AI youth policy framework calls for adult-only rules on AI companions

The Flourishing Generation whitepaper proposes pre-market and post-market safety testing for child-facing generative AI products, alongside age-appropriate design and research exemptions.

The Flourishing Generation whitepaper sets out a model policy framework for children, youth and generative AI.

Noēsis Collaborative and the Leverhulme Centre for the Future of Intelligence at the University of Cambridge have published a whitepaper calling for stronger policy controls on generative AI products used by children and young people.

The Flourishing Generation: A Whole-of-Society Strategy for Children, Youth, and AI sets out a model policy framework for lawmakers, developers and civil society groups working on child-facing AI systems.

The paper was co-authored by Ron Ivey, Founder and CEO of Noēsis Collaborative, and Henry Shevlin, a philosopher and AI ethicist at Google DeepMind, Associate Director at the Leverhulme Centre for the Future of Intelligence at the University of Cambridge and Program Director of the Kinds of Intelligence Program.

The framework focuses on generative AI that is conversational, human-like in behavior and social in function. It covers AI in toys, gaming, mental health, assistants, social media and AI companions, as well as general-purpose chatbots such as ChatGPT, Claude and Gemini.

Shevlin shared the project on LinkedIn, writing: “How should we design AI policy for children and young people if the goal were flourishing rather than just harm-minimisation?”

Framework targets AI companions and human-like chatbots

The strongest policy recommendation is to treat some AI companions and human-like chatbots as adult products by default. Shevlin described the framework as including “pharma-style pre- and post-market testing for children's AI products, and treating AI companions and chatbots that mimic rich human inner lives as adult products by default.”

The whitepaper’s model policy framework calls for generative AI systems to be held to pre-market and post-market safety audits similar to those used for high-stakes consumer goods such as pharmaceuticals, toys and car seats.

Those audits would test whether child-facing AI products support healthy development or, at minimum, do not degrade a child or young person’s capability development before minors are given market access.

The framework would also restrict to adults only conversational generative AI products that mimic “a rich human-like inner life,” including emotions, internal states and motivations, where those products encourage emotional dependence.

A second adult-only restriction would cover social AI products that primarily function as companions or are specifically designed, marketed or optimized to form ongoing social or emotional bonds with users.

The paper does not argue for a blanket ban on children using AI. Shevlin described its approach as “graduated rather than prohibitionist,” with car seats and film ratings used as the model. The framework allows for age-appropriate product versions, privacy-preserving age assurance, and supervised therapeutic and educational exemptions where evidence supports benefit.

Report cites risks but avoids a simple harm-only frame

The whitepaper argues that child and youth AI policy should not be limited to reducing harm. Its design paradigm states: “We believe technologies should be tools to enhance, develop, and help integrate a young person’s full range of human capabilities to achieve their aspirational goals in the communities and contexts where they live.”

That positive framing sits alongside a detailed risk map. The paper identifies potential harms including child abuse and exploitation, negative mental health impacts, self-harm, suicide and harm to others, data collection without meaningful consent, addictive design, psychologically targeted advertising, emotional dependency, social isolation, disruption of relational learning and developmental stunting.

The authors are careful about the evidence base. The paper says the rapid pace of generative AI development and the lack of large-scale, high-quality social science research make definitive assessments premature.

At the same time, the whitepaper says children and young people are already using AI systems in educational, play and social contexts. It cites one study finding that 72% of U.S. teens have used AI companions and 24% have shared personal information with them. It also cites another study finding that 75% of European children aged 9 to 16 have used generative AI.

The education angle is not treated as a separate safe zone. The paper says risks can arise even when children first use generative AI tools for educational purposes, particularly when products can engage in emotional relationships.

Shevlin’s own concern is focused on influence and moral development: “Another major concern is the role these systems may play in shaping opinions and moral development. Single conversations can have outsized impacts on teenagers because of the neuroplasticity of the adolescent brain. If a young person has a close relationship with a social AI system, the influence of that system could become extremely significant. That creates a serious responsibility for developers around the values embedded in these systems.”

Researchers call for audits, standards and youth voice

The whitepaper was developed from a Gen AI and Youth Policy Workshop hosted at Jesus College, University of Cambridge, in October 2025, and from a Gen AI and Youth Working Group launched at the event.

Contributors named in the paper include Nathanael Fast of the USC Marshall School of Business, Amina Fazlullah of Common Sense Media, Sam Hiner of the Young People’s Alliance, Ravi Iyer of the USC Neely Center and Psychology of Technology Institute, Thao Ha of Arizona State University and Timothy Lomas of the Harvard Human Flourishing Program.

The model policy framework draws on listening sessions and focus group discussions with children and young people, expert input, conversations with policymakers and academic research.

The paper also points to AI companies that have expanded parental controls, strengthened moderation of sexual and harmful content, improved reporting mechanisms and acknowledged emotional risks in research. It says gaps remain around minors’ access to AI companions, human-like conversational design, emotional attachment, age assurance and psychological impact evaluation.

The recommended whole-of-society strategy calls for impact audits, age assurance and safety-by-design standards for child and youth-facing generative AI products. It also calls for independent longitudinal research, data access, independent benchmarking, public and private monitoring capacity, youth and family input in governance, plain-language disclosures, and multi-stakeholder work on positive visions for children’s AI use.

The paper names Omidyar Network, Einhorn Collaborative, Lumina Foundation, Workday Foundation and Risman Foundation as funders of the Gen AI and Youth Policy Workshop.

The executive brief and full whitepaper are available now. Shevlin wrote on LinkedIn: “I'd welcome thoughts, especially from those working in education, child development, and AI policy.”