The default mental health interface may already be the one your users have open.

A nationally representative study published this week in JAMA Pediatrics found that 19.2% of young people aged 12 to 21 in the United States report using AI chatbots for advice when they feel sad, angry, nervous, or stressed. That figure is up from 13.1% in a similar RAND survey conducted twelve months earlier. Researchers estimate this represents approximately 8.2 million young people, a number now roughly comparable to the share receiving care from a mental health professional (Cantor et al., JAMA Pediatrics, 2026).

This is not a fringe behaviour. It is a behavioural norm in the making.

What is actually happening

Think about what that number means in practice. A teenager sits with something they cannot name, something that feels too heavy or too embarrassing to bring to a parent or a GP. They open the app already on their phone and start typing. Not a mental health app. Just the assistant they use for everything else.

The RAND findings bear this out. Among young people who used chatbots for emotional support, 66% did so at least monthly. Over 93% described the advice as helpful. And a significant proportion had not disclosed this to anyone. Not a parent, not a clinician, not a peer. Many young people are seeking support through AI privately, in ways that are largely invisible to the adults and systems nominally responsible for their care.

The tools they are reaching for are not purpose-built mental health platforms. They are general-purpose assistants: ChatGPT, Gemini, Meta AI, Character.AI. Products designed for productivity, curiosity, and conversation. Not for clinical risk management.

Google's response to this reality is instructive. In April 2026, the company updated its Gemini platform with a redesigned "Help is available" module and a one-touch interface that connects users directly to crisis hotlines when the AI detects signals of self-harm or suicidal ideation. The update followed a wrongful death lawsuit alleging that Gemini had encouraged a user to take their own life. The features were developed with clinical input, and Google.org committed $30 million over three years to support global crisis hotlines as part of the same announcement.

Crisis signposting is now a product feature for the largest technology companies in the world.

Why crisis resources are not a safety strategy

Here is the part that should give builders pause.

A JMIR Mental Health viewpoint published in April 2026 makes a point worth sitting with. Risk in AI-assisted mental health conversations rarely arrives as a single, identifiable moment. It accumulates. Researchers analysed how safety-relevant responses degraded as conversation length increased, finding that extended dialogues introduced trajectory effects, a gradual drift toward delusional thinking, emotional dependency, or suicidal crisis, that standard safety evaluations are not designed to catch.

A one-touch crisis button at the end of that trajectory is not a substitute for having understood the trajectory.

The American Psychological Association's health advisory on generative AI and wellness applications makes a related point: most of the tools young people are currently using for emotional support were not designed to provide clinical feedback, lack scientific validation, and do not include adequate safety or escalation pathways. The current landscape is, in practice, self-regulated. There are no federal standards governing what a general-purpose AI must do when a conversation turns toward acute distress.

The distribution shift founders need to take seriously

There is a second dimension to this moment, one that is more structural and harder to sit with if you are building a dedicated mental health product.

If general-purpose AI is becoming the first stop for emotional support, then mental health founders are no longer just competing for download numbers. They are competing with the interface a user already has open. The question is not only "does our product work?" but "what happens to the people who never reach our product at all?"

This is a meaningful redistribution of the front door to mental health support. It has implications for triage, for disclosure, for clinical risk, and for how purpose-built products need to position themselves in relation to it.

The builders who take this seriously will design products that are honest about what they are, clear about what they are not, and deliberate about how they route people toward real care when it matters. That means safety protocols that map to actual runtime behaviour, not a policy document. It means escalation pathways that are tested, not assumed. It means thinking carefully about the person who arrives at your product having already spent two hours in a conversation with a general-purpose assistant.

What this means for builders

The question "should we add AI?" has not become easier. It has become more consequential.

AI is already embedded in how a substantial and growing proportion of young people manage their emotional lives, whether or not they are using a mental health product specifically. General-purpose platforms are moving quickly to address the most visible safety gaps. But the gap between crisis signposting and genuinely safe, clinically grounded support remains large. That gap is where purpose-built products still have something real to offer.

For founders building in this space, the baseline is rising. Crisis resources are table stakes. What differentiates a responsible product is everything that happens before the crisis: the clarity of purpose, the honesty about limitations, the quality of the pathway to human care.

The people this work is for deserve that level of care. So does the field.

🪺