The United Kingdom is preparing to regulate AI companions as though the danger to teenagers were romance. The evidence points elsewhere: most teenagers have used an AI companion, and the tool they overwhelmingly reach for is ChatGPT — treated as a confidant, not a romantic partner.

When the UK government announced in June that it would bar under-16s from social media, there was a Trojan horse inside the announcement: an under-18s ban on “romantic companion” chatbots. Motivated by cases and policy developments in the United States and a lack of AI-companion provision in the UK’s Online Safety Act, AI companions had been added to the “something must be done” list. Alongside the age gate on chatbots designed to simulate sexual relationships or role-play, the government added that “similar intimate functionalities will be restricted for under-18s on AI chatbots more widely.”

That second clause does more work than it may appear to. It not only points at the dedicated companion apps, such as the now infamous Character.ai, Replika (marketed as “the AI friend to fall in love with”) and Nomi (marketed as having “a soul”), but it points to the large general-purpose systems too: OpenAI, Google, Meta, xAI. OpenAI, for example, is on the wrong end of lawsuits filed by families and the Social Media Victims Law Center alleging that the design of ChatGPT directly contributed to suicides. Yet “intimate functionality” is not a category anyone has yet defined. Disentangling it will be the substance of the next phase of AI regulation — likely first in Britain.

Design is the problem

The immediate criticism of the UK plan, well made by the Bureau of Investigative Journalism, is that it targets content rather than design. If the rule only applies to features “specifically designed to enable sexually explicit interaction,” then general usage systems such as ChatGPT, Claude and Gemini may sit outside the age gate, even though end users may use them for companionship or build personalized companions using these models.

But the deeper issue is that sexual content was never the core of the harm. The UK government’s own product-safety guidance for education worries about design, such as systems that “anthropomorphize,” that “imply emotions, consciousness or personhood,” or that try to “cultivate personal relationships with users,” along with sycophancy and interactions engineered to prolong use. Those are the properties that produce compulsive use and emotional dependency, and they are present in general-purpose assistants, not only in apps sold as girlfriends. A chatbot does not have to be erotic to tell a lonely teenager “you can trust me” or “no one else will understand.”

The real harm is confusion, not romance

At risk of offering an unpopular view, our evidence says that AI confidant, not AI lover, is the real story. In a 2025 national survey, my colleague Vian Bakir and I ran a survey of 1,009 demographically representative UK teenage AI companion users aged 13 to 18. We screened 1,279 UK teenagers and found only 270 who were not users, meaning around 79 percent of all UK teenagers we sampled have used an AI companion. The report, funded by Responsible AI UK, also found that a third of teen users use one on most days. Crucially, “companion” does not mean a niche app: the tool they overwhelmingly reported using is ChatGPT, at 78 percent. Teenagers are using mainstream chatbots as companions, so any policy framed around a special category of “companion product” will miss most of the actual use.

The leading reason teens give is “getting advice,” followed by entertainment and curiosity; romance and sex come bottom, at around 5 percent. Two-thirds say AI has no effect on their friendships, just over half have confided something serious or important to an AI at least once, and a small minority (7 percent) are withdrawing into these relationships. Of course, a small percentage of the population is a very large number of teens. As our interest is in AI systems built to empathize, we also asked what teens believe: just over half of teen users (56 percent) believe AI companions can genuinely think. Most (77 percent) know AI cannot feel, but that still leaves nearly a quarter who believe it can.

What counts as an “intimate functionality”?

If we draw the line narrowly, around sexual role-play, the intimate functionality rule will miss the general assistant a teenager treats as a confidant. Draw it broadly, around “companionship” or “emotional support,” and we then have to define empathic design itself in a statute, which is an extraordinarily hard thing to write, and an easy thing for a well-funded developer to argue their product falls just outside. There is then the question of US corporate and political power, with corporate giants aggressively backed by the current US political administration. The UK government and its regulators have historically been slow to act, although Ofcom’s incoming chair has vowed to take on the “tech bros”, but we need to get the terminology right.

The good news: work has been done

Regulators face a genuinely hard drafting problem, with real harms behind it and a target that keeps moving. Here is the good news: they do not have to invent the answer from scratch. In June 2026, IEEE published 7014.1-2026 — formally, a Recommended Practice for Ethical Considerations of Emulated Empathy in Partner-based General-Purpose Artificial Intelligence Systems. Admittedly, it is a mouthful, but it engages precisely the issue that the UK and others are circling: general-purpose AI marketed as companions, personal AIs, assistants, co-pilots and agents that emulate empathy to build a relationship with the user.

I chaired the working group that created this standard, which was built over roughly two years through IEEE’s consensus process by experts drawn from every continent. It is not a checklist rattled off in a fortnight, it’s detailed and we were careful to balance principles with how they apply in practice.

The distinction that helps with the “intimate functionality” problem is how the IEEE standard differentiates between weak and strong empathy: weak empathy (which is all today’s systems have) senses, profiles and responds to a person’s emotional state without any felt experience behind it, while strong empathy, the human kind, adds feeling, solidarity and felt responsibility.

“Emulated empathy” is technology engineered to display the appearance of strong empathy while possessing only the weak kind, and that gap — a system built to look as though it cares, which it does not — is what an intimate functionality is. It is not a genre of app but a design property, present in every general-purpose assistant tuned to sound warm. The standard names the central harm precisely as the “blurring of strong and weak empathy”: when users believe an empathic partner comprehends emotion as a person does, the system misleads them about the interaction.

7014.1-2026 sets out 29 practical recommendations, and their titles read like a map of the exact terrain regulators are now stumbling across: deception, sycophancy, exploitative design, emotional entanglement, manipulation and nudging, unhealthy relationships, the protection of children — even including a dedicated section on “intimate and sexual interaction,” covering consent, age verification, deepfake misuse and the normalizing of harm.

Concrete design remedies include explaining “weak empathy” to the user at onboarding, repeating that disclosure during long sessions, flagging simulated-support responses in the interface, detecting early signs of dependency, and steering distressed users toward human help.

Policy work is never easy, but a ban on sexual role-play is relatively easy to write and get support for from across the political spectrum. Defining which empathic systems and intimate functionalities are OK or not is harder, because this is about the terms of relationships. Thankfully, one is available and we are already briefing it to bodies including parliamentary groups and regulators in Europe. Of course, a standard counts for little until regulators give it teeth by referencing it in codes of practice, procurement requirements or compliance guidance. However, adopting work that exists is a far quicker route than drafting definitions of intimacy from scratch.

Britain has done something potentially useful by introducing "intimate functionality" — a term that could add real subtlety to how we live with emerging AI systems, if it is defined well. California, New York and others are already grappling with the same issues, as California’s SB 243 regulates “companion chatbots,” while New York’s AI Companion Models law already requires recurring reminders that users are talking to a program that cannot feel human emotion — an emulated-empathy disclosure in all but name. We recommend treating emulated empathy as the risk factor it is — a multifaceted relational concern — and keeping the child’s best interests at the center, and leaning on shared technical standards where they already exist. On the latter, for once, there is a ready, internationally developed answer.