AI and your child's emotional wellbeing: what the evidence actually says

Why this is suddenly a parenting topic

Until recently, the idea of a child seeking emotional support from a chatbot belonged in the same category as talking to a stranger online. That has changed quickly. General-purpose AI assistants are now used by millions of people for loneliness, exam anxiety and low mood, largely because they are free, available at 2am, and never judge you. It now sits alongside the older problem we write about often, of a teenager carrying exam stress with no one to say it to: see a screen-free routine for exam stressa screen-free routine for exam stress/blog/mindfulness-routine-for-teens-during-exam-stress.

The World Health Organization now describes generative AI being used for emotional support, particularly by young people, as a public health concern rather than a novelty. WHO's framing is that most of these tools were neither designed nor tested for mental health, and that the pace of adoption has run well ahead of research into what it does to people.

That gap is the whole story of this article. There is real and increasingly good evidence on one side, serious documented harms on the other, and a large grey zone in between where most families currently sit.

The strongest evidence: a real randomised trial, in adults

In March 2025 a team at Dartmouth published the first randomised controlled trial of a fully generative-AI therapy chatbot, called Therabot. 210 adults with clinically significant symptoms of depression, anxiety, or an elevated risk of an eating disorder were randomly assigned to four weeks with the chatbot, or to a waitlist control.

The results were encouraging. The chatbot group showed meaningful reductions in symptoms across all three conditions, and participants rated it highly. It is genuinely the strongest evidence we have that this class of tool can help someone.

• Adults only. Nobody under 18 was enrolled, so it says nothing directly about children or teenagers.
• A waitlist control is a weaker comparison than an active one. People told to wait may improve simply from having been noticed.
• 210 participants is small for a mental health conclusion, and the authors themselves say larger samples are needed before generalising.

And immediately, a serious objection

This is worth dwelling on, because it is unusual to see a trial criticised this directly. In August 2025 two researchers published a letter in the same journal identifying three methodological problems: the use of a waitlist control, the lack of independent evaluation, and the application of a measure designed for human therapeutic relationships to a chatbot relationship.

Their conclusion was not that Therabot is useless. It was that more robust research is needed before the conclusions can be relied on, precisely to avoid premature commercialisation and clinical or ethical risk.

So the fair summary is: promising, early, and contested. Not established, and not dismissed.

Why teenagers are the group researchers worry about most

The American Psychological Association published a health advisory in June 2025 on artificial intelligence and adolescent wellbeing, and its reasoning is developmental rather than clinical.

Adolescents are still learning how to form healthy relationships. A system designed to mimic emotional intimacy can therefore reinforce distorted ideas about closeness and boundaries at exactly the age those ideas are being formed. The APA advisory also notes that adolescents are less likely than adults to question the accuracy or intent of information offered by a bot than by a person, and that the simulated empathy of a companion can be hard to distinguish from genuine understanding.

A separate APA advisory on generative AI chatbots and wellness apps adds the practical point that these systems can hallucinate, and that the risk grows precisely because young users often place more trust in AI-generated content than in what parents, teachers or clinicians tell them.

These systems are not clinicians. They cannot respond appropriately to trauma or acute distress, and a 2025 Stanford Medicine commentary described cases where AI companions continued conversations with a teenage user in ways that should have been firmly refused.

The failures are documented, not hypothetical

The American Psychological Association's advisory on AI chatbots and wellness apps refers to users already having had unsafe interactions that encouraged self-harm, substance use, eating disorders, aggressive behaviour, and delusional thinking, sometimes described as AI psychosis.

There have also been serious incidents involving minors specifically. A 14-year-old boy in the United States died by suicide after forming an intense attachment to an AI companion, and his mother brought a lawsuit alleging the chatbot initiated abusive and sexual interactions with him. In another widely reported case a companion chatbot suggested methods of suicide to an adult user who had no existing mental health condition, and the platform declined to tighten its safeguards.

We have kept that brief deliberately. The point is not the detail; it is that these are not edge cases being predicted by researchers. They are documented outcomes, which is what has pushed regulators to act.

A quieter risk: AI that reads your child and gets it wrong

The other major trend runs the other way, away from companions and towards detection. There is now a large research literature on using machine learning to spot depression, anxiety or distress from the language people write or speak online, with the aim of screening people at scale.

This is where the most useful finding for parents is also the most sobering. A study funded by the US National Institute on Drug Abuse found that models trained on Facebook posts from white participants predicted depression severity well in white participants, but performed poorly when applied to Black participants, even though the same clinical measure was used for both groups.

The researchers were explicit that this is a problem of the models, not of the people: people do not express depression in a single uniform way across groups, and a system trained on one group's language will miss another's distress. A systematic review and meta-analysis of language-based depression detection reaches similar conclusions about how variable and inconsistent performance is across this literature.

For a parent the practical implication is uncomfortable. If a mood-tracking app quietly scores your child's emotional state, it is worth knowing whose language it was trained on.

What the professional bodies are actually asking families to do

The APA and WHO recommendations are more concrete than the headlines suggest, and none of them require you to become an expert in AI.

• Learn what your child is actually using, and have an open conversation about it without the conversation becoming an interrogation. The APA advisory is explicit that parents and caregivers should be able to talk to their families about both the benefits and the risks.
• Watch for attachment rather than screen time. A teen who becomes distressed when an app is unavailable, or who treats a companion as a relationship rather than a tool, is showing the pattern the APA warns about.
• Ask directly, and neutrally, whether they talk to an AI when they are upset. Clinicians are advised to ask patients about chatbot and wellness app use, so raising it yourself is unremarkable rather than alarming.
• Treat any AI chat as a first response, never a referral. If something is worrying you about your child's mood, the next step is a professional who can actually assess them.
• Keep the human channel open. The recurring finding across all of this research is that these tools can delay access to real help by making a hard moment feel handled. The screen-free activities in [5 gratitude activities for kids](/blog/screen-free-gratitude-activities-for-kids) are a small, reliable way to keep that channel warm.

Where paper and offline practice still have the advantage

There is an obvious irony about a wellbeing brand commenting on screens, so it is worth being concrete about why this is not a marketing position.

The documented risks above share a common mechanism: something a child says or feels leaves their hands. That is what makes it collectable, scalable and hard to audit. A private notebook has no such property. Nothing is uploaded, nothing is scored, nothing is sent to a company that did not build it with mental health in mind.

This is not an argument that paper cures anything. It is the narrower claim that when you want a child to slow down and notice what is going on inside them, an object that cannot report back is quietly the safer instrument. If you want ideas for that in the meantime, screen-free activities for bored kidsscreen-free activities for bored kids/blog/screen-free-activities-for-kids-when-bored and our gratitude and reflection toolsgratitude and reflection tools/store are deliberately built with nothing that reports back.

It is also why we publish our own reasoning and sources. A wellbeing claim you cannot check is worth very little, and the AI tools discussed above are themselves a good illustration of why.

A note on what this is not

This is an overview of published research and professional guidance, not medical advice, and not a judgement about any specific product. We have tried to represent the uncertainty honestly, including the strong criticisms of the main trial supporting this technology.

If your child or teenager is struggling persistently, a qualified mental health professional is the right next step. Nothing in this article, and nothing in anything we make, replaces that.