Working draft. Everything on this site is open for public comment and subject to revision.

← Library

NB-07 · Note

When the Parent Does Not Act

Enrollment as a duty in European law

10 November 2026 · S. Harrison Knudson, J.D.

A proposal for discussion. This is a draft, and comments are welcome.

Summary

Enrollment protects the child whose parent does nothing. It is not parental supervision. Under the amendment we propose, it would be a duty on the provider: no AI companion, general conversational chatbot or functionally equivalent system may begin a personalised conversation with a minor until a person holding parental responsibility has enrolled that child for a defined role. Until then, the answer is no. The only exceptions are systems run by helplines and confidential services that the law already lets children reach alone.

The strongest objection to enrollment is that measures relying on parents fail the children who get the least support at home. That objection is right about supervision. It does not apply to a duty that holds when the parent is absent, careless or the danger. This note answers it in four parts:

  1. The default is closed. A child nobody enrolled cannot be drawn into a conversation with a chatbot at all. Parents can open one defined role; they cannot switch protection off.
  2. Safety duties do not depend on enrollment. The provider owes every one of them to every minor, enrolled or not.
  3. Help never waits for a parent. Helplines, confidential services and abuse referrals stay open to the child without a guardian's consent.
  4. Parental responsibility is not only parents. Where parents are absent or unfit, a guardian or a youth welfare authority holds it and can enroll.

Enrollment answers one question: may this provider begin a conversation with this child, in this role? Safeguarding governs everything after that. Both are the provider's responsibility, and the first can be audited before any child is harmed.

A note on status. This paper distinguishes three things throughout: law in force, the Commission's KIDS Act proposal (not yet law), and the amendments IFC-SiS has proposed. Enrollment is our proposal. It is not already in the Commission's text; the paper argues that it fits the machinery the text builds.

The objection

The children most at risk online are often the children with the least support at home. A rule that works only when parents set controls, read notices and check settings will protect the children who were already protected. It will miss the rest.

That objection is consistent with the finding of the European Commission's Special Panel on Child Safety Online. Its report of 13 July 2026 says the burden of proof needs to be on providers, not regulators, parents and children. The panel extended its scope beyond social media to AI chatbots and companions, and flagged AI companions for the emotional dependency they can create in young people.

We agree. Parental controls, as built today, put the burden on parents: the service is open by default and the parent must find the switch to close it. A child whose parent never finds the switch gets the open service. Enrollment reverses that. The service is closed by default, and the burden of opening it lawfully sits with the provider.

The leading US child-protection body says much the same in its own guidance to parents. The National Center for Missing & Exploited Children's Gaming Safely Playbook (2026) warns that parental controls “can create a false sense of security” and that “children can easily find workarounds”; that controls “don’t always carry across platforms”; and that “a factory reset can silently undo your protections.” Of more than 60,000 CyberTipline reports with a gaming connection, nearly 67% referred to moving the conversation from a game to another app, game or website. The playbook then gives parents 26 separate steps to compensate. That is the burden enrollment is designed to remove.

Enrollment is a duty on the provider, not a task for parents

Enrollment is a rule addressed to the provider: a provider shall not engage a minor in conversation unless a person holding parental responsibility has enrolled that minor for a defined role. The parent is not asked to watch, configure or report anything. If no one acts, the rule still applies, and the answer is no.

The offline analogy is familiar. A school, a sports club, a doctor and a youth group each receive a child from the person responsible for that child, under their own rules of admission. None of them may ordinarily begin a relationship with a child on the child's own say-so, and none of them waits for a parent to object before that rule applies.

Three features keep the burden off parents:

The practical effect is simple. Under parental controls, the child of an absent parent gets the open service. Under enrollment, the same child gets no anthropomorphic companion at all, and every protection described below.

Protections that hold without any parent

Every protection in this section applies to every minor, enrolled or not. None of them requires a parent to know, act or consent. Some already exist in EU law; others are amendments we have proposed, and each is labelled.

Safety duties for every provider (proposed). Our feedback on the KIDS Act proposes two tiers: safety duties for every provider, and enrollment before conversation. A provider cannot escape the Article 14 duties by declining to enroll. Coverage should follow function, not label, so a specialised tutoring or wellbeing system that holds a sustained, personalised conversation with a minor is covered whatever its provider calls it.

No human pretence (existing law, plus proposal). Existing EU law already sets two floors that no parent can lower:

Enrollment builds on those floors. An enrolled system fills a stated role and says what it is. It may not present itself as a person, a friend or a therapist, and a role a parent cannot grant, such as a romantic companion, cannot be enrolled at all.

