Macron Wants AI Banned Where It Threatens Children. The Paris Charter Asks for Testing Instead

France is trying to turn children’s safety into the first piece of what President Emmanuel Macron calls “frontier AI regulation.” In an interview published on the eve of the Paris Peace Forum, Macron said AI should be banned wherever it threatens children’s mental health or brain development. On November 10, 2026, the Forum’s Child Summit published a Charter of Engagement for Safer AI for Children, endorsed by Anthropic, Black Forest Labs, Microsoft and OpenAI. The charter is voluntary and creates no legal obligations. But it sets out testing, evaluation and transparency practices that governments say they will use to shape future law, and it is a practical guide for any company whose AI product a child might use.

What Macron said, and where

The remarks come from an interview with the French daily Les Echos, as reported by Euronews and the Associated Press on November 10. “If there is a risk for children with certain uses of AI, then it must be banned,” Macron said, adding that “just because it’s AI doesn’t mean it should be the Wild West.” He said companies would accept “the principle of a pre-audit and pre-certification, with a prior review of the models before putting them in children’s hands,” and set out a broader ambition: “We are going to advance European regulation and push for the Americans, the Canadians, and others to do as we do. We are in the process of creating the beginnings of frontier AI regulation.”

Macron grouped AI dangers into three categories: children, democratic security and major national security issues. He also held a videoconference with around 50 heads of state and government, international organizations, companies and researchers. Euronews reported that the British, Canadian, Norwegian, Dutch and Japanese prime ministers took part, along with the chief executives of OpenAI, Anthropic, Mistral and Microsoft. The same push sits alongside his campaign to bar under-15s from certain social networks in France, and his broader complaints about Europe’s pace on big tech, which we covered when Macron criticized the EU’s geopolitical lag.

What the charter actually says

The charter comes from the Coalition for Children’s Rights and Protection in the Age of AI, which France, Spain and Kenya lead with ten UN agencies participating. Spain’s involvement follows its own push on the issue, including the call we reported when Spain’s youth minister urged a UN working group on big tech and children. It builds on earlier political texts, including the G7 Evian Call for a safer digital space for minors and principles adopted by G7 digital ministers in June 2026. The four endorsing companies are listed on the last page. Mistral took part in Macron’s call but is not listed as an endorser of the charter.

Its ten shared principles include safety and privacy by design and by default, protections against excessive and compulsive use and emotional dependency, age-appropriate design, parental tools, data minimization for children, and age assurance that does not rely solely on self-declaration. The operational commitments fall into three groups.

• Testing and incident response. Companies developing or deploying AI accessible to minors will run “robust evaluation and testing” for risks to children’s health, safety, fundamental rights and physical, mental and emotional well-being “before deployment where relevant and at regular intervals thereafter.” They will put mitigations in place, monitor after launch, and detect and respond to serious incidents involving children.
• Independent evaluation. Signatories will work with government-backed evaluation institutes, research institutes and child-safety experts, and may give evaluators appropriate access to systems. Companies can audit the assessment methods applied to them and raise documented objections, but they get no authority over safety criteria or conclusions and “no right to delay or suppress findings.”
• Research and transparency. Companies contribute money, expertise and access to models and relevant data “when lawfully permitted.” Methods, measurement tools and results, “including null, negative or inconclusive findings,” are to be published, and a scientific committee will publish an annual progress report. Contributions give “no privileged access to children’s personal data.” The Forum said AI developers including OpenAI and Anthropic, together with Common Sense Media, would pledge 8 million euros to research on AI and child development.

The text does not mention certification, licensing or approval. The gap between “pre-certification” in Macron’s interview and “evaluation and testing… where relevant” in the signed document is the most important point for readers. Macron’s office described the charter as intended to inform national and European legislative debates. Turning it into a pre-market approval requirement would take legislation, at the EU level or in France, that does not yet exist.

