick et al. (2017) in JMIR Mental Health, found that students who used Woebot over a two-week period reported significantly greater reductions in anxiety and depression symptoms than a control group who received psychoeducational reading materials. The study was small and short in duration, but it was methodologically rigorous for its time and established the empirical foundation for a rapidly expanding field.
What AI Can Do Well in This Context
The case for AI-assisted mental health tools rests on a specific set of affordances that are genuinely distinct from what human practitioners can provide. These tools are available at any hour of the day or night, at no marginal cost per session, with unlimited patience and without the social anxiety that many people experience in face-to-face therapeutic contexts.
For individuals who cannot access professional support — because of cost, geography, waiting times, or stigma — a well-designed AI tool that delivers structured psychoeducation or guides a CBT thought-challenging exercise may represent the difference between some support and none.
The evidence base for these affordances is most robust in a specific range of applications: structured delivery of CBT techniques for mild to moderate anxiety and depression, mood tracking, psychoeducation, and bridging support between sessions for people who are already working with a human practitioner. These represent genuine, if bounded, contributions to the landscape of mental health support.
The Limits of Simulation
The more serious questions arise when AI mental health tools extend beyond this bounded role — when they are positioned, implicitly or explicitly, as substitutes for clinical care rather than supplements to it.
The therapeutic relationship — what Carl Rogers (1957) described as the set of necessary and sufficient conditions for therapeutic change — is not primarily a vehicle for technique delivery. Rogers identified three core conditions that the evidence has consistently supported as drivers of positive therapeutic outcomes: empathic understanding, unconditional positive regard, and congruence (the therapist's genuine presence in the relationship).
These conditions are relational and dyadic — they exist in the space between two people, and they require a therapist who is genuinely affected by the client and genuinely present with them. Subsequent meta-analytic research has consistently confirmed the therapeutic alliance — the relational bond between client and therapist — as one of the strongest predictors of treatment outcome across modalities.
AI systems can simulate the language of empathy with considerable fluency. They cannot be genuinely present, genuinely affected, or genuinely changed by the encounter. This is not a technical limitation to be resolved in the next generation of models. It reflects a fundamental difference in the nature of the systems involved.
The Crisis Problem
The most urgent limitation of AI mental health tools is their capacity to manage acute psychological crisis. Conversational agents operate through pattern recognition and conditional logic, however refined their underlying models may be. A person in genuine psychological crisis does not present in predictable, classifiable ways — the signals of distress are often indirect, contextually embedded, and distributed across history, tone, and behaviour rather than a single statement.
A skilled clinician will notice a shift in affect, a hesitation, a change in engagement that precedes explicit disclosure of suicidal intent or self-harm risk. They will also draw on knowledge of the client’s history, social context, and previous patterns of coping to interpret those signals and decide when to probe further, when to slow down, and when to mobilise additional support. A text-based AI system may not surface these cues in time, and even systems that flag obvious crisis phrases can miss the quieter forms of despair that clinicians recognise as high risk. In this domain, the consequences of failure are not recoverable in the way that a missed recommendation about exercise or sleep might be.
For this reason, most professional bodies and safety guidelines position AI tools as inappropriate for managing acute risk, recommending instead that any system offering mental health support provide clear, persistent signposting to crisis services and encourage users in distress to seek human help immediately.
Drawing the Line
The ethical architecture required, then, is not a prohibition on AI tools in mental health contexts but a clear and enforced differentiation of roles. On one side sit AI systems as first-line resources, psychoeducational supplements, and triage tools: they can lower the activation energy for seeking help, provide coping strategies for mild to moderate distress, and route people more efficiently toward appropriate human care. On the other side sits the domain of primary clinical responsibility, where diagnosis, treatment planning, and risk management require an accountable human professional who can exercise judgement in the full sense of the term.
At present, that boundary is blurred in both regulation and commercial messaging. Some products market themselves in ways that imply parity with therapy — promising “24/7 AI clinicians” or “therapy without the therapist” — even when the underlying evidence base is narrow and escalation pathways are unclear. Policy and professional standards are only beginning to catch up, and different jurisdictions are moving at different speeds. Clarifying this line — in regulation, in clinical guidelines, in product design, and in the expectations set for users — is among the more important tasks facing mental health systems in the years ahead. How we draw it will determine not only who gets access to support, but what we are prepared to count as care when the machine enters the room.
References
Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19. https://doi.org/10.2196/mental.7785
Rogers, C. R. (1957). The necessary and sufficient conditions of therapeutic personality change. Journal of Consulting Psychology, 21(2), 95–103. https://doi.org/10.1037/h0045357
Wampold, B. E., & Imel, Z. E. (2015). The great psychotherapy debate: The evidence for what makes psychotherapy work (2nd ed.). Routledge.
World Health Organization. (2022). World mental health report: Transforming mental health for all. WHO. https://www.who.int/publications/i/item/9789240049338
Author Note (AI Usage): This article was drafted with assistance from a generative AI system to organise structure and suggest phrasing. All facts, citations, and final editing have been verified and approved by the author. The AI did not access any private health data.
Continue in this series: At The Threshold: The Attention Economy Evolves. Or return to the At The Threshold overview.
Comments
Post a Comment