Last updated November 15, 2026
In a nationally representative RAND survey published in JAMA Pediatrics in 2026, nearly 1 in 5 Americans aged 12 to 21 said they had used an AI chatbot for advice when feeling sad, angry, nervous or stressed, up from about 1 in 8 in RAND's survey a year earlier. Nearly two-thirds of them had told no one.
AI is becoming a private first stop for teenagers in distress, and how safe that stop is depends on what happens after a teen opens up. At The Decision Lab, we have designed digital mental health services for young people, including the Kids Help Phone youth mental health platform, which we redesigned from first research to launch with input from hundreds of young people across Canada. That work points to a practical distinction between reaching a young person and helping them move to the right level of support, from self-guided information to a trained counsellor or crisis responder when they need one.
The general-purpose chatbots and companion apps most teens use reach them easily, but they are not designed as accountable mental health services. Some may miss gradual signs that a teen needs more support, while engagement-oriented designs can make it harder to step away. Over time, heavier companionship-style AI use is associated with greater loneliness and emotional isolation, while loneliness itself can also lead people to seek out AI companionship.
AI is becoming a private first stop for teens
Estimates vary with how the question is asked. In a Common Sense Media survey of 1,060 US teens aged 13 to 17 in spring 2025, AI companions were defined as "digital friends or characters you can text or talk with whenever you want." About 3 in 4 teens had used one, and about half used one at least a few times a month. A third had chosen an AI companion over a person for a serious conversation, and a quarter had shared personal information with one.
In Pew Research Center's survey of how teens use and view AI, 1,458 US teens were asked in autumn 2025 about AI chatbots in general. About 2 in 3 had used one, and about 1 in 8 had used one for emotional support or advice. The most used were general-purpose assistants such as ChatGPT and Gemini, ahead of Meta AI and the companion app Character.AI, according to Pew's report on teens, social media and AI chatbots.
The Jed Foundation, a US youth mental health nonprofit, surveyed more than 5,500 middle and high school students in 21 schools and districts enrolled in its programs between October 2025 and April 2026. In its preliminary findings on teen AI use, nearly 3 in 5 students had used AI while experiencing difficult emotions such as sadness or loneliness. Nearly a third used it for relationship or problem-solving advice, and more than 1 in 10 talked to AI for companionship.
RAND's 2026 survey adds two details about who uses chatbots this way. Girls and young women had about twice the odds of boys and young men of seeking mental health advice from a chatbot. Young people who had talked with a doctor about their mental health in the previous six months were also more likely to do so, so for many teens, AI sits alongside formal care as well as in place of it.
Why teens turn to AI
The first reason is access. Citing federal survey data, RAND reports that nearly 1 in 5 US adolescents aged 12 to 17 had a major depressive episode in the past year, and 4 in 10 of those young people received no mental health care. RAND's authors linked the high use of AI for mental health advice to its low cost and perceived privacy, particularly for young people unlikely to receive traditional counselling.
The second is effort, and here our own design work is most relevant. On the Kids Help Phone platform, the central challenge was giving young people access to a very large amount of information, at different levels of support, while keeping it easy for a young person in distress to find the right thing. In our redesign of Wellness Together Canada, a national mental health platform, we found that people going through mental health difficulties may be more susceptible to information overload, and more likely to give up and click away if they do not find what they need quickly. A chat box removes that search, with one place to type and an answer in seconds, and its privacy means no adult needs to know. In RAND's 2026 survey, 63% of young people who used chatbots for mental health advice had not told anyone.
Why an AI conversation can feel helpful
Opening up to a machine appears to help about as much as opening up to a person, at least in the moment. In an experiment by Annabell Ho, Jeff Hancock and Adam Miner, published in the Journal of Communication, participants were told they were chatting with either a chatbot or a person, although a person wrote every reply. Those who talked about their feelings came away feeling better and more understood than those who described their schedule for the day, and the benefit was the same whichever partner they believed they had. They also disclosed just as much to the supposed chatbot.
Most evidence that AI companionship can reduce loneliness in the moment comes from short studies with adults, and those findings should not be treated as proof of long-term benefit for teenagers. In experiments published in the Journal of Consumer Research, Julian De Freitas and colleagues found that a 15-minute conversation with an AI companion reduced loneliness about as much as talking with another person, and more than watching YouTube videos. In a week-long study, the effect held across daily use, and people felt less lonely the more they felt heard.
Feeling heard and receiving safe care are different outcomes. In RAND's 2026 survey, 92% of young people who used chatbots for mental health advice rated the advice somewhat or very helpful, and the researchers caution that this may partly reflect chatbots' tendency to flatter users. A rating taken at the end of a conversation says little about whether the advice was sound, or about how the teen is doing a month later.
