Canada's young people asked AI for mental-health help. The gap answered first

THE SIGNAL IN ONE SENTENCE

The chatbot is not the whole story. The empty chair beside it is. Jack.org published its 2026 Youth Voice Report on October 9 after surveying young people in its Canadian network about generative AI and mental health. Among the 395 eligible respondents who completed the survey and said they had used generative AI for mental-health purposes, roughly one in three said they did so at least monthly. Three in five had used it for advice about a specific situation. Almost half, 49.9 percent, said they had not told anyone about that use. Those are striking numbers. They are also easy to turn into a bad headline. This was not a random sample of every young person in Canada. Respondents were recruited through Jack.org's network of youth mental-health advocates. To enter the main survey group, they had to report already using generative AI for mental-health purposes. The report says plainly that its findings identify patterns within this group, not population-level estimates. So the honest sentence is not, "One in three Canadian youths use AI for mental health." The honest sentence is, "Among 395 eligible respondents in Jack.org's network who reported using AI for mental-health purposes, one in three used it monthly or more." Less catchy. Much more useful. The plain signal is this: young people are not choosing a machine in a vacuum. They are choosing an immediate, private and nonjudgmental doorway when a human doorway feels closed, expensive, slow, embarrassing or unavailable. That does not make generative AI a therapist. It makes its appeal diagnostic. The report's strongest finding is not that young people prefer AI. They generally did not. Half rated peer support as an effective source of mental-health support. Forty-seven percent did the same for mental-health professionals, and 43 percent for trusted adults. Respondents who had used professional services rated human professional support more highly and AI support less highly than respondents without that experience. Young people described AI as a situational tool. Social relationships were the most common specific use, followed by academic challenges and self-care. Some treated it like search. Some used it like a journal. A smaller group described a conversation that felt more relational. The appeal was often the shape of the interaction: always available, no appointment, no visible flinch, no bill, no awkward opening sentence. That last point matters. Nearly half of respondents kept their AI use to themselves. Among those who disclosed, friends were the most common audience. Mental-health professionals and teachers were much less common. Reasons for staying quiet included privacy, fear of judgment, embarrassment and uncertainty about how others would react. There are therefore two possible adult responses. The first is to announce that chatbot use is foolish and banish the phone. That can turn a technology risk into a disclosure risk. A young person who expects punishment or ridicule has one more reason not to say what they asked, what the system answered or what support they still need. The second is curiosity with boundaries. Start with: What did you want help with? What made this easier than asking a person? Did the answer make you feel calmer, more frightened or pushed toward anything risky? Did you share identifying information? Would you like help checking the advice or talking with somebody who can actually know you? That conversation does not endorse every chatbot answer. It creates a route out of a private loop. The route matters because generative AI can be confidently wrong, flattering in dangerous ways and blind to the context that a person notices. It may validate a harmful belief because agreement looks like support. It may misunderstand sarcasm, coercion, abuse, medication, eating behavior or immediate danger. It can remember or transmit sensitive details under terms a young user has not read. A model may produce the correct warning in a test and still fail in the messy wording of an actual crisis. Jack.org's respondents recognized many of those limits. The report describes concerns about inaccurate information, privacy, overreliance, bias and AI reinforcing harmful thoughts or behavior. Respondents were less comfortable using AI in a crisis than for an everyday struggle. Many described it as a starting point, not a replacement for real people. The report itself offers the right policy clue. When a smaller subset of 73 respondents ranked government priorities, protecting young people's data and privacy came first. They also wanted practical knowledge about data rights, emotional risks, accuracy and when to move from AI to real-world support. Banning AI for people under 16 ranked last among the supplied options. That is not a referendum on one law. The subgroup is small and drawn from the same network sample. It is still a useful design brief: reduce the reasons for secrecy, make the risks legible and build a human handoff that works before somebody is in acute danger. For families, the handoff can be simple. Agree that asking an AI question will not trigger automatic punishment. Review sensitive answers together without making the young person defend every word. Decide which topics always require another person, such as immediate safety, self-harm, medication changes, abuse, threats, severe symptoms or a plan that could cause harm. Put trusted names and numbers somewhere easier to find than the chatbot history. For schools, do not tuck this into a generic plagiarism assembly. Mental-health use is different from homework use. Train counselors, teachers and administrators to ask about AI without shaming the student. Give staff a short protocol for privacy, documentation and escalation. Explain clearly what remains confidential and what must be shared when safety is at risk. For youth services, make the first contact less punishing. Offer text, chat, phone and in-person routes where possible. Publish wait times honestly. Let a young person start with a small question. If intake forms, office hours and costs make a chatbot the only door open at midnight, the service design is part of the technology story. For AI developers, a cheerful disclaimer is not a safety system. Mental-health interactions need age-appropriate language, strong privacy defaults, easy deletion, conservative memory, testing with indirect crisis language, and escalation that does more than paste a hotline into every difficult conversation. The system should make its limits visible and help the user reach a person without pretending it completed a clinical assessment. Measure the handoff, not just the warning. Did the user understand why the system could not safely continue? Could they contact a service in their location? Did the model preserve agency while responding to immediate danger? Did it avoid inventing a diagnosis? Did the service collect less sensitive data after the risk became clear? Did a human become reachable? Governments have the wider job. Privacy rules and age protections matter, but they will not fix the access gap that makes instant machine attention attractive. A safe policy has two columns. One covers product duties: data minimization, truthful claims, independent safety testing, incident reporting and enforceable protections for minors. The other covers human capacity: affordable care, school counselors, culturally appropriate support, rural access, disability access, peer programs and services outside ordinary business hours. Leave out the second column and the policy scolds young people for noticing the obvious. The machine answered. The report should not be romanticized. Its participants came through a mental-health network and may be more aware of these issues than the broader population. The sample cannot tell Canada how common this behavior is overall. Self-reported use can be remembered or interpreted differently. Generative AI changes quickly, so a snapshot can age before the ink dries. It should not be dismissed either. Mixed methods let respondents describe the experience behind the count. Young people helped shape the questions and analyze anonymized material. The result is not a national prevalence survey. It is a detailed map of why a particular group used a tool that many adults discuss without them. One path through that map is clear: A young person has a difficult thought. Human support feels unavailable, costly or judgmental. AI offers instant private language. The response may help organize the thought, or it may be inaccurate, reinforcing or unsafe. Disclosure determines whether anybody else can check. A calm adult or service can turn the private exchange into a human next step. Every break in that chain is a design decision. Canada's official 9-8-8 Suicide Crisis Helpline is available by call or text at 9-8-8, 24 hours a day. Its site says that if someone's safety is at immediate risk, call 9-1-1 or go to the nearest emergency department. A generative AI system is not an emergency service. For everybody else, the question is not whether a young person should ever type a hard feeling into a machine. That debate arrived after the behavior. The useful question is what happens next. If the answer is a lecture, the conversation may close. If the answer is open curiosity, a privacy check, a reality check and a reachable human, the chatbot can become a doorway instead of a room with no exit. Young people did not ask adults to pretend the risks are harmless. They asked adults to understand the need that got there first. That is the gap talking. We should probably listen.