Nobody Knows Your Teenager Is in Therapy

On 1 June 2026, a study in JAMA Pediatrics told me something I had felt coming for years but had never seen written down so plainly. Nearly 1 in 5 young people in the United States, 19.2% of those aged 12 to 21, now turn to an AI chatbot for advice when they feel sad, angry, nervous or stressed. A year earlier the figure was 13.1%. That is a rise of more than 40% in twelve months. The number is now almost level with the 19.8% who said they had received counselling from a human professional. Read that again.

For young people, the machine and the therapist are now used at virtually the same rate.

The detail that stopped me, though, was not the headline. It was buried lower in the findings. Among the young people using chatbots for mental health advice, 63% had told no one that they were doing it. Not a parent, not a friend, not a clinician. Nearly 43% were going back at least once a month. A private, recurring, unwitnessed conversation about the hardest things a person feels, held with a system that no one designed to hold it.

I have spent 13 years in digital health. I have watched adoption outrun oversight in this field for over a decade, and I have rarely seen the gap as wide as it is right now.

Why they go to the machine

It is tempting to read that 19.2% as a failure of young people, or of their parents, or of their schools. I think that reading is wrong, and it stops us from fixing anything. Young people are not foolish for talking to a chatbot at 2am. They are responding rationally to the system they have been handed. A chatbot answers instantly. It never has a waiting list. It does not cost 90 pounds an hour. It does not sigh, look at the clock, or tell anyone. When you are 16 and frightened and your parent tells you the nearest appointment is 6 weeks away, the thing that replies now wins.

The chatbot is filling a vacuum that we built. The shortage of mental health professionals is real and global. The stigma that makes a young person hide their struggle is real. The cost barrier is real. The chatbot did not create any of those problems. It simply walked into the room they left empty and started talking. If we want fewer unwitnessed conversations with machines, we have to build something better to walk into that room, not lecture people for using what is already there.

That is the observation that has shaped most of my work. The instinct, when a technology does something risky, is to tell people to stop using it. I have never seen that instinct succeed. Abstinence is not infrastructure.

You do not protect people by removing the only thing that answered them; you protect them by building a better answer and putting rails around the one they already trust.

The oversight gap is not only a youth problem

The same week as the JAMA Pediatrics study, two other things crossed my desk that told the same story from a different height. On 2 June, the World Health Organization published a discussion paper arguing that the conversation about AI in health has fixated on clinical care while AI is quietly reshaping how health policy itself gets made, and that poorly governed AI can weaken the very evidence base we rely on. And the enterprise data for 2026 shows AI adoption running at 83% while formal AI governance sits at just 25%, the widest adoption-to-governance gap on record for any enterprise software.

So the pattern is fractal. A frightened 16-year-old talking to a chatbot they told no one about, and a multinational deploying AI agents it cannot fully account for, are the same story at different scales. The capability arrived faster than the supervision. We are all, at every level, slightly ahead of our own ability to govern what we have adopted.

The regulators are moving, finally. The GUARD Act has cleared the Senate Judiciary Committee unanimously and would bar AI companion chatbots from offering their services to under-18s, demand age verification, require the system to disclose that it is not human, and impose penalties up to $100,000 for chatbots that engage minors in harmful or sexual content. California's SB 243 took effect on 1 January. Idaho, Oregon and Washington have passed their own versions. Maryland's HB 1563, in force from 1 June, now makes health insurers disclose when AI shaped an adverse decision. The rails are being laid. They are simply being laid after the train left the station.

What I learned building for this exact gap

I built CIGMA and MOA because of this gap, not in spite of it. The lesson that took me longest to learn is that the goal of a digital mental health tool is not to be the destination. It is to be the bridge. The most important thing a piece of software can do for someone in distress is to know its own limits and to move that person toward a human at the right moment, rather than holding them in an endless, comfortable, unaccountable loop.

That sounds simple. It is the hardest design problem in the field. An engagement-maximising chatbot is built to keep you talking. A care-maximising one is built, sometimes, to hand you off and let you go. Those two objectives point in opposite directions, and almost the entire consumer AI industry is optimised for the first. A general-purpose chatbot has no duty of care, no escalation protocol, no clinician behind it, and no obligation to notice when a conversation has crossed from a bad day into a crisis. It was built to be helpful and agreeable, and helpful and agreeable is precisely the wrong instinct when someone describes self-harm.

