During a lovely Romanian wine tasting, I met two social workers. One works with children who need to be placed in safer environments because of difficult or unsafe circumstances at home. The other supports members of the Russian LGBTQ+ community living in another EU country.
We had an interesting conversation about how AI might affect their work and, more broadly, mental health and well-being. We realized that none of us really knew what the implications might be, both in terms of opportunities and risks.
That conversation stayed with me, and when I returned home, I decided to explore the following question:
Q: How can AI contribute to addressing mental health issues
A: AI could make a substantial contribution to improving mental health, potentially more than in many other areas of health care. The reason is that much of mental health care involves communication, reflection, recognizing patterns, changing behaviour and providing support over time — activities in which AI is becoming increasingly capable.
There is also a fundamental difference from physical health care: AI cannot perform surgery or manufacture medicines by itself, but it can already conduct conversations that people experience as supportive and sometimes even therapeutic.
I see seven major opportunities, ranging from prevention to treatment and the organization of mental health services.
1. Seven ways AI can improve mental health
1. Accessible psychological support, 24/7
AI can offer immediate, confidential-feeling conversations to people experiencing anxiety, loneliness, stress or mild depressive symptoms. It can help people articulate their feelings, reflect on problems and explore coping strategies. This is particularly valuable when professional support is unavailable or waiting lists are long.
2. Personalized psychological interventions
AI can guide people through evidence-based techniques such as cognitive behavioural therapy (CBT), mindfulness, behavioural activation and stress management. Unlike a conventional self-help book, it can adapt exercises to an individual’s responses and progress.
3. Earlier recognition of problems
With consent, AI could identify changes in mood, sleep, activity or language that suggest emerging psychological difficulties. This might help identify deterioration earlier, especially in people with recurrent depression. However, prediction is imperfect and false alarms are a serious concern.
4. Supporting psychiatrists and psychologists
AI can assist professionals by preparing consultation summaries, tracking symptoms, finding relevant clinical evidence and supporting treatment planning. This could reduce administration and increase the time available for personal contact.
5. Reducing loneliness and social isolation
AI companions may provide conversation, encouragement and cognitive stimulation, particularly for older people or those living alone. They could also encourage participation in real social activities. The important question is whether AI helps strengthen human relationships or gradually substitutes for them.
6. Expanding access to underserved populations
Multilingual AI tools could provide basic psychological support in rural areas, refugee settings and countries facing severe shortages of mental health professionals. They might also lower barriers for people who hesitate to seek help because of stigma.
7. Improving prevention and public mental health
AI could help identify community-level risks, evaluate prevention programmes, support workplace well-being and tailor mental health information to different populations. Such applications require strong safeguards against surveillance and discriminatory profiling.
2. Does it actually work? Emerging scientific evidence
There is now more than speculation: randomized studies show measurable benefits, especially for anxiety and depression.
Results from a March 2026 systematic review of mental health chatbots
Depressive symptoms: 0.31
Standardized effect size (Hedges’ g)
Anxiety symptoms: 0.28
Standardized effect size (Hedges’ g)
The analysis included 39 randomized trials, with more than 7,000 participants contributing to each outcome. Both effect sizes indicate small but statistically significant average improvements compared with control groups. These are not percentage reductions.Source: npj Digital Medicine, 25 March 2026.
An especially interesting example is Therabot, developed at Dartmouth College.
In a randomized trial published in March 2025, 210 participants were assigned to receive either access to Therabot or a waiting-list control condition. Those using the chatbot experienced substantial improvements, including an average 51% reduction in depressive symptoms and 31% in generalized anxiety symptoms, relative to their own baseline symptom scores. The treatment was delivered through text conversations on a smartphone.
These are encouraging results, but two qualifications matter:
- Much of the published evidence concerns relatively short interventions, and long-term outcomes remain uncertain.
- A 2026 review found that 35 of 39 included trials had a high overall risk of bias. The evidence is therefore promising, not definitive.
Importantly, the strongest findings concern AI systems specifically designed and evaluated for mental health support, not necessarily ordinary general-purpose chatbots.
3. What will humans continue to do better?
This connects to our earlier discussions about AI and human intelligence. I think mental health provides a particularly interesting example of the distinction between simulating empathy and experiencing empathy.