The declaration governments signed

Separately, more than 50 countries, six international organizations and about 20 philanthropies and companies signed the Children First Declaration, which covers health, nutrition, education, protection and psychosocial well-being in the digital age, with targets for 2045. Canada, the United Kingdom, Australia and most EU member states are on the Forum’s published list. The United States is not. In the Forum’s release Macron went further than the charter text: “no AI should be placed in the hands of children without first being tested, evaluated, and therefore authorized.” This is a political commitment that signatories may act on, not a treaty.

How it fits existing law

Several of the charter’s themes are already law in Europe, which is why European companies should not treat it as a purely voluntary exercise.

• EU AI Act. Article 5 has prohibited AI practices that exploit vulnerabilities due to age to materially distort behavior in harmful ways since February 2, 2025. Enforcement depends on national authorities, several of which were designated late, as we noted in our report on member states that missed the AI Act authority deadline.
• Digital Services Act. Article 28 requires online platforms accessible to minors to ensure a high level of privacy, safety and security for them, and the Commission’s guidelines under that article address age assurance and recommender design.
• GDPR. Children’s data gets specific protection, and France sets the age of digital consent at 15. Charter commitments on data minimization and on keeping evaluators away from children’s personal data line up with existing duties.

In the United States, the momentum is at the state level, with laws such as the one we covered when California passed its child-safety chatbot and social media law. Federal proposals for pre-deployment testing, like the bill we examined in our piece on the AI Evaluation Act, have not become law. That is the gap Macron says he wants to “push” the Americans to close.

A five-step child-safety evaluation plan

Appendix 2 of the charter sets out six requirements for evaluation methods. Condensed into five steps, they work as a checklist for any team assessing a product that children could reach, whether or not the company signed.

1. Define the harm and the failure. Write down which harms you are testing for, what counts as a failure, and examples on each side, separating harmful behavior from appropriate discussion of sensitive topics.
2. Test by age and context. Cover the ages, languages, abilities and settings you expect, and report where results differ or evidence is thin.
3. Test real use and foreseeable misuse. Include children’s own phrasing, indirect disclosures, curiosity and attempts to bypass safeguards, and run long, repeated conversations where risks build over time.
4. Test the whole stack. Evaluate model safety, age assurance, product safeguards, account actions and parent notifications together, and check whether under-18 users are actually identified. Report results for correctly configured teen accounts separately from tests where age was unknown or misclassified.
5. Make the results reviewable. Record prompts, configurations, scoring criteria, evaluator qualifications, uncertainty and false positives so an outside reviewer could repeat the work. Feed the results into your data protection impact assessment.

Teams that already run DPIAs will find most of this familiar. The difference is the focus on developmental outcomes and the expectation that outside evaluators can see the method. Our article on the children left out of privacy design covers why products built for adults so often fail these tests.

Frequently asked questions

Does the charter bind companies that did not sign it?

No. It is a voluntary commitment by its endorsers. Non-signatories are still bound by existing law such as the AI Act, the DSA and the GDPR, and the charter shows the direction future rules may take.

Is there now an approval process for children’s AI products in France?

No. Macron spoke of pre-certification and authorization, but no French or EU law creates such a process for AI products today. Any approval regime would need new legislation.

Does it apply only to products made for children?

No. The testing commitment covers AI systems “accessible to minors,” which includes general-purpose chatbots and image tools that children can reach, not only children’s apps.

What should a smaller company take from it?

Use Appendix 2 as a testing template, document the results in your DPIA, and make sure age assurance and parental controls are tested together with the model.

A voluntary text that may become a template

Voluntary codes often become reference points for regulators once lawmakers act. This charter gives regulators a ready description of what responsible pre-deployment testing looks like, written and endorsed by the companies that would have to do it. Companies that build those practices now will be better placed if the evaluations later become mandatory. Read the full charter and the Paris Peace Forum announcement, and compare your own testing against Appendix 2.