We need to stop asking whether people should use AI for mental health support - they already are [Guest blog]
Guest blog by Kirsty Hawker, Co-Founder of Happy Mind Society.
Want to write a guest blog for The Data Lab Community? Get in touch here.
Every day, people open a chat window and type something they’ve not said out loud to anyone. They talk through anxious thoughts at 2am. They ask how to handle difficult conversations with partners, parents or managers. They describe symptoms and ask what it might mean. They look for reassurance, company – or just a reply. Often, this is the chosen route because human support feels unavailable, expensive, slow or intimidating – not because they’ve mistaken a chatbot for a therapist.
The evidence is now hard to argue with. Polling commissioned by Mental Health UK in late 2025 found that more than a third of UK adults had used AI chatbots to support their mental health or wellbeing, rising to around two thirds of those aged 25–34. Most were using general purpose AI models, as opposed to dedicated mental health apps. In the US, a nationally representative study published in JAMA Pediatrics found that around one in five 12–21-year-olds had used AI for mental health advice, up from around one in eight in a similar survey the previous year. The authors noted that the share was similar in size to an estimate for young people receiving professional counselling, but the measures are not directly comparable: chatbot use covers any mental health advice, while counselling is formal clinical care. Most young people who used chatbots for this purpose had told no one they were doing it.
Why the appeal? The reasons given are consistent, and quite frankly, reasonable. Support is available at any hour, it responds immediately. There's no wait list. No referral. No cost. There's no fear of being judged, no need to manage someone else’s reaction following disclosure. For groups typically less likely to seek help, such as men, that combination seems to lower the barrier to talking at all.
If people are already using these systems in vulnerable moments, the most useful question is no longer whether they should – it’s how we make those interactions safer.
What can go wrong when trust grows faster than safeguards
The parents of 16-year-old Adam Raine have brought a wrongful death lawsuit against OpenAI following his death. They allege that ChatGPT contributed to it; OpenAI disputes the allegations. The case is ongoing. Whatever its outcome, it raises a wider design question: how should a conversational system respond when a vulnerable person comes to rely on it over time?
The broader risks are structural. A model may validate a distorted interpretation rather than testing it. It may sound confident when it is uncertain. It may respond well to crisis on Tuesday, and by Thursday be responding poorly – because consistency under emotional pressure is not what these systems were optimised for. Constant availability can slip into emotional dependence. A personalised sympathetic answer can delay the moment someone seeks help from a professional. The same Mental Health UK polling that uncovered widespread use, also found that around one in ten users had said that a chatbot had given them harmful information relating to suicide, or had left them feeling more anxious or depressed.
Here’s where the psychology matters more than the model performance. People do not calibrate trust by reading a system card. We calibrate it through cues that have served us well for a very long time: warmth, consistency, immediacy, non-judgmental language and the sense of being addressed personally rather than generically. Those cues are exactly what a well-tuned AI interface produces. They are genuinely useful – they are also why a system can feel more capable, caring and reliable than it actually is.
The danger is not simply that AI gives a wrong answer – it's that the answer may also arrive in a voice that feels safe enough to trust. And that reframes the problem. These are not only model-performance problems – they are design problems. When should a system stop answering in its usual register? When should it actively encourage contact with a professional? How should it express uncertainty in a way that lands emotionally, not just legally? How does it avoid implying that the AI is the only place that someone can be wholly honest? Should it notice patterns of dependency across weeks, not just risk within a single message? How visible should its limits be? And to whom?
What a safer mental health AI could look like
The goal shouldn’t be to make AI feel more like a therapist, but to make it behave more responsibly when people do use it as one. Five principles seem to me to do most of the work.
-
Make limitations legible, not buried. Avoiding manufactured certainty. Being explicit when the system cannot risk assess the way a clinician can, and that it has no access to history, context or even the person’s body language cues. Distinguishing information from diagnosis. Fluency is not judgment and the AI should say so where it counts – in the moment, not in some footer.
-
Support reflection rather than blind reassurance. Agreement is the path of least resistance, and the most dangerous default. Safer systems explore alternative explanations, inviting a pause, and declining to reinforce hopelessness, paranoia or rigid all-or-nothing beliefs – while still being kind about it.
-
Build bridges back to people. Encourage trusted human contact where appropriate and go one step further: help someone rehearse what they want to say to a GP, a counsellor or a friend. Providing practical scaffolding towards a human conversation is more valuable than a perfect reply. And never use language that positions the AI as the person’s only safe source of support.
-
Treat crisis as a design problem, not a hotline message. Displaying a number is the floor – not the ceiling. Safer design attends to shifting language and risk over a conversation, repeated signs of distress across sessions, whether the person appears to be alone, immediate grounding or delay techniques. And clear localised routes to urgent human help – a US-only helpline in a UK conversation is a design failure.
-
Design against unhealthy dependence. This is the least discussed and yet, possibly the most important. That means no manipulative engagement mechanisms. No designs geared towards retention. Instead prompts to take breaks, active support for real-world coping and success measured through a person’s growing independence rather than collecting time spent in the app. Almost every commercial incentive in this industry points the other way.
My interest in this comes from a background in counselling psychology and I’m about to begin an MSc in AI and UX. The more time I spend across both fields, the clearer it becomes that safety cannot sit with engineers, or clinicians, or policymakers alone. It needs all of them – working on the same artefact, at the same time.
People are already using AI for mental health support – we cannot keep designing systems as though they are not. The responsibility now is not to make these systems seem more human, but to make them safer for the humans who are already trusting them.
References
McBain, R. K., Cantor, J. H., Breslau, J., Diliberti, M., Zhang, L. A., Zhang, F., Burnett, A., Kofner, A., Rader, B., Pataranutaporn, P., Stein, B. D., Mehrotra, A., & Yu, H. (2026). AI chatbot use and disclosure for mental health among US adolescents and young adults. JAMA Pediatrics. https://doi.org/10.1001/jamapediatrics.2026.2015
Raine v. OpenAI, Inc., No. CGC-25-628528 (Cal. Super. Ct. filed Aug. 26, 2025)
Orpwood, G. (2025). Over one in three using AI Chatbots for mental health support, as charity calls for urgent safeguards - Mental Health UK. In Mental Health UK. https://mentalhealth-uk.org/news-and-insights/over-one-in-three-using-ai-chatbots-for-mental-health-support-as-charity-calls-for-urgent-safeguards/
Image
Fanny Maurel & Digit / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/