The Workshop and Who Showed Up
On November 3, King's College London and OpenAI convened a workshop, backed by the King's Institute for Artificial Intelligence, to examine how generative AI affects young people's learning, wellbeing and development. The guest list signals how seriously both organizations are treating the question: researchers, educators, clinicians, policymakers and youth organizations including UNICEF Innocenti, European Schoolnet, Parent Zone and Kooth UK, alongside academics from the University of Oxford and University College London.
The named speakers set the tone for a working session rather than a product showcase. Dr Nessa Keddo, a King's senior lecturer in media, diversity and technology, Rafaela Nicolazzi, OpenAI's EMEA privacy and consumer protection lead, and Professor Elena Simperl, co-director of the King's Institute for AI, each addressed a specific angle of the same underlying question: whether a single safety approach can actually serve the full range of teenagers using AI tools today.
The Core Finding: One Safeguard Fits No One Well
The central conclusion cuts against how most AI safety settings are currently designed. Protections calibrated for younger children unnecessarily restrict older teens who need more latitude to do real research and schoolwork, while safeguards built for older, more independent users can leave younger children genuinely under-protected. The same AI system produces meaningfully different experiences depending on a user's age, language, education level, digital skills and access to a trusted adult, which makes a single age gate a blunt instrument for a problem that is not actually binary.
Professor Simperl framed the stakes clearly: the impact of AI on young people cannot be understood by looking at the technology in isolation. That is a direct challenge to how most vendors currently validate safety features, typically against the product in a lab setting rather than against the actual variation in family circumstances and support networks that determine how a safeguard performs in practice. A safety feature that tests well in isolation can still fail the teenager who has no adult to ask for help.
Why Comfort Does Not Equal Capability
One of the workshop's sharper distinctions: being comfortable using a chatbot is a different skill from being able to judge whether its answer is correct. Spotting an error, protecting personal data from being shared inadvertently, and recognizing when a question needs a human adult rather than an AI response are each separate, teachable skills that comfort with the interface does not automatically confer. A student who uses AI fluently every day can still be no better equipped to catch a wrong answer than one who rarely touches it.
That distinction has direct implications for how schools and vendors frame AI literacy training. Teaching students to operate a chatbot is a different curriculum from teaching them to evaluate its output, and the workshop's framing suggests institutions investing only in the former are solving the easier, more visible half of the problem while leaving the harder judgment skills largely untaught. A district that reports high AI literacy scores based on tool fluency alone may be measuring comfort rather than the judgment skills the workshop identifies as the actual safeguard, and that gap will only surface later, when a student encounters a confidently wrong answer with no one around to catch it.
The Parental Monitoring Trade-Off
The group also flagged a genuine trade-off in parental monitoring tools that vendors often market as an unambiguous safety win. Monitoring can help a parent catch serious concerns early, but it can also discourage a teenager from asking the sensitive questions, about health, identity or relationships, that AI tools are otherwise well positioned to help with privately. A monitoring feature that technically works as designed can still produce a worse outcome if it pushes a teenager away from asking a question they needed answered.
The workshop's answer was to judge every safeguard by its actual effect on a child's life rather than by whether the feature functions as specified, while keeping parental tools in place. Parents, the group argued, need clear information on how a given AI system works, how a child's data is handled and which protections actually exist, information that lets families make an informed choice rather than treating any monitoring feature as self-evidently protective.
What OpenAI Is Committing To
OpenAI used the workshop to detail two concrete commitments. Its European Youth Safety Blueprint covers age-appropriate experiences, privacy-preserving age assurance, protections against manipulative outputs and parental controls, which Rafaela Nicolazzi summarized as a commitment to treat teens as teens, rather than as either small adults or large children. Separately, OpenAI's AI and Teen Development Research Grant program will fund independent studies specifically on how AI affects 13- to 17-year-olds, rather than relying only on internal research.
Dr Keddo pushed the commitment further, arguing that including young people as co-designers and participants in future research is essential, beyond treating them only as survey respondents after the fact. King's and OpenAI said they are exploring additional work through the Wellbeing Council on AI, positioning this workshop as the start of an ongoing relationship rather than a single publicity event. For institutional buyers, that ongoing structure matters more than the initial announcement, since a research grant program and a standing council give both organizations a mechanism to update their guidance as evidence accumulates, rather than leaving today's recommendations frozen while the technology and teen usage patterns keep changing underneath them.
What This Means for Edtech and AI Vendors Selling Into Schools
For any vendor selling AI products into K-12 or higher education, this workshop hands procurement reviewers a new, specific question to ask: does the age-safety design differentiate meaningfully between a 13-year-old and a 17-year-old, or does it apply one setting across an entire teen band labeled under 18. A vendor that cannot answer with specifics about how its safeguards vary by developmental stage is likely relying on the kind of blunt, uniform approach this research calls out as inadequate.
Institutional buyers should also expect this kind of evidence-based scrutiny to become a genuine competitive differentiator over the next procurement cycle, not just a compliance checkbox. Vendors who can point to independent research, rather than internal marketing claims, about how their safety features perform across different ages and family circumstances will have a real answer when a school board or university risk committee asks the question this workshop just made impossible to avoid.