Help never waits for a parent (proposed). Our filed amendments keep three routes open to every child:

These routes follow General Comment No. 25 of the UN Committee on the Rights of the Child, which says providers of preventive or counselling services to children online should be exempt from any requirement of parental consent (para. 78), warns that safety monitoring can stop a child from reaching a helpline (para. 76), and asks for confidential access to psychological counselling (para. 94). The KIDS Act proposal points the same way: Member States must establish national strategies ensuring minors have easy, free and confidential access to channels through which they can seek assistance with the harms it addresses (Article 33(1)(a)).

When the parent is absent, or is the danger

Enrollment is given by a person holding parental responsibility, not only by a parent. That distinction matters most for the children the objection is about.

The KIDS Act proposal already points this way. Its recitals say providers remain primarily responsible for the safety of their services when used by minors, and define a guardian as any person holding parental responsibility under applicable national law (Article 3(5)(b)), with recital 15 tying that notion to Council Regulation (EU) 2019/1111, which covers persons, institutions and other bodies that hold it, not only parents (COM(2026) 681, recital 15). Article 26 sets out how providers verify parental responsibility for guardian accounts and tools for guardians; enrollment is not yet among its purposes, and our filed amendment would add it. Enrollment would give that existing machinery an admission function.

When parents are absent. European family law already names who holds parental responsibility when parents cannot. In Germany, for example, the youth welfare office (Jugendamt) can act as guardian and then exercises the whole of parental custody for the child (Amtsvormundschaft; BGB §§ 1773 ff., 1786; SGB VIII § 55). The same holds for a court-appointed guardian. Under enrollment, that guardian enrolls; a child in care is not left outside the system, and not left to the open service either.

When the parent is the danger. General Comment No. 25 recognises that abuse may be perpetrated within a child's circle of trust, including by family (para. 81), and that a child's privacy may be vital where parents themselves threaten the child's safety (para. 77). Enrollment is built for that case:

When a parent refuses. Confidential help survives a parent's refusal; ordinary education may not. A child in a coercive household could be denied a useful tutor for no safeguarding reason. Our feedback lets a school enroll a pupil for an educational role with the guardian's consent, and the Commission's text already excludes systems designed primarily for educational purposes and operated by or for educational establishments (Article 2(4)(c); recital 13). An open question we want tested: where a guardian cannot be reached or refuses without reason, should a school or youth service be able to enroll for a narrow educational role on its own authority, and under what safeguards?

This follows the structure of the UN Convention on the Rights of the Child. The child holds the rights; the person with parental responsibility exercises them in line with the child's evolving capacities (Article 5); and the state must protect the child from abuse while in anyone's care (Article 19).

The principle beyond chatbots

The immediate proposal is narrow: enrollment before a personalised conversation with an AI companion, a general conversational chatbot or a functionally equivalent system. That is the amendment we filed, and it can be adopted on its own.

The principle behind it is broader, and we name it here as a future extension, not as part of the present amendment. When the law requires authorisation before a relationship with a child begins, it creates a duty that could apply to any person or system forming a relationship with a child online: a game that pairs a child with adult strangers, or a messaging service that lets an unknown adult reach a child directly. Extending it would need its own functional definition of covered contact, with clear treatment of one-off questions, moderated groups, classroom discussion and contact the child initiates.

Offline, this duty is so ordinary that nobody names it. A stranger cannot enroll a child in a sports club, take a child on a trip or tutor a child behind the parent's back. The European Union already states the principle that what is illegal offline should be illegal online. Online, the duty was never written down, so a stranger, or a company, can begin a relationship with a child unilaterally.

The duty sits at the door, before contact, not after harm. It does not wall children in:

Legislation should reference the norms that already exist, not reinvent them. The rules for adults who work with children are already written in the norms of teaching, medicine, counselling and coaching, including the duty to refer disclosures of abuse. An enrolled system should meet the safeguards of the profession that ordinarily fills its role. Europe already has the machinery to do this: Regulation (EU) No 1025/2012 lets the Commission request a European standard, and the KIDS Act lets providers rely on an adequate code of conduct (Article 14(3)). The law points to the standard; it does not have to write it.

Who the relationship is for

The relationship a child forms with a commercial AI system is not between equals. On one side is a child. On the other is a service built by specialists in engagement, whose revenue often depends on the child's attention, sold on to advertisers or converted into subscriptions. Parents should not need a degree in computer science to protect their children from that relationship.

