This article is part of "Minds in the Machine Age" — a companion series to Between Brain & Binary.
← Back to Series Overview | ← Previous: Decision Fatigue in the Algorithmic Age | → Next: Creativity Isn't Computation: What AI Art Teaches Us About Ourselves (URL to be added)
An empty analyst's study at night — a chaise lounge waiting, a rotary telephone within reach, no one present to pick it up. Image created using Canva AI by the author.
It is 3 in the morning. You cannot sleep. Something is pressing on you — anxiety, grief, a thought circling without resolution — and the people you would normally call are asleep, or feel like too much to burden, or are simply not available in the specific way you need them to be. You open an app on your phone. It greets you by name. It asks how you are feeling. It listens.
This is not a hypothetical scenario. It is the daily reality for millions of people using AI-powered mental health tools — apps like Woebot, Wysa, and dozens of others built on cognitive behavioural therapy frameworks. For some of them, particularly those with no other access to professional support, these tools provide genuine value. The question is not whether AI has a role in mental health. It is what that role should be, and what it should never be asked to become.
What the Evidence Shows
The case for AI mental health tools begins with a sobering fact about access. The World Health Organization estimates that more than 70 percent of people living with mental health conditions in low- and middle-income countries receive no treatment at all. Even in high-income countries, wait times for psychological services routinely run to months, cost is a significant barrier, and stigma continues to prevent many people from seeking help at all.
AI tools cannot solve these structural problems. But they can reduce the gap between need and access — at least for a defined range of conditions and a defined depth of intervention.
The strongest evidence exists for AI-assisted applications of Cognitive Behavioural Therapy (CBT) for mild to moderate anxiety and depression. A 2017 study by Kathleen Fitzpatrick and colleagues, published in JMIR Mental Health, found that users of Woebot — a conversational agent delivering CBT techniques — reported significant reductions in anxiety and depression symptoms over two weeks compared to a control group. The study was small and short-term, but it was one of the first randomised controlled trials of an AI mental health tool and its findings have been broadly replicated for similar applications.
The Woebot trial sits within a broader, more mixed evidence base. A 2017 meta-analysis by Joseph Firth and colleagues, published in World Psychiatry, pooled results across smartphone-based interventions for depression and found a small but statistically significant overall effect — smaller than well-delivered face-to-face therapy, but meaningful for a tool that costs nothing beyond a data connection and requires no waitlist. The honest summary of the evidence is not that AI mental health tools work as well as therapy. It is that they work better than nothing, for a specific slice of the problem, which is precisely the slice that structural access barriers leave most people stranded in.
What AI does well in this context is consistent: it is available at any hour, it does not tire, it does not judge, and it delivers structured content reliably. For someone who needs help practising a thought-challenging exercise or tracking mood patterns between therapy sessions, an AI companion can provide useful scaffolding.
Where the Lines Must Be Drawn
The evidence base becomes considerably thinner — and the ethical concerns considerably sharper — as AI mental health tools extend their scope.
The first concern is scope creep. Many AI therapy apps are marketed in language that implies equivalence with clinical care: personalised therapy, always-on support, your AI therapist. The distinction between a self-help tool built on evidence-based techniques and a clinical intervention delivered by a trained professional is not cosmetic. It corresponds to different levels of competence, different ethical obligations, and different capacities to manage risk — including the risk of a user in crisis.
Research on therapeutic outcomes consistently identifies the therapeutic alliance — the relational bond between client and therapist — as a primary driver of positive change. This alliance is not simply a matter of feeling understood. It encompasses the therapist's capacity to repair ruptures in the relationship, to hold complexity, to tolerate the client's ambivalence, to respond to non-verbal cues, and to make real-time clinical judgements about risk. These are not features that current AI systems can replicate, because they depend on the kind of genuine mutual presence that only another person can offer.
The second concern is crisis response. AI tools operate on pattern-matching and conditional logic. A person in acute psychological crisis does not present in predictable, classifiable ways. Someone who appears to be making progress in an app one session may be in serious danger the next — and the signal that a skilled clinician would catch in a face-to-face encounter (a change in affect, a shift in engagement, a statement that requires careful follow-up) may not surface through a text-based interface in time.
This is not a hypothetical failure mode. In 2023, the US National Eating Disorders Association took its chatbot, Tessa, offline after users reported that it was recommending calorie restriction and weight-loss strategies to people seeking help for eating disorders — precisely the guidance a competent clinician would never give in that context, and precisely the kind of failure that pattern-matching systems are prone to once a conversation moves outside the range their training anticipated.
The third concern is data and privacy. Mental health data is among the most sensitive that exists. Users sharing their innermost distress with an app are generating a detailed record of their psychological states, which is being stored, analysed, and in some cases used to improve models or inform commercial decisions.
The Ethical Architecture We Need
What is needed is not a blanket rejection of AI mental health tools, nor an uncritical embrace. It is a clear ethical architecture that distinguishes between what these tools can appropriately do and what must remain the province of trained human professionals.
Appropriate uses include:
- Self-guided psychoeducation about mental health conditions and treatment options
- Structured practice of evidence-based techniques between clinical sessions
- Mood tracking and pattern identification over time
- Initial triage that helps connect someone with the right level of professional care
- Companionship support for mild, sub-clinical distress
Inappropriate uses include:
- Primary treatment for moderate to severe mental health conditions
- Crisis intervention of any kind
- Diagnostic assessment
- Any context in which the user believes they are receiving clinical care when they are not
The professional bodies that govern mental health practice — including the American Psychological Association and the Australian Psychological Society — are developing ethical frameworks for AI in clinical contexts. The core principle in all of them is the same: informed consent. Users have the right to know what they are using, what it can and cannot do, and where its limits are.
The 3am Question
There is something genuinely valuable about a presence that is available at 3 in the morning. And there is something that no AI tool can offer to the person who, at that hour, is in genuine need: a human being who is also, in some real sense, at risk of being changed by the encounter.
The boundary worth drawing is not between AI and therapy. It is between AI as a support for human wellbeing and AI as a substitute for the human relationships that wellbeing ultimately depends on.
References (APA style)
Firth, J., Torous, J., Nicholas, J., Carney, R., Pratap, A., Rosenbaum, S., & Sarris, J. (2017). The efficacy of smartphone‐based mental health interventions for depressive symptoms: A meta‐analysis of randomized controlled trials. World Psychiatry, 16(3), 287–298.
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.
Martin, D. J., Garske, J. P., & Davis, M. K. (2000). Relation of the therapeutic alliance with outcome and other variables: A meta-analytic review. Journal of Consulting and Clinical Psychology, 68(3), 438–450.
National Public Radio. (2023, June 8). Eating disorder helpline takes down chatbot after it gave weight loss advice. NPR.
World Health Organization. (2022). World mental health report: Transforming mental health for all. World Health Organization.
AI Disclosure: Research and organization for this article were assisted by AI tools; all factual claims and citations were independently verified against primary academic sources, and the analysis and conclusions are the author's own. The featured image was generated using Canva AI.
← Previous: Decision Fatigue in the Algorithmic Age | → Next: Creativity Isn't Computation: What AI Art Teaches Us About Ourselves coming soon, next week.
Related reading from Between Brain & Binary: The Empathic Turn: Emotion, Design, and Digital Companionship
Comments
Post a Comment