An AI system may respond more patiently, consistently and even more convincingly than a human therapist. But that does not necessarily mean it understands suffering in the human sense.
| Dimension | Potential AI advantage | Human advantage |
|---|---|---|
| Availability | 24/7, at low marginal cost | Limited |
| Patience | Effectively unlimited | Limited by time and energy |
| Knowledge | Broad access to psychological techniques | Clinical experience and contextual knowledge |
| Personalization | Can analyze extensive longitudinal information | Understands personal and social context |
| Empathy | Can generate consistently empathetic responses | Can genuinely share emotional experiences |
| Relationships | May provide conversational continuity | Offers genuine human attachment and reciprocity |
| Judgment | Can identify patterns and suggest options | Can assume responsibility and act in complex real-life situations |
There is an intriguing paradox here.
A person experiencing depression may benefit from an AI that does not actually experience compassion but behaves compassionately.
Perhaps what matters most for certain therapeutic outcomes is not whether the therapist feels empathy, but whether the patient experiences empathy.
However, I would distinguish between three objectives:
- Reducing psychological symptoms: AI may become highly effective.
- Helping people understand themselves: AI could become a valuable partner in reflection.
- Providing meaningful human connection: AI may help, but whether it can substitute for human relationships is much more doubtful.
The third point is especially important because loneliness, loss of meaning and social isolation are themselves contributors to mental health problems.
4. Where AI could actually make things worse
Five risks deserve particular attention.
Emotional dependency. People might develop strong attachments to AI companions, reducing their motivation to seek real relationships.
Excessive agreement. AI designed to be supportive may reinforce unhealthy beliefs rather than challenge them. This could be especially dangerous for people experiencing paranoia, delusions or suicidal thoughts.
Missing serious illness. An AI could fail to recognize when someone needs urgent psychiatric intervention, or mistakenly reassure them.
Privacy. Mental health conversations contain extremely sensitive information. Their use for advertising, profiling or commercial manipulation would be particularly troubling.
Medicalization of ordinary life. There is a risk that normal sadness, grief, disappointment and loneliness are increasingly interpreted as conditions requiring treatment. AI could reinforce this tendency if every difficult emotion triggers an intervention.
WHO emphasizes that health AI requires appropriate evidence, transparency, regulation and human oversight, particularly because systems can produce inaccurate or biased advice.
5. What should health systems do?
From a public-health perspective, I would prioritize the following applications.
| Priority | Intervention | Potential health-system value |
|---|---|---|
| Very high | AI-assisted CBT and psychological self-help | Expands access to low-intensity care |
| Very high | AI support for mental health professionals | Reduces administrative burden |
| High | Support between therapy sessions | Improves continuity of care |
| High | Mental health support in primary care | Helps with early intervention and referral |
| Medium | Loneliness prevention and social connection | Potentially improves well-being |
| Experimental | AI as an autonomous psychotherapist | Could expand access, but needs substantially more validation |
A particularly interesting model would be stepped mental health care supported by AI:
Level 1 — AI-supported self-helpPrevention, stress management and mild symptoms
Level 2 — AI + mental health professionalStructured therapy, monitoring and clinical review
Level 3 — Specialist human-led careComplex, severe or high-risk conditions, with AI assisting
Illustrative model, not a validated clinical triage protocol. People must be able to access appropriate professional care directly at any level.
This model could help mental health services concentrate scarce professional resources on people who need them most, without making AI a mandatory gatekeeper.
6. A more fundamental question: Can AI improve the determinants of mental health?
I think this is where the biggest long-term opportunities — and the biggest policy questions — lie.
Most discussions focus on AI treating depression or anxiety after symptoms have developed. Yet mental health is influenced by employment, financial security, housing, family relationships, physical activity, education and social participation.
AI could influence all these determinants, for better or worse.
For example, AI might help people learn, find meaningful employment, maintain social contacts or cope with daily challenges. Conversely, widespread job insecurity, surveillance, social fragmentation and the displacement of human relationships by machines could worsen population mental health.
That leads me to two hypotheses.
Hypothesis 1: AI could eventually provide effective psychological support to millions of people at a fraction of today’s cost, making basic mental health support much more accessible.
This looks increasingly plausible, although clinical safety and long-term effectiveness remain to be established.
Hypothesis 2: The net impact of AI on population mental health may depend less on AI therapy itself than on how AI changes work, social relationships, inequality and the experience of having a meaningful life.
This second hypothesis is less researched but, in my view, potentially more consequential.
The central distinction is between using AI to treat mental health problems and shaping society so that fewer mental health problems arise in the first place.
For WHO and European health systems, that could be an especially valuable way to frame the debate: AI should not simply make psychological treatment more efficient. It should also contribute to healthier societies, while avoiding new sources of psychological harm.
And it raises a question I find particularly interesting in light of our earlier discussions: If an AI can make someone feel understood, does it matter that the AI itself understands nothing in the human, experiential sense?