EU law already recognises the commercial pressure. The DSA bars online platforms from showing advertising based on profiling to a recipient they know with reasonable certainty is a minor (Article 28(2)), and General Comment No. 25 goes further, asking states to prohibit by law the profiling or targeting of children of any age for commercial purposes (para. 42). Yet a 2026 audit of TikTok found that commercial content outside the DSA's narrow definition of advertising, such as influencer and brand posts, was profiled to minors' interests five to eight times more strongly than the formal advertising profiling permitted for adults. Rules aimed at the advertisement miss the relationship that delivers it. Enrollment governs the relationship itself.

German law recognises the underlying principle

German tort law has long put the duty on whoever controls a hazard that attracts children, not on the children or only on their parents. Under the Verkehrssicherungspflicht (BGB § 823(1)), the Federal Court of Justice holds that an owner or operator may not rely on children staying out of a danger zone that tempts a child's play instinct, and must take effective, lasting protective measures, even where children enter despite a prohibition (BGH, 4 May 1999, VI ZR 379/98; BGH, 13 March 1995, VI ZR 34/94). A firefighting pond left unfenced, although the building permit required a fence under DIN 14210, breached that duty of the owner, the architect and the contractor (BGH, 12 November 1996, VI ZR 270/95, NJW 1997, 582). A sign reading "Eltern haften für ihre Kinder" does not exclude the owner's liability (Haufe). Courts do weigh the supervision that parents can be expected to give, but the operator's own duty remains. The Court has also held that this duty can demand more care than public-law rules prescribe (BGH, VI ZR 379/98, headnote). Meeting one technical requirement does not exhaust the duty to protect against foreseeable harm, which is why enrollment can never stand in for continuing safety duties.

These judgments do not establish any duty of digital enrollment. They establish a principle: whoever creates or controls a foreseeable hazard to children must secure it. That is the offline pool fence, and enrollment would be a legislative application of the same principle. A service designed to draw children into a relationship is a hazard that attracts children; its operator must secure it at the entrance, whatever parents do. Applying this doctrine to digital services is an argument by analogy, not settled law, but German youth-protection law already points the same way: JuSchG § 24a requires covered providers to take structural precautionary measures, including safe default settings, with a parental accompaniment tool as only one measure among several (BzKJ criteria).

The constitutional frame is the same. Article 6(2) of the Basic Law makes the care and upbringing of children the natural right of parents and the duty incumbent on them first, and charges the state with watching over how that duty is carried out. Enrollment respects that order. The parent's right comes first, and the closed default protects it: no company may decide on the family's behalf which relationships begin. The state's role stays secondary; it does not choose roles for the child, it only ensures that the door is secured until the person with parental responsibility chooses. German courts have not applied the Verkehrssicherungspflicht to conversational AI, and this paper does not claim they have.

What enrollment does, and does not, establish

Enrollment is a condition of lawful admission, not a transfer of responsibility from the provider to the guardian. It answers one question: has the provider been authorised to engage this minor in a defined conversational role?

It does not establish that any later conversation is safe. It does not waive the child's rights, excuse a breach of the provider's statutory duties, or decide whether a product was defective or caused harm. Those questions stay with the applicable law.

Nor is enrollment data-protection consent. A role-specific authorisation to interact is a different legal act from consent to processing under the GDPR. The provider must separately establish a lawful basis for processing, respect data minimisation and retention limits, and protect the child's confidentiality.

The resulting duty is narrow and enforceable. Before a covered conversation begins, the provider must establish that an authorisation exists and covers the role. After that, it must stay within the role and meet every applicable safeguarding obligation. The first question can be audited without waiting for injury; the second remains subject to ordinary supervision and remedies.

What the duty gives courts and regulators

Today a court or regulator has to work out, after the harm, what a provider owed a child it was never supposed to be talking to. Enrollment writes the authorisation question down in advance: who may speak with this child, in what role, under which profession's norms. Whether a provider conversed with an unenrolled minor then becomes an objectively auditable compliance question, though verification, role classification and record integrity can still be disputed.