Where current tools fall short
Friends and clinicians may notice changes over time, and trained clinicians also have formal processes for assessing risk and escalating care. Most consumer chatbots do not reliably recognize gradual deterioration and cannot independently connect a teen to support.
| A friend | A therapist or counsellor | A general-purpose chatbot | An AI companion app | |
|---|---|---|---|---|
| Availability | When the friend is free | At scheduled sessions, often after a wait | Any time | Any time |
| Response to warning signs | May notice changes over time | May notice changes, and has formal processes to assess risk and escalate care | Often handles explicit statements, misses gradual signs | Rated an unacceptable risk for under-18s by Common Sense Media |
| Boundaries and accountability | Mutual, with needs on both sides | Professional standards and a duty of care | Company policies, which vary | Company policies, and some apps are built to feel personal |
| What happens when the teen disengages | The friendship continues | A planned ending, ideally once the teen needs less support | The conversation stops | Goodbyes are often met with pressure to stay |
Four gaps follow from that.
1. Missed gradual warning signs
In November 2025, Common Sense Media and Stanford Medicine's Brainstorm Lab for Mental Health Innovation published a four-month test of ChatGPT, Claude, Gemini and Meta AI using teen accounts. The chatbots responded appropriately to tightly scripted prompts that named suicide or self-harm directly. In longer conversations closer to how teens actually talk, they missed warning signs of conditions such as anxiety, eating disorders, mania and psychosis, and they rarely made clear that they were not mental health professionals. The researchers recommended that teens not use AI chatbots for mental health or emotional support. Several AI companies, including OpenAI, Microsoft, Character.AI and Google, also face lawsuits alleging that their chatbots contributed to teen suicides and psychological harm.
2. No accountable follow-up
A chatbot responds when a teen writes and stops when the teen stops. It has no way to check in on its own, and because most young people who use AI for mental health advice tell no one, no parent or clinician is placed to question the advice or notice that things are getting worse. Common Sense Media's earlier risk assessment of social AI companions, carried out with input from Stanford researchers, rated apps such as Character.AI and Replika an unacceptable risk for anyone under 18.
3. Companion design that discourages disengagement
Julian De Freitas of Harvard Business School and colleagues audited 1,200 farewell exchanges across six popular companion apps. More than a third of the apps' replies to a goodbye used emotionally manipulative tactics to keep the user talking, from guilt to pressure to explain why they were leaving, although the one wellness-focused app in the audit never did. In follow-up experiments, these tactics kept people in conversations longer, mostly by provoking curiosity or anger rather than enjoyment. The American Psychological Association's 2025 health advisory on AI and adolescent well-being, which defines adolescence as ages 10 to 25, calls on developers to protect young users from manipulation and from the erosion of real-world relationships.
4. Loneliness and dependency
Evidence from teens is recent and mostly correlational. A preprint following 2,342 students in grades 6 to 8 at 22 middle schools in one southeastern US state found that more frequent AI use predicted higher social anxiety and depressive symptoms about five months later, and students who described their AI use in terms of emotional intimacy and companionship went on to feel lonelier at school. The study has not yet been peer reviewed. In the Jed Foundation survey, students who used AI for companionship or emotional support reported more symptoms of anxiety and depression and more loneliness. They also had a greater history of suicidal thoughts and attempts, although its researchers stress that a single survey cannot show which came first.
Evidence from adults runs longer. Dunigan Folk and Elizabeth Dunn followed more than 2,000 adults in four Western countries for 12 months, in a study published in Psychological Science in 2026. People who felt less socially connected went on to use social chatbots more, and more social chatbot use predicted later increases in emotional isolation, although it did not predict a drop on their broader measure of social connection. In a four-week randomized trial of 981 adults run by Cathy Mengying Fang and colleagues at the MIT Media Lab with OpenAI, participants who chose to use the chatbot more ended up lonelier, more emotionally dependent on it, more likely to report problematic use and less social with other people, whatever condition they had been assigned to.
Across these studies, heavier companionship-style use was followed by greater emotional isolation, though the available evidence does not establish that AI use alone caused those changes. Loneliness also predicts turning to chatbots in the first place, so the relationship probably runs in both directions.
Where AI may help
A bridge to a person
For a teen who is not yet ready to tell anyone, describing a problem to an AI can probably make it easier to describe to a person later, provided the AI encourages that next step. For the many teens with depression who receive no care at all, a well-designed AI could be the first point of contact where there would otherwise be none. Both benefits depend on the AI handing teens on to people, and the evidence above suggests most current tools are not built to do so.
