The European Commission’s proposed EU KIDS Act would impose dedicated safeguards on AI companions and chatbots accessible to minors. Medical ethicist Theodoros Zarkos, who has mapped its Article 14 against the scientific literature, argues that some of these rules rest on solid evidence, others on precaution – and that the law should be honest about the difference.
A chatbot used to retrieve information is one thing. A system that remembers a young person’s disclosures, responds in an intimate or affiliative style and presents itself as a steady source of reassurance is another. The regulatory problem changes when interaction begins to resemble a relationship.
That question is now before European lawmakers. On 17 September 2026, the European Commission proposed the EU KIDS Act. Article 14 would impose specific obligations on providers of AI companions and general conversational chatbots accessible to minors. Providers would have to avoid design features likely to create emotional dependency, apply safer default settings, restrict the reuse of information from previous interactions, evaluate risks before market entry and monitor emerging harms afterwards. When these systems are embedded in social networks, video-sharing platforms or online games, they would not be activated automatically or promoted to minors, and young users would have to be able to opt out.
The direction of travel is clear. The harder question is whether the evidence is equally clear.
What the evidence actually shows
In a recent DOI-registered preprint, I examined how closely the safeguards in Article 14 correspond to the available scientific evidence. The review included 36 scientific sources, represented across 111 mechanism-level extraction rows, and 20 legal or normative sources analysed separately. Nine studies met a stricter threshold for direct, minor-specific empirical evidence.
The result was not a simple case for or against regulation. Evidence was uneven across the mechanisms examined. Of the seven relational-risk mechanisms examined, four showed direct correspondence with Article 14’s safeguards, two partial correspondence and one indirect correspondence.
The clearest signals concern relational presentation, emotional reliance and unsafe responses. In a preregistered experiment with adolescents aged 11 to 15, relationally framed chatbot responses were perceived as more human-like, trustworthy and emotionally close than explicitly non-human responses, although both styles were judged similarly helpful. Other research indicates that adolescents and young adults already use AI chatbots for mental-health advice when feeling sad, angry, nervous or stressed. A simulation study of therapy and companion bots also found that some systems endorsed clearly ill-advised proposals presented by fictional distressed adolescents.
These findings matter, but they should not be stretched beyond what they establish. They do not show that every emotionally responsive chatbot harms minors, nor do they establish a long-term causal pathway from relational interaction to dependency. Some conversational systems may also provide useful support, particularly where access to human help is limited. The evidence base remains relatively young and methodologically heterogeneous, with much of it relying on short-term experiments, self-report, simulated scenarios or mixed-age samples.
One gap is especially important. Persistent conversational memory can make an interaction more personalised and continuous, yet minor-specific evidence about its long-term effects remains thinner than the evidence on relational style, trust or unsafe responses. Article 14 therefore reaches into areas where precaution is doing part of the regulatory work.
The bioethical question is relational
The bioethical concern is not limited to screen time, inaccurate information or isolated harmful content. It also concerns the conditions under which a minor forms trust, discloses personal information and makes decisions.
A conversational system can influence autonomy without explicitly instructing a user to do anything dangerous. If it remembers intimate disclosures, adapts its language to emotional cues and consistently responds as though it were a dependable social presence, the design itself may affect how a young person interprets the interaction. For minors, whose judgment and self-direction are still developing, that possibility deserves particular attention.
Protection, however, should not be reduced to paternalism. Young people also have legitimate interests in privacy, access to information, emotional expression and developing independence. A proportionate framework must distinguish between restricting manipulative or dependency-forming design and blocking uses that may be benign or beneficial.
Precaution can be justified when potential harms are serious and uncertainty remains, but precaution should not be presented as proof. Regulation is more credible when it makes uncertainty visible and creates mechanisms for learning from it.
What Article 14 can – and cannot – settle
Article 14 is better understood as a risk-governance proposal than as the legislative expression of settled science. Several safeguards have a plausible empirical basis: limiting dependency-inducing design, providing safer defaults, testing systems before deployment and monitoring real-world harms afterwards.
Important questions nevertheless remain. How should emotional dependency be defined and measured? At what point does useful personalisation become manipulative? When does conversational memory improve safety, and when does it deepen vulnerability? How should providers demonstrate that safeguards remain effective after models and product features change?
The Commission’s proposal does not answer all of these questions, and it should not be expected to. The EU KIDS Act has only begun the ordinary legislative process and may change before adoption.
What matters is the standard that follows. Where evidence is reasonably direct, safeguards can be correspondingly firm. Where evidence is partial, regulation should remain precautionary but open to evaluation. Where evidence is indirect, the law should require better testing, transparency and post-market evidence rather than treating uncertainty as if it had already been resolved.
For relational AI and minors, that is more defensible than either technological optimism or blanket alarm. The relevant question is not whether conversational AI is simply good or bad for young people. It is which design features create which risks, for whom, under what conditions, and how confidently we know.
Disclosure: This article draws on the author’s DOI-registered preprint, “Bioethical Governance of Relational AI for Minors: Evidence–Regulation Alignment in Article 14 of the Proposed EU KIDS Act” (DOI: 10.5281/zenodo.23064572). The preprint has not yet undergone peer review.
About the author
Theodoros Zarkos, MSc, is a Medical Ethicist and Social Scientist based in Kiato, Greece. He holds an MSc in European Social Policy and Social Rights and conducts independent research in bioethics, medical ethics, child rights, digital health, and AI governance.
This post is also available in: FR