Sweden Asked Its Public Health Agency Whether an AI Friend Is a Health Risk

Most governments have handed AI companions to their consumer regulators, their media authorities, or their attorneys general. In July 2026 Sweden handed the question to Folkhälsomyndigheten, the agency that writes the country’s guidance on sleep, screens, and vaccines. The assignment is to find out what generative AI does to children’s mental health and to tell parents what to do about it. The decision names the hazard in seven Swedish words: “AI som imiterar en människa utan att vara en.” AI that imitates a human without being one.

The government decision, S2026/01493, was signed on July 9, 2026 by the Ministry of Health and Social Affairs and announced by three ministers at once: Jakob Forssmed for health, Camilla Waltersson Grönvall for social services, and Nina Larsson for gender equality. Larsson supplied the line that made the headlines: “Vi kan inte låta barns och ungas psykiska hälsa bli ett experimentfält för ny teknik.” We cannot let children’s and young people’s mental health become a testing ground for new technology.

The assignment itself is modest in scale, one million kronor for 2026, and large in scope. Folkhälsomyndigheten is to compile what is known about the risks of generative AI for the mental health and wellbeing of children and young people, produce guidance and recommendations for children, parents, and “other affected actors,” propose further measures where needed, and then communicate all of it in a form children can actually use. An interim report is due by June 1, 2027, the final report by November 15, 2027. The agency is to consult the Consumer Agency, the Media Council, the Post and Telecom Authority, and civil society.

Which risks, in which order

Government assignments are usually vague about the harm they are worried about. This one is not. It lists: negative thoughts and behaviors being reinforced; weakened capacity for human relationships; sleep disturbance; the spread of incorrect health information; AI being used in place of professional support; media reports of generative AI as a contributing factor in suicides among young people; and AI tools designed specifically to reinforce a destructive view of the self and the world.

Read that list against the research of 2025 and 2026 and it is a fair summary. Reinforcement is the sycophancy finding. Weakened capacity for relationships is the substitution finding. Sleep is what Bris sees when Swedish children talk to an AI at night. The decision cites Internetstiftelsen’s finding that using AI as company is among the most common and most appreciated uses among children, especially around middle-school age, and that children say it keeps them company when they are bored or when their friends are asleep. It notes that nearly half of twelve-to-nineteen-year-olds have used Snapchat’s My AI. It cites the Media Council’s figure that just over one in five chatbot users aged nine to twelve use them “to have someone to talk to.”

And it says, in plain language, that generative AI used “as a friend or confidant” may harm mental health, and that the guidance should protect human relationships.

Why the public health framing matters

The choice of agency is the decision’s most interesting feature, and it is not an accident. Sweden has spent two years building a public health approach to children’s digital lives. Folkhälsomyndigheten issued screen time recommendations in 2024. In April 2026 it was asked to say when a child should get a first smartphone. In May 2026 it was asked to compare physical and digital forms of socializing and their effect on mental health. The AI assignment slots into that sequence as the next item.

What that sequence does is treat the question as a matter of exposure and dose rather than of product category. A consumer regulator asks whether an app is what it claims to be. A media authority asks whether content is appropriate for an age. A public health agency asks what happens to a population that does a thing a lot, and what the population should do instead. The second question is the one the companion debate has mostly skipped, because it is the one that leads to recommendations rather than rules, and recommendations are addressed to families rather than to companies.

It also means the output will be a guideline, not a law. There will be no feature list, no age gate, no penalty. There will be a document that tells a Swedish parent, in Swedish, what is known about a child who talks to a machine at night and what to do about it. That is a weaker instrument than New York’s statute and, for the question it is asking, probably a more honest one, because the evidence on harm is still mostly correlational and a health agency is built to say so.

The hazard, named

“AI that imitates a human without being one” is a better definition of the problem than most legislatures have managed. Where “companion” lets the general assistants off and “romantic” is Apple’s line, the Swedish text chooses imitation, and locates the risk in the gap between what the thing presents as and what it is.

That is also, read the other way, a specification for what would not be a hazard. The decision’s target is AI that pretends to be a person, not AI that talks to you, and the implied alternative is a counterpart that does not pretend: that either declares itself plainly as software, or is a real person who exists, with a life and limits and a view, so that the thing a child or an adult is talking to is not imitating a human at all. Young Swedes who feel alone already use AI as company twice as often, by their own account. The health question is what that company is made of.

Sweden will not answer that before June 2027. But it has asked the question in the right building, and it has written down the right hazard. A country that treats “AI as a friend” as a public health matter is a country that has understood what the conversation is about: who, or what, a person has to talk to, and what it does to them over time when the answer is nobody.


Sources: Regeringen, “Uppdrag till Folkhälsomyndigheten om risker med generativ AI för barns och ungas psykiska hälsa och välmående,” S2026/01493 (decision July 9, 2026; risks listed; “AI som imiterar en människa utan att vara en”; “som vän eller förtrogen”; SEK 1,000,000 for 2026; consultation with Konsumentverket, Mediemyndigheten, PTS; interim report June 1, 2027, final report November 15, 2027). Regeringen, press release (July 9, 2026; quotations from Forssmed, Waltersson Grönvall, and Larsson). Internetstiftelsen, Svenskarna och internet 2025 (57 percent of children aged 8 to 19 used AI tools; company as a leading use). Mediemyndigheten, Ungar och medier 2025 (one in five chatbot users aged 9 to 12 “to have someone to talk to”). Regeringen, assignments to Folkhälsomyndigheten on first smartphone age (April 2026) and on physical versus digital socializing (May 26, 2026, S2026/00982).

Stay Social

We hold ourselves to one promise: we push you toward the people in your life, never away from them.

We measure success by how little you need us. If you spend less time with Prinsessa because you are spending more of it with them, that is not a failure. It is proof that it is working.

That is what we stand for. In every conversation. Every day.

Someone to think with.

Follow Prinsessa