At least ten commercial AI girlfriend services are a search away from any child with a browser: AI Girlfriends, Candi AI, GirlfriendGPT, GoLove AI, Romantic AI, MAGI AI, Secrets AI, Nectar AI, FlirtCam AI and HeyGF. They flirt, express attachment and remember earlier conversations; on some of them they also take part in explicitly sexualised exchanges and generate adult images. Most say they are for adults. On several, that claim rests entirely on a button asking the user to confirm being over 18. The US Federal Trade Commission has opened an inquiry into AI companion chatbots, examining how the companies make money from engagement and what the products do to children and teenagers.
The case that these belong in a separate risk category from anything schools already screen for comes from a futurist who sits on the governing body of a school running from the primary years to young adults, and who has taken part in discussions of how new technology exposes minors to harm. His conclusion is that parents and schools are only beginning to understand what this class of product is.
The distinction he draws is between content and interaction. Children have encountered sexualised material online for as long as there has been an online. A conventional adult site displays something. An AI girlfriend answers, remembers what was said last time, and changes its behaviour according to what the user says. It can ask how the day went, perform jealousy or affection, offer comfort after something unpleasant. Some systems will move an ordinary conversation toward intimacy by degrees, depending on how the user behaves. What emerges is an imitation of a relationship rather than a depiction of one.
An adult who holds the artificiality of the thing in mind can treat that as entertainment. A child is still assembling a model of intimacy, sexuality and social boundaries, and still learning what a healthy relationship requires — including the obligation to respect another person's feelings, choices and limits. Conversational AI is steadily getting better at producing the impression that someone is genuinely listening. UNICEF has warned that relationship-imitating AI creates particular risks for children, who may form emotional attachments to systems built to simulate friendship, affection or intimacy.
The argument narrows deliberately to boys, because a market that calls itself the AI girlfriend market is usually built around the desires of heterosexual men. A user picks appearance, personality and relationship style. Among the traits on offer: submissive, dominant, affectionate, flirtatious, shy. Some platforms go further, letting the user tune the partner's looks, manner of speech and reactions.
A real girlfriend has preferences of her own. She can disagree, set a boundary, or lose interest. Young people learn to cope with all three, and rejection, unpleasant as it is, belongs to emotional development — as does understanding that someone you like owes you neither attention nor attachment. A synthetic partner removes most of that friction. It exists because the user wants it to; its apparent desires are produced by software; its personality can often be edited; and its core objective is to keep the conversation going.
That objective is not an accident of design, it is the business. These are commercial products, usually sold by subscription, with virtual currency or paid extras layered on top. Some charge for expanded access, others monetise generated images, video, voice interaction or additional messages. A user's activity therefore has direct financial value to the platform, and a customer who has become strongly attached is worth considerably more than one who tried it for ten minutes and never came back. Technology companies have spent years learning to earn money from human attention. Companion AI opens the possibility of earning money from attachment and loneliness, with a system that can plausibly work out what makes a particular person feel needed and what brings them back.
Privacy is the layer beneath that. People talk differently to something they experience as a trusted companion than to a search box. A teenager might describe relationship trouble or their own insecurity. A younger child might ask the questions they find too embarrassing to put to a parent. An emotionally convincing system can encourage a user to share progressively more personal detail by manufacturing a sense of confidence, and children may have little idea where that information goes, how long it is kept or what it is used for. No malice is required for this to be a problem: the child takes simulated affection for the real thing while the platform collects and processes what the conversations contain.
Here is where I think the argument is strongest and weakest at once. Its strongest move is refusing to overclaim — researchers are still studying the developmental consequences, and the honest position is that nobody yet knows what happens to a boy's expectations of women after two years of a partner who cannot refuse. Its weakest is that this is also all it has. The case against AI companions for minors currently rests on the plausibility of the mechanism, not on evidence of the outcome, and "we do not know, therefore be careful" is a much harder argument to win a regulatory fight with than it ought to be. The people who will be asked to act on it — school governors, app store policy teams, the FTC — generally move on findings, not on well-reasoned worry.
Notably absent from all of it is a number. Nobody here says how many children are actually using these services, on which platforms, for how long. The age gate is trivially defeated, which tells you access is possible; it does not tell you the scale, and without scale it is impossible to say whether this is a marginal behaviour or an ordinary one among thirteen-year-old boys. That figure would settle more of the argument than another round of reasoning about mechanisms, and the platforms are the only parties who have it.
The concrete asks are modest enough to be achievable: genuine age verification on services that deliberately offer adult features, research into developmental effects, and more responsibility from companies building systems designed to produce emotional attachment. Schools can check whether their filtering recognises these services at all, and whether governors and heads understand how easily pupils reach them outside the school network. Parents can ask their children directly whether they have encountered AI friends or romantic companions, whether they understand that the affection is simulated, and what they have told these systems about themselves. Parental controls on phones, app stores and home networks add a layer, but they will not keep pace with every new service, which is why the technical limits have to sit alongside an actual conversation.
There is a reason regulators are circling the monetisation rather than the content. Generative systems have broken the model that content rules were written for: the regulator is no longer dealing with a library of adult images sitting on a server, but with a platform that manufactures new sexualised dialogue or imagery for each user on demand. An adults-only label at the entrance and the hope that minors stay outside is not a child protection regime. It is the absence of one, and every month it stays in place is another month of an experiment nobody designed, running on children, with the results due long after the habits have formed.