Children Already Confide in Chatbots. A New Survey Tests Whether the EU KIDS Act Draws Its Age Lines in the Right Place

The European Union is about to decide how old a child must be to use social media, and the debate has mostly been about phones, feeds and teenagers. A new survey from France suggests the harder question sits one step over, in the chatbots and AI companions that children already talk to every week, often below the age the services themselves allow.

In an opinion piece published by Tech Policy Press on November 4, 2026, Jessica Galissaire, a senior policy researcher at the Berlin-based think tank interface, argues that the Commission’s proposed EU KIDS Act treats AI chatbots and social media differently without enough evidence to justify the split, and that regulators should start from how children actually use these tools. Her conclusion is a balancing act: “regulation must preserve the opportunities generative AI can offer children, including learning, creativity and access to information, while ensuring they are not left alone to navigate potentially harmful systems.” What follows sets out her argument, checks it against the text of the proposal, and draws out what it means for companies building AI products that children can reach.

What the survey found

The data comes from a YouGov study commissioned by interface and published as Growing Up With Generative AI. Between June 24 and 30, 2026, YouGov surveyed 970 French children aged 11 to 17 and 1,138 parents of children in that range. Key figures reported in the op-ed:

• 40% of 11- and 12-year-olds say they use general-purpose generative AI at least weekly; parents estimate 26%.
• Among 11- and 12-year-olds who use generative AI, 14% say they use it to confide.
• 23% of 11- and 12-year-olds regularly use AI built into social media, and 16% regularly use AI companions, although most of these services’ terms bar users under 13.
• 89% of teenagers familiar with generative AI see at least one benefit, while 44% of parents whose children use it are not at ease with that use.
• 92% of teenage users say they spend more time than they would like on at least one tool, rising to 59% for AI companions and 57% for AI in social media, against 45% for general-purpose assistants.
• Only 32% of parents feel sufficiently supported in guiding their children’s AI use, and 42% say they get no support at all.

Galissaire is careful with her own numbers. She writes that spending more time than intended “is a subjective experience, not a clinical diagnosis,” and that a higher rate of weekly negative emotions among AI companion users (50%, against 38% for all teens surveyed) shows correlation, not cause. The survey covers France only and is self-reported. Those limits matter when the figures are used to argue for or against a specific legal rule.

Her argument, in her terms

The core of the piece is a comparison. For social media and video-sharing services with risky features, the KIDS Act proposal sets a staircase of age limits. For AI companions and general chatbots, it relies on parental tools and design duties instead. Galissaire asks why a child under 13 would be kept off certain platforms but allowed onto an AI companion with a parent’s permission, and says the distinction “should be supported by evidence, not assumptions about which categories of technology are inherently more dangerous.”

She does not call for a ban. Her view is that age limits can be justified when proportionate and evidence-based, but that “restricting access should not become a substitute for addressing the features that make a service potentially harmful.” She also doubts that parental controls can carry the weight the proposal places on them, given how few parents in the survey feel equipped, and says primary responsibility should stay with the companies designing the systems. That is an opinion, and a contestable one: others argue that a clear age floor is easier to enforce than design standards that regulators must assess case by case.

What the proposal actually says

We checked the op-ed’s description against the Commission’s proposal, COM(2026) 681 final, adopted on September 17, 2026. The summary holds up, with one nuance on social media.

Article 6 does not set a flat minimum age. It bars users under 15 from accounts on social networking and video-sharing services where the service poses a risk to minors, and lists the features that trigger that test, such as livestreaming to unknown audiences, contact with strangers and profiling-based recommendations. Between 13 and 15, guardians may set up limited accounts with supervision tools switched on. Few mainstream platforms would escape the risk test, but the structure is conditional.

For AI, Article 14 applies to providers of “AI companions” and “general conversational chatbots” and requires, at a minimum:

1. Avoiding design features and behaviours that simulate interpersonal relationships likely to create emotional dependency, plus the anti-compulsive-use measures that apply to social media.
2. Safe default settings, including a default that the system does not use information from a minor’s earlier conversations in later ones, except where needed for the minor’s safety.
3. Transparency about economic transactions.
4. Access for children under 13 only when enabled and controlled through guardian tools.
5. State-of-the-art evaluation and testing for risks to minors before the system is placed on the market.
6. Post-market monitoring, including detection of serious incidents involving minors, with an exemption for micro and small enterprises.

