When young people discuss mental health questions with AI

What the JAMA study examined

McBain and colleagues surveyed a total of 1,009 adolescents and young adults in the US in November 2025. The participants were between 12 and 21 years old. Population weighting was intended to make the sample representative of approximately 42.8 million people in this age group in the United States.

Participants were asked whether they had ever asked an AI chatbot for advice on mental health topics. For those who had used such systems, the study also recorded frequency, perceived helpfulness, and disclosure to other people. In addition, the survey assessed whether respondents had spoken with a physician about mental health needs in the preceding six months.

This is a representative cross-sectional survey. It describes self-reported use at a specific point in time. The study did not observe conversation trajectories, did not assess the quality of individual responses, and did not examine clinical effects. This limitation is important because prevalence, subjective helpfulness, and actual benefit are different questions.

19.2 percent – not 13.1 percent

In the current survey, 19.2 percent reported having used an AI chatbot for mental health advice at least once. Extrapolated, that corresponds to approximately 8.2 million adolescents and young adults in the United States. This means that usage is not a rare fringe phenomenon, even though “at least once” does not yet say anything about intensity or duration.

The figure of 13.1 percent comes from a similar survey by the research team conducted in 2024. In the publication, it is cited as a comparative value. Because the wording differed slightly and because the same individuals were not observed over time, the comparison is not a classic longitudinal study. Nevertheless, the authors note an increase by nearly half.

Separating the two survey years is more than a statistical subtlety. Anyone who presents 13.1 percent as the current figure for 2025 underestimates the prevalence reported in the study. Conversely, anyone who derives an exactly measured individual increase from the difference claims more than the design can demonstrate.

Frequency and perceived helpfulness

Among those who had used AI for mental health advice, 42.8 percent did so at least monthly. Relative to all respondents, that corresponds to approximately one in twelve people. The study thus shows a group for whom such conversations are a recurring part of everyday life, alongside occasional experimentation.

91.7 percent of users rated the advice they received as at least somewhat or very helpful. This figure describes the respondents’ perception. It does not demonstrate medical correctness or an improvement in well-being. Even a response that reassures, agrees, or is immediately available can seem subjectively helpful without being reliably accurate in content.

This distinction is particularly crucial for product development. A high satisfaction rate can show that an offering meets a need. But it must not be used as the sole evidence of efficacy. For that, controlled studies, independent quality assessments, and longer-term observations would be necessary.

Most told no one about it

63.3 percent of users had not told anyone that they had used an AI chatbot for mental health advice. Extrapolated, that corresponds to approximately five million people. Of those who disclosed their use, 28 percent named a friend; 16.4 percent reported it to a trusted adult such as a parent, teacher, or medical professional.

Not every failure to disclose indicates shame, secrecy, or danger. Some conversations may have been casual. Other people may fundamentally prefer to keep private thoughts to themselves. The survey asks about behavior, not reasons. Nevertheless, the high proportion shows that adults and professionals are unaware of a substantial portion of this use.

This does not imply a mandate for surveillance. More sensible are opportunities for conversation without blame: Which systems are being used? What was helpful about them? Were there answers that exerted pressure, caused uncertainty, or asserted a diagnosis? An open inquiry can strengthen media literacy without devaluing the low-threshold need for a private conversation.

Who reported use more often?

In the adjusted statistical models, female respondents reported use more often than male respondents. The odds ratio was 2.10. Among 18- to 21-year-olds, reported use was also higher compared with 12- to 14-year-olds; here the odds ratio was 3.65.

People who had spoken with a physician about mental health needs in the past six months also reported chatbot use more often. The odds ratio was 1.89. This may mean that people with an existing need for support are trying out additional information or conversation pathways. The cross-sectional data cannot clarify the direction of this association.

Odds ratios are not direct percentage differences and are not statements about individual people. They must also not be misinterpreted as fixed characteristics of gender or age. The results refer to this U.S. sample, the questions used, and the survey time point in November 2025.

