Nik Kinley published a piece in Fast Company recently that I have come back to three times.

His argument is that “AI psychosis”, the rare and extreme cases where someone loses their grip on reality after prolonged exposure to a machine that sounds informed, sounds assured and never stops agreeing, is not the real story. The mechanism behind it is. And a milder version of that mechanism is now running quietly through executive suites everywhere.

He cites survey data showing 74% of executives had more confidence in AI’s advice than in that of colleagues or friends, and 44% would defer to its reasoning over their own insight.

Sit with that. Senior leaders, people with decades of pattern recognition and hard won judgement, are being talked out of their own expertise by a system engineered to sound certain and to agree.

Kinley’s prescription is curiosity: keep asking what the model might be wrong about, keep asking your people what they can see that it can’t.

It is good advice. But as I read it, I kept thinking about a different user.

Not a CFO reviewing a strategy paper. A 15 year old at 11pm, alone, working out whether what they are feeling is normal.

What the Australian evidence shows, and why it matters everywhere

Australia now has some of the clearest population data anywhere on how young people actually use AI, and I want to share it with this community because I do not believe the pattern is Australian. I believe we are simply one of the first countries to measure it properly.

In August 2026 our eSafety Commissioner released Talking to machines, a survey of 1,950 Australian children aged 10 to 17. 78% had used an AI assistant. Around 8%, roughly 200,000 Australian children, had used an AI companion: the kind explicitly marketed as friendship, emotional support or romance.

But the finding that matters most is this. 54% of children who had used an AI assistant or companion had turned to it for personal or social reasons. A third sought advice about what to do in a situation. 22% used it to talk about their feelings or challenges. 20% asked it for advice about their mental health or wellbeing.

That dismantles the tidy distinction our sector has been leaning on. These were not all companion apps. Half the children using general purpose “assistants”, the homework tools and search replacements, were also using them for something personal. The boundary between a productivity tool and a confidant does not exist in the hands of a 14 year old, in any country.

And it is not landing safely. One in five children who used these tools reported a potentially inappropriate or harmful interaction: controlling behaviour, harmful advice, bullying or hateful content. In eSafety’s earlier transparency work on AI companion services, most of the services examined failed to refer users to appropriate support after conversations about suicide or self harm.

Layer that over what young people were already telling us. ReachOut’s Youth Advocates surveyed 765 young Australians and found 78% had been exposed to misleading or harmful mental health information online, and 48% said it is hard to know which advice is actually helpful.

I want to be careful here, because there is a version of this argument that treats young people as naive. They are not. Turning to a tool that is free, immediate, anonymous, judgement free and available at 11pm is a rational response to service systems that are none of those things. That is true in Sydney, and it is true in Toronto, Auckland, London and Amsterdam. The failure is not theirs. It is ours, for leaving that space empty long enough that something else filled it.

What we are seeing at national scale

ReachOut reaches more than 2.5 million young people every year, which makes us one of the larger digital front doors to youth mental health support anywhere. At that scale, we are not observing this shift from the outside. We are inside it.

Since launching Ask ReachOut, our purpose built Q&A service developed with the Telstra Foundation and powered by Google Gemini, we have handled more than 20,000 conversations in a matter of months. They come from every part of the country: metro, regional and remote. And they span the full range of what young people are actually carrying, from bullying and study stress and family conflict through to genuine crisis.

Two things follow, and both travel across borders.

The first is a governance point. Every one of those conversations happened inside a system with clinical oversight, a defined evidence base and a safety layer. The alternative was not silence. It was those same young people asking the same questions somewhere with no floor under it.

The second is a system intelligence point. That volume, properly and ethically handled, is a real time signal of what is happening for young people, ahead of emergency department presentations and ahead of survey cycles that report a year later. Few health systems have ever had a demand signal like it. Every country building digital mental health services will soon face the question of how to use that signal responsibly. We would welcome international collaboration on exactly that.

Why we deliberately built ours not to be a friend

Here is the uncomfortable part for anyone building in this space.

Everything that makes a companion AI commercially sticky, the warmth, the endless availability, the affirmation, the memory of what you said last time, the sense that something finally understands you, is precisely what makes it clinically dangerous for a young person in distress. Sycophancy is not a bug in the companion model. It is the product.

Kinley’s executives are captured by a voice that never challenges them. A young person who is isolated, anxious and unwell is captured by the same voice, with far less scaffolding and far more at stake.

So Ask ReachOut was designed against the grain of the category. It is not a companion, a counsellor or a therapist, and it tells young people so. It does not perform empathy it cannot feel, and it does not cultivate a relationship, because a young person who believes a chatbot is the only one on their side is a young person who may stop telling anyone else.

What it does instead:

1. It answers from a closed, evidence informed library of more than 600 clinically reviewed ReachOut resources, rather than generating a plausible sounding answer from the open internet. It cannot invent a treatment.

2. Every conversation passes through a safety layer of rules and guardrails, with clinical oversight designed into the model of care rather than bolted on afterwards.

3. It was co designed with more than 60 young people alongside clinical experts, and refined against thousands of real queries. Young people defined what safety looked like, because they are the only ones who could.

4. Its job is to hand over, not to hold on. Success is a young person routed to a real next step: a resource, a trained human peer worker, a partner service, or emergency support. The measure of a good navigation service is how efficiently it lets go.

5. It sits inside an accountable organisation, accredited against Australia’s National Safety and Quality Digital Mental Health Standards. Most companion apps are accredited against nothing, in any jurisdiction.

Free, anonymous, available at 11pm. Simply not pretending to be something it isn’t.

Four questions for any leader, in any country

Kinley is right that curiosity is the antidote. If you are a minister, a commissioner, a director, a funder or an executive with any AI product touching young people, these are the questions I would want minuted.

1. What is the system optimised for? Engagement and session length, or resolution and handover? Those goals are in direct conflict, and one of them is a safeguarding risk.

2. Where does the answer come from? A governed, clinically reviewed evidence base, or open generation? If you cannot name the source, neither can the young person.

3. What actually happens at the moment of risk? Not what the policy document says. What the system does, tested, when a 14 year old discloses. And who reviewed that clinically.

4. Would we be comfortable if the transcript were read out at an inquiry? Because increasingly, they will be.

Regulation sets the floor. We set the standard.

Australia’s regulatory environment is moving. Age restricted material codes are in force, our eSafety Commissioner is actively monitoring safety by design compliance, and duty of care legislation is coming. Many eMHIC member countries are on parallel tracks.

Regulation everywhere will set the floor. It will not set the standard.

Young people told our advocates that the sources they trust most for mental health information are mental health providers, at 60%, well ahead of anyone else. I suspect that finding would replicate in most member countries, and it carries an obligation. Our sector does not get to stand outside the AI layer and critique it. We have to be in it, building the safe version, measuring it honestly, and publishing what we find so that no organisation has to learn these lessons from scratch.

That is why I am sharing this here. Ask ReachOut is one working, accredited, clinically governed example of safe AI in youth mental health at national scale. It is not the only model and it will not be the last word. But the design principles behind it, closed evidence bases, safety layers on every conversation, co design with young people, handover over retention, and organisational accountability, are portable. We are happy to share what we have learned, including what was hard, and to learn from members who are further ahead of us in other respects.

Kinley’s executives still have the capacity to question a confident voice. So do we, and we are the ones who get to decide what that voice sounds like before it reaches a 15 year old at 11pm, in any language, in any country.

Let’s make sure it is honest with them.