In 2025 about 5.4 million young people reported using generative AI tools for mental health or emotional support when they felt sad, angry, or nervous. For many young people, these conversations feel comforting and supportive. But they also come with real risks. Emotional dependency, misleading or inaccurate advice, and the inability of current AI systems to respond appropriately in crisis situations, such as self-harm or suicidal thoughts, have already led to devastating consequences. Several major AI companies now face lawsuits following teen suicides linked to harmful chatbot interactions.

At the same time, policymakers and researchers are struggling to keep up with the pace of technological change. Technology companies are asking for clearer guidance on how to better protect young users. And yet, young people themselves are often missing from the conversation.

To help bridge this gap, we convened a workshop at the Stanford Center for Youth Mental Health and Wellbeing in early November 2025, with support from Hopelab. Our guiding question was to understand how generative AI can be used to support, rather than harm, young people’s mental health.

We brought together an unusually diverse group of voices, researchers, clinicians, policymakers, tech leaders, civil society organizations, and young people, to ensure this conversation reflected real experiences, not just theory.

Participants included Stanford faculty; representatives from the California State Assembly; AI companies such as OpenAI, Anthropic, Character.AI, Google, and TikTok; and organizations including Koko, the Family Online Safety Institute, the Young People’s Alliance, The Rithm Project, and the Center for Youth and AI.

Above all, we wanted to understand what responsible AI should look like from the perspective of young people themselves, and what they believe must change.

Young participants shared the many ways teenagers and young adults are using generative AI. This includes helping them with social skills, like how to respond to a friend asking for help; dealing with professional situations (how to write a polished email) and as emotional confidantes (how to deal with anxiety). AI tools are free, always available, and feel private. For young people facing long therapy waitlists, high costs, or stigma around counseling, AI can feel like the only option. Second, many don’t feel safe opening up to friends. They worry their stories will be shared or misunderstood.

Generally, young people found conversations with AI helpful. They also raised critical questions: What is the right balance between digital and human mental health support? How can we make sure young people can benefit from AI, but do not become too emotionally dependent on AI systems? How can we protect young people without infringing on their rights or autonomy? 

Here are three recommendations that came out of the workshop. 

There was strong agreement on the need for more structural ways to involve young people. AI systems are often developed without meaningful engagement with young people, especially racial and ethnic minority youth, LGBTQ+ youth, and those from socioeconomically disadvantaged backgrounds. This is problematic because young people have the right to be heard on issues that affect their lives, and including them leads to better digital services. Young people have their own ways of communicating, distinct developmental needs, and unique digital realities. They may be able to foresee ethical issues that adults may not fully understand. 

Young people told us that safeguards should focus on positive design rather than bans. Outright restrictions are often ineffective, and young people can find their way around them. Prosocial design includes AI systems that stimulate social connection, promote healthy behaviors, and integrate AI literacy. Young people need support to critically evaluate information, understand bias in AI systems, and protect their data privacy. 

As one youth panelist put it:

The discussions also showed the clear limitations of current AI systems as therapeutic tools and the urgent need for stronger protocols around crisis situations, such as suicidality and self-harm. Can AI systems balance talk with users about difficult mental health topics, while consistently directing users to professional support? Where is the line between a helpful and harmful conversation, and when should an AI refer to other people or resources? 

Answering these questions requires better data, transparency, and shared accountability across companies, researchers, and policymakers.

We are now working on the next steps, including establishing working groups and launching new research projects. 

Stay tuned for what comes next and contact me if you want to join our efforts!