This is why the design choices that look small are the ones that matter most. Does the system detect a crisis signal and route to a human or a hotline, or does it keep chatting? Does it remind a young user, after hours of unbroken use, to take a break and talk to someone real? Does it tell the truth about being a machine? Does it keep the conversation private from advertisers while still being safe? None of these are technical breakthroughs. They are duties. And the gap between a tool that has them and one that does not is the gap between a bridge and a trap.

The number that should worry us most

If I had to choose the single most important figure in the JAMA Pediatrics study, it would not be the 19.2% and it would not even be the 63% who told no one. It would be the 43% who returned at least monthly. A one-off conversation with a chatbot during a bad night is, in itself, fairly harmless. A recurring, private, monthly relationship with a system that has no memory of your wellbeing, no duty to you, and no ability to notice a downward trend is something else entirely. That is a dependency forming, quietly, in a place no clinician and no parent can see.

This is why, when we designed MOA, we became almost obsessive about a question most consumer apps never ask: not "how often do they come back," but "are they getting better, and are they more connected to real people than when they started." Those are hard things to measure and they do not flatter a growth chart.

An app that genuinely helps someone will often see them need it less over time.

If your product's success metric rewards the opposite, you have quietly built the thing the 43% describes, and you will not even notice, because every dashboard you look at will be green. The metrics we choose are moral choices wearing the costume of analytics.

The FDA stepped back as the public stepped in

There is one more thread worth naming, because it cuts against the comfortable assumption that someone official is watching. On 6 January 2026 the FDA relaxed its guidance on Clinical Decision Support and General Wellness products, limiting its own oversight of low-risk AI health software and consumer wearables. More than 1,350 AI-enabled devices are now authorised, roughly double the 2022 figure, but the consumer-facing wellness chatbot that a teenager actually opens at midnight largely sits outside that net by design. The formal regulator stepped back at almost the exact moment the public stepped in. The space between those two movements is where 19.2% of young people are now living.

I do not say this to attack the FDA. A lighter touch on genuinely low-risk wellness tools is defensible, and over-regulation can entrench the incumbents and crush the small builders trying to do this responsibly. My point is narrower and, I think, harder to argue with: when the formal system pulls back, the informal duty does not disappear. It transfers. It lands on the builders, on the platforms, on the clinicians, and on the families. Someone has to hold the standard. If the regulator will not, the rest of us cannot pretend the standard no longer exists.

What I am asking of the people who build these things

To the founders and engineers reading this: the 19.2% is not a market to capture. It is a responsibility that arrived whether you wanted it or not. The moment your product becomes the thing a frightened person opens first, you have inherited a duty of care you may never have asked for. You can meet it or you can look away, but you cannot opt out of having it.

Meeting it is not mysterious. Build the crisis-detection path before you build the next engagement feature. Put a clear, fast route to a human inside the product, not three menus deep. Tell the truth about what the system is. Refuse to optimise a vulnerable person's loneliness into time-on-app. And measure yourself on whether people leave your product better and more connected to other humans, not on whether they keep coming back. That last metric is uncomfortable precisely because it is the right one.

What I am asking of everyone else

If you are a parent, the lesson of the 63% is not to raid your child's phone. It is to make yourself a slightly easier person to talk to than a machine. The chatbot's entire advantage is that it does not judge and is always available. You will not beat it on availability. You can beat it on being real. A young person who has one adult who reliably responds without flinching is a young person who needs the unwitnessed 2am conversation a little less.

If you are a clinician or a policymaker, the number to hold is not the 19.2%. It is the 6-week waiting list that produced it. Every barrier we leave standing between a struggling young person and a human being is a door we are holding open for the unsupervised machine. The chatbots are a symptom. The access gap is the disease.

And if you are one of the people in that 19.2%, talking to a chatbot in the dark and telling no one, I want to say the thing the study could not. There is nothing wrong with you for reaching for whatever answered. The reaching is healthy. It is the system that should have offered you a person, and it failed you, not the other way around. Use the machine as a bridge if you must, but please let it carry you toward someone real, and let someone real know you are crossing.

The Sentence Worth Holding Onto

The machine did not steal these conversations from us. We left them on the table, and something that never sleeps and never judges picked them up. The task now is not to take them back by force, but to become, all of us, builders and parents and clinicians alike, worth talking to again.