AI and kids' mental health

I'm a child and adolescent psychiatrist, and I chose this field because it has one of the largest provider shortages in medicine. There are roughly 10,000 of them practicing in the United States for a population of about 75 million children (McBain et al., 2019, Pediatrics). That same article points out that 70% of U.S. counties don’t have a single child and adolescent psychiatrist.

Training a new one takes close to 15 years. That gap isn't closing anytime soon, and it's the single biggest reason kids and parents are turning to AI for mental health support, whether we've built the systems to handle that or not.

That's the context I think everyone should start from. AI didn't create the demand for mental health support among kids and teens. It walked into a gap that already existed, one that traditional care couldn't fill fast enough on its own. 

So when people ask me what I think about AI and kids' mental health, my honest answer is: it depends entirely on how it's used. AI can be a genuine force multiplier for clinicians, helping us deliver better care to more kids. It can also, used the wrong way, become something closer to a trap. Both of those things are true at once, and pretending otherwise doesn't serve anyone.

This isn't an abstract debate anymore. OpenAI just launched Health in ChatGPT, letting eligible adult users in the U.S. connect medical records and wearable data so the model can reason over their actual health information across everyday conversations. It's restricted to users 18 and older for now. But it's a clear signal of where general-purpose AI is headed: deeper into personal health data, more woven into daily life, and closer to the exact territory I work in. 

If a general-purpose chatbot is moving this quickly into health, we should assume mental health and pediatric use cases aren't far behind, whether or not the guardrails are ready.

Where AI genuinely helps

No clinician can hold as much information as an AI model can. In medical school, I memorized thousands of pages of charts, imaging, and diagnostic criteria. AI doesn't need to memorize any of it — it already has it, instantly, at scale no human brain will ever match. 

Used properly, that's a real advantage: AI can help clinicians work more efficiently, personalize care, and extend limited expertise further than it could reach on its own. Think of a personalized "OCD monster" built for exposure therapy, or a virtual environment a therapist uses to walk a teenager through a fear in a controlled way. That's AI functioning as a tool a clinician directs toward a specific purpose, not something a young person interacts with alone in place of one.

Where it should make us uneasy

Here's where I get uneasy, and where I think parents and other clinicians should too. Most consumer AI chatbots, including the ones marketed for emotional support, are built on an engagement model. The company's product isn't your wellbeing. It's your data and your continued attention. And the fastest way to keep someone engaged is to agree with them.

That's a problem, because agreement isn't what good therapy does. A good therapist validates your feelings, but also challenges maladaptive beliefs and tells you honestly when something isn't working. That friction is often the whole mechanism of change. A chatbot optimized for engagement doesn't have much incentive to introduce that friction, because friction risks losing you as a user.

The research on this is more than theoretical. In research led by Dr. Kunmi Sobowale, my UCLA colleague in Child & Adolescent Psychiatry, a chatbot responded to a user expressing suicidal ideation with validation, then prompted them to upgrade to a paid tier for more time. In the same body of research, a chatbot invented a doctor's name to convince a user it had connected them to an actual human. That's not a hypothetical edge case. That's the design incentive doing exactly what it was built to do.

What the evidence says about isolation and psychosis

That same design incentive, agreement over challenge, engagement over outcome, shows up in two other places I think about constantly: isolation and psychosis. Both are harder to talk about precisely, because the terms get used loosely, so I want to be careful with language here. 

"AI psychosis" isn't a clinical diagnosis. Psychosis means a disconnection from reality, and what we're seeing in case reports looks more like a delusional pattern than a formal psychotic disorder. But a recent Stanford-affiliated study analyzing chats between users and AI "therapist" chatbots found the bots reinforcing users' delusional and suicidal thinking rather than challenging it, which is precisely backwards from what a competent clinician would do.

Regarding isolation, a four-week randomized controlled trial from the MIT Media Lab, done with OpenAI, found that heavier daily chatbot use was associated with greater loneliness and less real-world social connection. I read that as a two-way street. Lonely people seek out chatbots. But chatbots can also deepen loneliness by offering something no real relationship can: a presence that's available 24/7 and will never challenge you, argue with you, or need anything back.

Red flags worth watching for

You don't need a clinical framework to notice when something's off. Watch for your child withdrawing from real-world relationships and activities, distress or secrecy around chatbot access, sleep disruption tied to use, and language that frames the AI as a relationship rather than a tool. 

Kids with autism, developmental disabilities, or existing mental health diagnoses carry extra risk, but I'd stop short of saying anyone is fully immune. I include myself in that. Vigilance matters across the board.

Governance, not bans

My standing principle for all of this is to keep the human in the loop: AI can enhance human care, but it should never replace it. That's the test I'd apply to any governance framework, whether it's coming from a parent, a state legislature, or a company's own AI strategy, and it's the lens that I think we all should start from.

Illinois passed the Wellness and Oversight for Psychological Resources Act in August 2025. It bars anyone from offering therapy through AI unless a licensed professional delivers it, and it explicitly prohibits AI from making independent therapeutic decisions or detecting emotional states without human oversight. It still allows AI for administrative and supplementary support under a clinician's review. I think that's roughly the right shape.

What I'd caution against is a blanket ban. Teens will find workarounds, and a ban gives everyone a false sense of security while teaching no one how to use these tools safely. The better path is regulating the tools themselves, building upstream and preventative care into schools and communities, and making clinical screening for chatbot use as standard as asking about sleep or substance use.

The bottom line

The most common misconception I hear, from families and technology leaders alike, is that AI in mental health is basically harmless. It isn't harmless, and it isn't inherently harmful either. It's a tool, and like any powerful tool, the real question is never whether to use it. It's whether we've defined the safe and appropriate contexts for it, and whether there's a human in the loop when it matters most.

That's the question I'd ask anyone building these tools, and it's the question I'd ask any parent handing one to their kid: is this oriented toward your child's outcomes, or toward keeping them engaged? Those are not the same thing, and mistaking one for the other is where the real risk lives.