In the United States, the natural enforcer of such a duty is a parent's lawsuit. In Europe, enforcement runs mainly through regulators and collective redress, and the duty fits each route:

Route Where it stands What enrollment adds
Regulator enforcement The KIDS Act proposal (COM(2026) 681, 17 September 2026, not yet law) applies AI Act enforcement to AI companions and general chatbots, with fines of up to 6% of worldwide annual turnover for intentional or negligent breaches (Article 34(2)) A test a regulator can check before harm: was this child enrolled, and did the system stay within its role? That serves the Special Panel's call to put the burden of proof on providers.
Collective redress The proposal itself would add the KIDS Act to Annex I of the Representative Actions Directive (EU) 2020/1828 (Article 41), alongside the DSA (DSA Art. 90) and the AI Act (AI Act Art. 110). Minors and guardians may also mandate a not-for-profit body to act for them (Article 21) A clear basis for qualified entities to act on conversations with unenrolled minors, if enrollment is written into the Act.
Product liability Directive (EU) 2024/2853 covers software and AI systems placed on the market or put into service after 8 December 2026 (Article 2(1), as corrected on 7 May 2026). Article 7 lists the needs of the intended users among the factors for defectiveness. Article 10(2)(b) presumes defectiveness where the claimant shows the product does not comply with mandatory product safety requirements under Union or national law intended to protect against the risk of the damage suffered A written safety expectation for child users, and nothing more automatic. A claimant still needs legally relevant damage and causation, and whether an enrollment rule counts as a mandatory product safety requirement is unsettled. The regulatory breach is easy to show; a damages claim is not.
Data protection GDPR Article 8 applies where consent is the legal basis for processing in an information society service offered directly to a child; national thresholds run from 13 to 16 None directly. Enrollment is a separate authorisation to interact and must itself comply with the GDPR. Article 8 shows only that EU law already verifies who holds parental responsibility.

A bounded safe harbour. Article 14(3) of the proposal already lets providers show compliance through an adequate code of conduct. Our feedback builds on that, asks the Commission to request a European standard under Regulation (EU) No 1025/2012, and proposes, as a new step the co-legislators would have to write in, a presumption of conformity. Any such presumption should be rebuttable, limited to the obligations the standard covers, and unable to extinguish independent statutory duties or remedies. A provider can enroll a child properly and still manipulate or harm that child; conformity is evidence of care, not immunity.

The Commission's text gives enrollment a precise starting point (COM(2026) 681):

Enrollment would extend the under-13 rule into a uniform, role-specific condition on personalised conversation with any minor, using the verification machinery the proposal already builds.

Objections and answers

"A child whose parent never enrolls is left with nothing." The child loses one thing: an open-ended conversation with an AI system no responsible adult has admitted. Search, reading, school, helplines and confidential services all remain. Where parents are absent, a guardian or youth welfare authority can enroll.

"Teenagers need more freedom than this." They do, and the rule is built to be proportionate to age. What changes with age is the range of roles a guardian may open, not whether an authorisation exists: a 17-year-old can be enrolled once, for broad study help, for a full school year. The burden is one decision, not a permission slip per question. Confidential help is open at every age without consent, and the rule governs personalised conversation with a commercial system, not what a teenager may read, search or learn. The KIDS Act itself defines a minor as anyone under 18 (recital 15).

"General Comment No. 25 requires respect for adolescents' evolving autonomy." It does, and this is the point on which the rule is most open to challenge. The Committee asks states to treat evolving capacities as an enabling principle, notes that children engage more independently of parents online, and asks that services be appropriate to children's evolving capacities (paras. 19–20). It also asks that any restriction be lawful, necessary and proportionate (para. 59). Enrollment is designed to meet that test. It would be set by law; it answers a specific risk the Special Panel identified, emotional dependency on AI companions; it reaches only personalised conversation with commercial systems, not information, expression or association; and it involves no monitoring, which the Committee warns can cut children off from help (para. 76). The Commission's own text gives older teenagers more say elsewhere, letting minors over 15 consent to changes in default settings (Article 11(2)). Whether 16- and 17-year-olds should be able to open some roles themselves is a fair question, and one the pilots should test rather than assume.

"This shifts responsibility to parents." It places the burden on the provider. Every safety duty stays with the provider whether or not the child is enrolled. A guardian, or the authority holding parental responsibility, still makes one decision, and that agency is deliberate: the alternative is a company deciding for the family. If no one decides, the child is protected by the default rather than exposed by the gap.

"Children will lie about their age, or use an adult's account." The provider does not need to guess who is a child. It needs to know whether a guardian has enrolled this child for this role. Under our proposal the child is recorded by a child-specific credential, not a legal name, and a provider that claims it was deceived must show it tested its verification and fixed known weaknesses.