When a chatbot is built into a social network, video platform or game, the proposal adds that it must not switch on automatically, must not be pushed to minors and must be easy to turn off. Supervision of chatbot providers would run through the AI Act’s enforcement system. All of this is still a proposal under negotiation in the European Parliament and the Council; none of it binds anyone yet.

The rules that already apply

Waiting for the KIDS Act would be a mistake, because several existing laws already reach children’s use of AI:

• EU AI Act. Article 5 has prohibited since February 2, 2025, AI systems that exploit vulnerabilities due to age in ways that materially distort behaviour and cause significant harm. Separate transparency rules in Article 50 require people to be told when they are interacting with an AI system.
• Digital Services Act. Platforms accessible to minors must ensure a high level of privacy, safety and security under Article 28, and the Commission’s July 2025 guidelines on minors also cover AI features, such as chatbots built into platforms.
• GDPR. A child’s conversation history is personal data, and some of it, about health or sexuality for example, may be special category data. Lawful basis, minimisation, retention and Article 8 consent rules for information society services all apply.
• UK Online Safety Act. Ofcom’s children’s safety codes have applied since July 2025, and the regulator has already opened an investigation into an AI companion chatbot service under the Act. Services in scope need highly effective age assurance to keep children from the most harmful content.

The United States is moving on its own track, mostly through state law. New York’s statute on AI companion chatbots and minors is one example of the state-level rules companies with global products must now track alongside Europe’s.

What product and privacy teams can take from the debate

Whatever the final age thresholds look like, the direction is clear enough to plan for. A company offering a chatbot or companion feature that children can reach should be able to answer four questions with evidence rather than assertions.

Who is actually using it? The survey suggests terms of service age limits do not keep 11-year-olds out. Usage analytics, support tickets and app store ratings can show whether a product has a child audience in practice. Any age assurance method has privacy costs of its own, and critics warn that heavy-handed checks can undermine privacy and free expression, so the choice of method needs its own assessment.

What does it remember? The proposed memory default goes to the heart of how many assistants work. Teams should map what conversation data is stored, for how long, who can access it, whether it feeds personalisation or model training, and whether a user can see and delete it. The commercial pull runs the other way: the secrets users tell a chatbot are becoming a product feature, and that is precisely what the proposal wants to switch off for minors by default.

Is it designed to keep children talking? Follow-up prompts, streaks, flattery and simulated friendship are design choices. Evidence on emotional reliance is still thin, but the concern is not new; work on AI chatbots and mental health has made similar points about systems that sound supportive without the safeguards of real care.

Can you prove you tested it? Pre-market testing and incident monitoring mean records. A risk assessment document is not proof that controls work; logs of red-team tests on child scenarios, escalation paths and fixes are. Those monitoring records must themselves be minimised, so that safety evidence does not turn into an archive of children’s conversations.

Frequently asked questions

Is the EU KIDS Act law?

No. The Commission adopted the proposal on September 17, 2026. It must be agreed by the European Parliament and the Council, and the text can change. Under the proposal as drafted, it would generally apply six months after entry into force, with some parts later.

Would the KIDS Act ban children from AI chatbots?

Not as proposed. There is no general age ban for AI companions or general chatbots. Access for under-13s would have to be enabled and controlled through guardian tools, and providers would face design, memory, testing and monitoring duties.

Do customer service bots fall under Article 14?

The proposal’s definition of a general conversational chatbot excludes systems limited to specialised tasks, including customer service, technical support and transactional uses. AI companions are covered whatever their breadth if they simulate a social or emotional relationship.

Who is Jessica Galissaire?

She is a senior policy researcher in interface’s programme on strengthening the digital public sphere and platform regulation. Her Tech Policy Press piece is an opinion article, and the survey was commissioned by her organisation.

Does the survey prove AI companions harm children?

No. It is a self-reported survey of French children and parents. It shows patterns of use and associations, such as higher rates of negative emotions among companion users, but the author herself says it cannot establish cause.

The age question is the easy one

Lawmakers will spend the coming months arguing about numbers: 13, 15, 16. Galissaire’s point is that the number may matter less than what the product does once a child is inside. A rule that keeps 12-year-olds off one app and lets them confide in another, with nothing checking what that second app remembers, would protect the headline more than the child. Negotiators now have the data to decide which of those outcomes they are writing into law.