Key points
- ✓AI in mental health has the potential to extend access and personalise care — and the potential to cause serious harm if deployed carelessly
- ✓YouMindo uses AI for pattern detection and personalisation, not for clinical decision-making
- ✓Every AI feature has a clinician in the loop — AI surfaces information; humans interpret and act on it
A few months ago, a journalist asked me whether YouMindo would eventually replace therapists with AI. I said no — and not because of some vague discomfort about technology, but because of a specific, evidence-based belief about what AI can and cannot do in a clinical context. This post is an attempt to be precise about where we've drawn those lines and why.
What AI is genuinely good at in mental health
AI is good at pattern recognition across large datasets in ways that exceed human capacity. In mental health, this creates real opportunities. An algorithm can notice that a user's mood scores have been declining gradually over six weeks in a pattern associated with depressive episodes — more reliably than a clinician reviewing a weekly session, and weeks before the person might self-identify the shift.
AI is also good at personalisation at scale. A clinical team can't individually tailor exercise recommendations for 280,000 users. An algorithm that learns which types of exercises a specific person tends to complete, at what times, under what conditions, and with what results — and uses that to surface the right recommendation at the right moment — can provide a level of personalisation that wasn't previously possible.
These are genuine contributions to care, and we use AI for both of them. Our mood trend detection, risk flagging, and exercise recommendation engine are all AI-driven.
What AI is not good at
AI is not good at the things that are most central to clinical care. It cannot form a therapeutic relationship. It cannot exercise the clinical judgment that comes from years of supervised practice, an intuitive read of what someone isn't saying, or the capacity to make a risk judgement in a moment of uncertainty and live with that responsibility.
The failure modes of AI in mental health also tend to be serious ones. A poorly calibrated risk detection model can both miss people in crisis (false negatives with potentially fatal consequences) and flag people who aren't in crisis (false positives that undermine trust and create unnecessary alarm). These are not the kinds of errors you can simply iterate your way through when real people's safety is at stake.
Our actual framework
The framework we use internally is: AI surfaces, humans decide. AI can flag a pattern, alert a therapist, or surface a recommendation. A human — the user themselves, a peer facilitator, or a clinician — is always in the loop before anything consequential happens.
Concretely: our risk detection model can identify language and mood patterns associated with suicidal ideation. When it does, it doesn't automatically send an alert or restrict access. It surfaces the information to the user first ('We noticed something in what you shared — are you having thoughts of hurting yourself?') and to their therapist if they have one, and it offers a range of human support options. The AI has detected something. A human is making every decision that follows.
What we've chosen not to build
We've made deliberate choices not to build things that are technically feasible but clinically irresponsible. We don't use AI to generate therapeutic content — exercises, psychoeducation, or in-app messages — without clinical review. We don't use AI to make or suggest diagnoses. We don't position any AI feature as a substitute for clinical judgment.
We've also chosen not to use user data to train AI models without explicit informed consent — and we give users a meaningful choice, including the option to use YouMindo without any AI personalisation features.
The honest uncertainty
I want to be clear that our framework isn't perfect, and that reasonable clinicians disagree about where the right lines are. Some of our advisors think we're too conservative and are leaving real clinical value on the table. Others think we should be more cautious still. We take both views seriously.
What we're committed to is being explicit about what we're doing and why, publishing our outcome data including negative findings, and updating our approach as the evidence develops. The worst thing we could do is treat AI in mental health as either a panacea or a threat — it's a tool, with specific strengths and specific risks, that requires ongoing, honest, evidence-based calibration.
We don't claim to have that calibration perfect. We claim to be taking it seriously. In this space, that distinction matters.
Dr. Sarah Chen is a clinical psychologist with 14 years of practice. She co-founded YouMindo and leads the company's approach to ethical AI in care.