"Enrollment means surveillance." It means the opposite of monitoring. Enrollment is one record of one authorisation. It imposes no general obligation to monitor conversations, and revoking enrollment deletes the minor's conversation records and stored memory, subject to legal retention duties and the child's own data-protection rights; records needed as evidence of a safety incident are kept under restricted access. The authorisation record should disclose only what is needed to show that a role was authorised.

"This will stifle innovation." Schools, clubs, doctors and counsellors all work with children under rules of admission and duty. A service that wants to work with children can do the same, and the safe harbour tells it how.

"This will block legitimate speech and learning." Over-blocking is a real risk, and the rule is drawn to limit it. It applies only to personalised conversation with AI companions, general chatbots and functionally equivalent systems, not to content, search or publication. Systems designed primarily for education and run by or for schools are already outside the Commission's scope (Article 2(4)(c)). The opposite risk is the larger one: the proposal's definition of a general conversational chatbot excludes specialised educational systems (Article 3(5)(d)), so a commercial tutoring bot could escape Article 14 altogether. Our filed amendment would cover any system that holds a sustained, personalised conversation with a minor. The pilot measures below include the rate at which ordinary educational questions are wrongly refused, because a rule that chills learning would fail on its own terms.

"Providers cannot prove who is a minor." They do not have to. The Commission's text already applies its child-safety design to unregistered users until a provider has established through age assurance that the user is an adult (Article 8). Enrollment follows the same logic: a user not shown to be an adult is treated as a minor, and a provider that converses with that user without an authorisation must explain why. The burden sits with the party that controls the system and its records.

What it would take to implement

Enrollment needs a credential that does not yet exist at scale. Its minimum requirements:

The residual risks are real: impersonation, account sharing, fraud by adults posing as guardians, delay for children in care, and disputes where parental responsibility is contested. None of them is a reason to keep the default open. Each is a reason to test the credential before the rule takes effect, which is what the next section proposes.

Draft amendment text for the safe harbour (proposed). Building on our filed Amendment 8 to Article 14(3):

Harmonised standards. The Commission shall request one or more European standardisation organisations, in accordance with Article 10 of Regulation (EU) No 1025/2012, to draft European standards on enrollment, the defined role and role-based safeguards. AI systems in conformity with harmonised standards, or parts thereof, the references of which have been published in the Official Journal of the European Union, shall be presumed to comply with paragraphs 1a to 1d of this Article to the extent that those standards cover those requirements. The presumption is rebuttable and is without prejudice to the other obligations of the provider under this Regulation and to remedies available under Union or national law.

What we want to test

We are setting this argument out at length in The Geography of Zero, a book for parents now in preparation, and we hope to test enrollment with partners in several countries. A pilot should answer questions a regulator and a child-protection practitioner would ask:

  1. How many conversation attempts with unenrolled minors does a closed default stop, and how many legitimate uses does it block by mistake?
  2. How often do enrolled systems decline out-of-role requests, and what do children ask for when they are declined?
  3. Do referrals to helplines and confidential services reach the child as fast without a guardian's involvement as with it?
  4. Can children in care be enrolled by their guardian or youth welfare authority without delay?
  5. Where no guardian can be reached, or a guardian refuses without reason, does a narrow school or youth-service enrollment route help children, and what safeguards does it need?

A sixth question is technical: does the child-specific credential hold up against impersonation, account sharing, delayed revocation, linkage across providers, and misuse by an abusive guardian?

The children the objection is about, those with the least support at home, should be measured separately in every one of these, as should children in care, children under contested or cross-border guardianship, and the rate at which ordinary educational questions are wrongly refused. We would welcome the view of practitioners who build safeguarding concepts for institutions on how to design those measures.

Sources

EU instruments and proposals

Children's rights

German law

Child-protection guidance

Commentary

Scope note. The KIDS Act is a proposal, not law, and article numbers may change in negotiation; references here were checked against the operative text in Council document ST 13321/26 of 18 September 2026 and should be re-checked against the latest Council and Parliament texts before formal circulation. Two items still rest on secondary sources: the 1996 pond judgment (full text as reproduced by a law-firm site; an official reporter citation is preferable) and the youth welfare office's guardianship powers (Haufe; check against BGB §§ 1773 ff., 1786 and SGB VIII § 55 on gesetze-im-internet.de).

This note is a proposal meant for discussion. Critical reactions — doctrinal, comparative or practical — are more useful than supportive ones.

Send a comment