What "mental health advice" means in the study

The research question was deliberately broader than psychotherapy. A conversation about stress, sadness, relationships, or coping with a difficult situation can count as mental health advice without constituting treatment. The study therefore does not reliably distinguish between a brief need for information, emotional relief, and repeated use that has a therapeutic effect.

Likewise, general chatbots and specially developed mental health offerings were not tested as equivalent interventions. The category describes a usage practice: young people turn to AI with mental health questions. Which models, prompts, data protection conditions, and conversational roles underlay this remain an open question when assessing individual products.

The authors compare the scale of use with the proportion of young people who received counseling from a mental health professional. They themselves emphasize that the two measures are not equivalent. Formal care and an arbitrary chatbot request fulfill different tasks, are subject to different responsibilities, and must not be offset against each other.

Why the US figures do not simply apply to Germany

The survey is representative of the United States, not of Germany. Access to care, costs, language, awareness of individual services, and cultural expectations differ. It would therefore be wrong to calculate a German user figure directly from the 19.2 percent.

For the German-speaking region, the study nevertheless provides an important research question. If comparable offerings are accessible with low barriers, their actual use must be examined: by age, reason, frequency, form of conversation, and possible supplementation or displacement of human contacts. This requires its own representative data.

Age groups should also not be lumped together prematurely. A twelve-year-old has different developmental, consent, and protection needs than a 21-year-old. Product rules, language, and data collection must take these differences into account, rather than treating “young people” as a homogeneous target group.

Conclusions for conversational offerings

A service should clearly show at first contact that the responses come from an AI system. This information must be understandable and must not appear only in a long legal text. At the same time, the labeling should not be turned into a wall of warnings that makes actual access to the conversation impossible.

Data minimization is particularly important because personal chat texts can contain sensitive information about relationships, feelings, everyday life, and health. A guest mode can lower the barrier to entry, but it must not be associated with false promises. What remains local, what is transmitted to a model provider, how long data is stored, and how it can be deleted must be explained in an understandable way.

Conversation quality requires its own checks. The system should not reflexively diagnose, immediately give advice under any strain, or confuse agreement with understanding. Equally important are corrections, a no that is respected, and a clear separation between emotional support, information, and professional treatment.

• Show AI labeling clearly and understandably at first contact.
• Use age- and situation-appropriate language instead of a single standard address.
• Do not present subjective helpfulness as evidence of correctness or efficacy.
• Explain storage, model provider, and deletion in simple language.
• Investigate usage and errors only with clear consent and in a data-minimizing manner.

The APA Recommendations as Supplementary Context

The American Psychological Association has published its own health advisories on generative chatbots, wellness applications, and adolescent well-being. These documents are professional recommendations, not an additional efficacy study. They emphasize that general AI chatbots should not be treated as a substitute for qualified mental health care.

For adolescents, the APA calls for age-appropriate protective mechanisms, independent research, and comprehensible AI literacy. At the same time, it does not describe AI categorically as good or bad. Risks depend on developmental stage, product design, usage context, and existing vulnerabilities.

This differentiated stance aligns with the JAMA data. The usage is already a reality. The practical question is therefore not only whether young people should talk to AI. What matters is what expectations a product creates, what data it processes, how it makes errors visible, and whether people can continue to autonomously access other forms of support.

What Research Is Now Missing

First, comparable data for Germany and other European countries are needed. Representative surveys could capture usage occasions, products used, conversation frequency, and disclosure. In doing so, well-being, mental health treatment, everyday self-reflection, and pure curiosity should be kept separate.

In addition, qualitative research is needed. Why do some young people tell no one about their usage? What do they experience as helpful, patronizing, or unpleasant? Which responses change decisions or relationships? Such questions cannot be answered from usage numbers or automated chat ratings alone.

Finally, longer-term trajectories are needed. A cross-sectional study cannot show whether regular use supplements human conversations, replaces them, or—depending on the situation—does both. But it can show that these questions are no longer theoretical. Millions of young people have already incorporated the technology into their personal information and conversation routines.