The UK plans to ban romantic AI chatbots for under-18s.

On June 15, UK Prime Minister Keir Starmer revealed the government’s intentions to ban under-16s from accessing social media sites, like Facebook, Instagram, and TikTok. As part of this policy, "so-called AI 'romantic companion' chatbots – designed to simulate sexual relationships or roleplay with users" will be restricted to ages 18+.

As researchers who specialise in young people's digital well-being and mental health, we welcome the government's recognition that young people face distinct risks online, and that AI systems—in addition to social media platforms—warrant targeted regulatory attention. The new minimum age of 18 for romantic and sexually simulative AI chatbots is a reasonable step, but it addresses only a narrow slice of the risks young people themselves identify as most pressing.

What young people actually think

Our research in the SHIFT-AI project was designed to address a gap that runs through most AI safety work: the frameworks intended to protect young people have largely been built without their input. Across a series of sessions this past spring, members of the NeurOx Young People's Advisory Group (aged 16–24) worked alongside our research team as co-researchers—not consultees—identifying, evaluating, and stress-testing adolescent AI risks and mitigation strategies.

These conversations with young people cast doubt on the efficacy of the announced policy. AI companionship and romance were seen by our co-researchers as potentially harmful but unlikely to occur in their own lives or those of their peers. Several risks that dominate adult-led AI safety discourse, including parasocial attachment to AI systems, were rejected outright as failing to reflect how young people actually experience AI in day-to-day life.

By contrast, our co-researchers consistently flagged high-likelihood, high-impact risks that the announced policy does not address. These include over-reliance on AI for emotional support and mental health guidance, unwarranted trust in expert-sounding AI responses, and cognitive de-skilling through habitual AI use. Perhaps most strikingly, our co-researchers identified a vulnerability not captured in existing frameworks: Many described turning to AI specifically because it feels like a lower-stakes alternative to the adults in their lives—teachers, parents, clinicians—when they need guidance or want to disclose something difficult. Critically, these are not findings about romantic attachment, and they point to risks faced by adolescents that the current proposed policy does not touch.

Age and agency

We are also concerned about how age thresholds are being applied in this policy. In the United Kingdom, 16-year-olds may consent to sex and to their own medical treatment. Under the newly proposed policy, however, these same young people would be prohibited from engaging in simulated romantic or sexual interaction with chatbots. On its face, this is puzzling. A 17-year-old may legally have sex with another person, but not roleplay romance with a chatbot. Given that there is no conclusive evidence that romantic interaction with AIs exceeds the risks young people are already trusted to navigate at this age, why might policymakers draw the line at 18?

The implicit logic, we think, runs as follows: because the Online Safety Act already restricts under-18s from accessing harmful online content, including pornography, romantic AI interaction is being treated as legally closer to pornography than to consensual sex. At present, we set aside whether this equivalence is defensible—but we do call for making it clear why certain cut-offs are being applied, and what evidence base supports them.

This matters beyond the case of romantic chatbots. Age thresholds applied without a clear evidence base risk signalling concern rather than targeting harm, treating young people as objects of protection rather than agents whose competence and perspectives deserve serious weight.

An alternative path to keeping young people safe

Rather than restricting access to generative AI (whether in a romantic context or wholesale), regulation should be building the conditions for competent and critical engagement. This includes investing in AI literacy across schools and care settings, and requiring developers to design systems that support rather than undermine adolescent agency. In practice, this means scrutinising the features that make AI systems potentially harmful for young users regardless of their content: the design choices that encourage emotional disclosure to conversational agents, that present AI outputs as authoritative, or that maximise engagement at the expense of critical reflection.

Bans and age gates risk removing agency from young people without creating incentives for better design. Our recommendations, grounded in co-production with young people, are twofold: First, that the government should prioritise restrictions on the anthropomorphic and engagement-maximising design features of AI systems used by under-16s, regardless of whether those systems involve romantic functionality; and second, that meaningful youth co-production should be incentivised at every stage of AI development and safety evaluation. Together, these would address a far wider range of potential harms, and do so in a way that reflects the actual landscape of risk as young people experience it.

We are continuing to conduct research with young people to build the evidence base that adolescent AI safety requires, and we welcome collaboration from researchers, AI developers, and policymakers who share a commitment to making technology safer for people of all ages.

To learn more about our research on keeping young users of generative AI safer, see here.