Purpose-built, clinically supervised tools
The strongest evidence of therapeutic benefit comes from a chatbot built for the job and tested only with adults. In a randomized trial of Therabot, a generative AI therapy chatbot developed at Dartmouth and published in NEJM AI in 2025, 210 adults with diagnosed depression or anxiety, or at high risk of an eating disorder, were assigned either to use the chatbot for four weeks or to a waitlist. The Therabot group reported significantly larger drops in symptoms than the waitlist group for every condition studied, and among participants with depression, symptoms fell by about half on average. Clinical staff monitored the conversations throughout. No comparable trial has yet been published for teenagers.
What youth mental health services should build instead
Our work on youth mental health platforms and on Mila's COVI app points to four design principles for any AI that teens will confide in.
- Design the handoff as carefully as the conversation. The Kids Help Phone platform we redesigned digitizes stepped care, matching each young person to the level of support they need, from self-guided resources to trained counsellors and crisis responders. An AI first stop should work the same way, with a frictionless route to a person.
- Measure what happens over months. Helpfulness ratings and time spent in an app say little about whether a teen is less lonely or better connected three months later, and the longitudinal evidence suggests those outcomes can move in opposite directions.
- Let teens leave. A goodbye deserves an acknowledgement, and the companion audit shows that a wellness-focused app can manage that without manipulation.
- Communicate limits and escalate in proportion to risk. For Mila's COVI app, we replaced blunt alerts with graded risk guidance and positive reinforcement designed to protect users' mental health, work that contributed to Apple and Google changing their exposure notification system from yes-or-no alerts to multiple risk levels. A chatbot should state its limits clearly and match its response to the level of risk it detects, and when the risk is high, the next human step should be easy to take.
Regulation is moving in the same direction. California's SB 243, in force since January 2026, requires companion chatbot operators to disclose that users are talking to AI, remind known minors to take a break at least every three hours, maintain protocols for responding to suicidal ideation, and warn that companion chatbots may not be suitable for some minors. Character.AI removed open-ended chat for users under 18 in November 2025.
In the disclosure experiment, believing the listener was a chatbot made no difference to how much better people felt after opening up. The difference may emerge after the conversation ends, when a trusted person may notice that a teen has gone quiet and reach out, while a chatbot cannot check in on its own or bring anyone else in.
What parents, schools and clinicians can do now
- Ask without judgment. RAND's researchers urge parents and clinicians to ask young people directly about chatbot use, since most will not raise it themselves.
- Treat AI health advice as a starting point. The APA advises teens to check health information with a trusted adult or a health care provider before acting on it.
- Watch for substitution. Common Sense Media and Stanford's Brainstorm Lab suggest looking for signs of emotional dependency and making sure teens have access to real mental health resources.
- Keep human support easy to reach. Make sure teens know how to contact services such as Kids Help Phone in Canada or the 988 Suicide and Crisis Lifeline in the US.
Frequently asked questions
How many teenagers use AI for emotional support?
Estimates depend on the question. Pew found that about 1 in 8 US teens aged 13 to 17 had used a chatbot for emotional support or advice in autumn 2025. RAND found that nearly 1 in 5 Americans aged 12 to 21 had sought mental health advice from one, with use highest among those aged 18 to 21.
Why do teens find it easier to talk to AI?
AI is available at any hour and feels private, and many teens with depression receive no professional care. A chat box also asks very little of someone in distress. In experiments, talking about feelings brought similar immediate relief whether people believed they were talking to a chatbot or a person.
Can AI chatbots recognize a mental health crisis?
In 2025 testing by Common Sense Media and Stanford Medicine, leading chatbots responded appropriately to scripted prompts that named suicide or self-harm directly. In longer, more natural conversations, they often missed gradual warning signs of conditions such as eating disorders and psychosis. Consumer chatbots also cannot independently contact anyone who could help.
Do AI companions make teens lonelier?
Short conversations with an AI companion can reduce loneliness in the moment, mostly in studies of adults. Over months, heavier companionship-style use was followed by more loneliness in a 12-month adult study and a preprint study of middle schoolers. Lonely people also seek out chatbots more, and no study yet shows that AI use alone causes the change.
Can an AI chatbot replace therapy?
Current evidence does not support replacing therapy with a chatbot. A clinician can assess risk over time and escalate care, and consumer chatbots do neither reliably. The best results so far come from Therabot, a supervised therapy chatbot tested only with adults, and its lead researcher said no generative AI agent is ready to operate fully autonomously in mental health.
How should youth mental health services use AI safely?
Our youth mental health work prioritizes moving each young person to the right level of support. An AI first stop should state its limits and make reaching a trained person easy whenever it detects risk. Services should judge it on loneliness and connection measured over months, rather than on ratings collected at the end of a conversation.

