Summary: An expert consensus statement titled “Towards responsible AI for mental health and well-being” came out of the World Health Organization on March 20, 2026, and it targets something already underway with virtually no oversight in place: people in large numbers — young people most of all — have turned generative AI tools into outlets for emotional support, even though those tools were neither built for mental health nor ever clinically tested in that role. Teams shipping general-purpose AI should drop the assumption that “we’re not a medical product” settles the matter — under WHO’s framing, emotional-companionship scenarios now sit inside product responsibility.
An expert consensus statement titled “Towards responsible AI for mental health and well-being” came out of the World Health Organization on March 20, 2026, and it targets something already underway with virtually no oversight in place: people in large numbers — young people most of all — have turned generative AI tools into outlets for emotional support, even though those tools were neither built for mental health nor ever clinically tested in that role. The experts’ verdict leaves little room for doubt: what we are looking at has crossed into public health territory; it is no longer a question of product design.
At the root of the consensus sits an online workshop held on January 29, 2026 and hosted by the Delft Digital Ethics Centre (DDEC) at TU Delft, drawing more than 30 international experts from AI, mental health, ethics, and public policy. DDEC holds the distinction of being the first WHO Collaborating Centre focused on AI for health governance, including ethics, and the workshop carried official pre-summit status for the India AI Impact Summit 2026.
Behind the Workshop: A Permanent Network Takes Shape
The statement deserves a read, but the institutional machinery around it deserves just as much. In the same announcement, WHO confirmed it is setting up a Consortium of Collaborating Centres on AI for Health, with member institutions covering every one of the six WHO regions; the candidate members had already completed a first gathering at TU Delft on March 17–19, 2026. Read together, these facts make the statement the opening deliverable of a standing governance network rather than a standalone declaration — standards, evaluation frameworks, and cross-border coordination are the natural follow-ups.
Nobody states WHO’s position more plainly than Dr. Alain Labrique, the organization’s Director of Data, Digital Health, Analytics and AI: as AI comes into contact with people “in moments of emotional vulnerability” ever more often, safety, accountability, and well-being have to sit at the top of these systems’ priorities. Sameer Pujari, WHO’s AI lead, was blunter: “We are at a critical juncture,” and the spread of AI into daily life “has far outstripped investment in understanding its impact on mental health.”
The Three Recommendations at the Center
Three principal recommendations emerge from the consensus:
- Treat the use of generative AI as a public mental health concern, met with coordinated responses from government, health systems, and industry. The scope is drawn on purpose to cover every generative AI solution, not only those marketed around mental health.
- Build mental health into impact assessments and continuous monitoring of AI solutions, with indicators reaching across health determinants, short-term clinical measures, and long-term outcomes.
- Design AI mental health tools jointly with mental health experts and people with lived experience, youth included, keep the work evidence-based, adapt it to cultural, linguistic, and contextual factors, and leave users empowered.
Where the Public Health Framing Comes From
Everything turns on the scope phrase “all generative AI solutions.” Regulatory regimes today mostly place medical-purpose AI in the high-risk bucket and control it tightly — the EU AI Act being the textbook case. The danger the experts single out, though, lives in the general-purpose chatbot: no medical rule binds it, yet it has quietly taken on the job of emotional companion. Dr. Caroline Figueroa of TU Delft, who took part in the workshop, pushed for shared consensus on crisis referral frameworks and accountability systems, and the workshop summary adds a longer-horizon warning: emotional dependence on AI companionship.
That is the logic behind the framing. When risky behavior sits inside a “non-medical” product, the medical-device regulatory path never picks it up; the public health route instead looks at exposure, consequences, and intervention points across whole populations. The pace at which AI has settled into daily routines — the thread running through the industry’s 2026 opening — has left governance trailing behind, and this document is WHO’s formal answer to that gap.
The Practical Read for Product Teams
Three consequences follow. Teams shipping general-purpose AI should drop the assumption that “we’re not a medical product” settles the matter — under WHO’s framing, emotional-companionship scenarios now sit inside product responsibility. Impact assessments will need mental health indicators, and crisis detection together with human referral paths moves from differentiator to baseline, most of all in products aimed at teenagers. And co-design carries real weight here: the document explicitly requires mental health experts and people with lived experience as participants, which lifts the minimum standard for research and design processes.
For developers and product owners, the sensible move is to run the next two years of plans against this consensus as a risk checklist. How your product treats people in their most vulnerable moments is a question you will be asked eventually — by a regulator, a journalist, or both.