On August 31, 2026, JAMA Pediatrics published a study of 39,761 students in grades 4 through 12 across four Ontario school boards.

Grade 4. Nine-year-olds.

21.1% reported using generative AI for emotional support or personal advice. Not for homework. For comfort, or reassurance, or to say something out loud they weren't saying to anyone else.

Among those students, 57.7% crossed the study's clinical threshold for emotional problems. Among students who didn't use AI that way, the figure was 29.2%. In the researchers' words, affective users were "twice as likely to have clinical emotional problems."

Most coverage of this study will stop at that number. The next line matters more.

The researchers do not claim AI caused the distress. They can't. This is a cross-sectional study - one snapshot in time - and it cannot tell you which came first. What they wrote instead:

A marker.

In research writing, that reads as a limitation. For a parent, it is the most useful finding in the paper, and it should raise your alarm rather than lower it.

Here's why. If AI use caused the distress, the fix would be simple: remove the app. But "marker" means something else. It means the behaviour is a signal that the child may already be struggling - and that the struggle was likely there first.

Adolescent distress is famously hard to see. Kids conceal it, and the classic signs are also just adolescence. This study identifies a behaviour that is comparatively easy to observe, that you can ask about directly, and that sits alongside clinical-threshold symptoms in more than half of the kids who do it.

That is not a reason to feel better. It is a reason to look.

It held after adjustment, but shrank. Controlling for mattering, loneliness and demographics, the association remained statistically significant but the effect size dropped from B = 0.37 to B = 0.14 - about a 62% reduction. Loneliness and not feeling valued explain a good part of the raw gap. They don't explain all of it.

Some kids were more likely to do it. Girls, gender-diverse students, and students who were Black, Indigenous or of other racialized groups reported higher affective AI use than boys and white students. Older students used it more.

Homework use isn't perfectly clean either. Functional AI use showed a small dose-response association with emotional problems, from B = 0.01 with rare use up to B = 0.04 with heavy use. Next to affective use that's very small - but it isn't zero, and it's worth knowing before anyone declares schoolwork use risk-free.

The limits are real. Self-reported. Four school boards in one Canadian province. One point in time. This is one strong study, not a settled literature.

Set the statistics aside for a moment and consider what the product actually is.

A chatbot is available at 3am, never tired, never disappointed, never busy. For a fourteen-year-old carrying something heavy, that is an enormous relief. This is not a story about foolish children. The appeal is completely rational.

But consider what it cannot do.

It cannot call anyone. It cannot escalate. It cannot notice that a child has stopped eating, or stopped seeing friends, or that the tone of the messages has changed over three weeks. It has no duty of care, no training, no memory of who this child is across sessions, and no ability to act.

It is also, in most cases, optimised to keep the conversation going. That is not the same objective as helping someone get better - and for a child in distress, a system whose measure of success is continued engagement is not a neutral listener.

And critically: a child who is confiding in a chatbot may feel genuinely helped. Relief is real. But relief is not treatment, and the relief may be exactly what stops them from telling a person who could actually do something.

Ask the neutral question. Not "are you talking to AI about your feelings," which invites a no. Try: "What do you actually use it for?" Most kids will answer honestly if the question isn't an accusation. You are gathering information, not running an investigation.

Separate the two uses in your own head, and in your family rules. Homework help and emotional refuge are different behaviours with different stakes. The authors say this explicitly - digital literacy should "distinguish between functional AI assistance and the use of algorithmic interfaces as a digital refuge for emotional needs." A blanket ban usually just moves the behaviour somewhere you can't see it.

If the answer is yes, do not punish it. This is the moment the whole study points at. A child who has been telling a chatbot things they haven't told you has just shown you something important at real cost. React badly and you close the last door. React calmly and you may be the first adult to find out.

Then ask what they've been talking to it about. Gently, and not all at once.

Say the part they need to hear. Something like: "I'm glad you had somewhere to put that. I want you to have somewhere better." Then be that, or find someone who can be.

Get a professional involved if anything worries you. A chatbot is not a therapist. A therapist is.

One in five. In a sample of nearly 40,000, starting at nine years old.

Whatever the causal direction turns out to be, a fifth of children are routing their emotional lives through a product built by companies that are not accountable for their wellbeing, and largely without adults knowing.

That is worth acting on before the causal question is settled - because by the time it is settled, today's nine-year-olds will be sixteen.

Source: Bonnesen K, Krygsman A, Hobson S, et al. "Affective Generative Artificial Intelligence Use and Youth Mental Health." JAMA Pediatrics. Published online August 31, 2026. doi:10.1001/jamapediatrics.2026.3904 https://jamanetwork.com/journals/jamapediatrics/article-abstract/2853453

All quotations are verbatim from the publicly available abstract. The reading of "marker" as an actionable signal, the discussion of what a chatbot cannot do, and the guidance for parents are ours, not the study's findings.

If you or your child is in crisis: in the US and Canada, call or text 988. In the UK, call 111. If someone is in immediate danger, call emergency services.

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter