Macron wants to ban AI where it threatens children: a charter with OpenAI and Anthropic at the Paris Peace Forum

Macron wants to ban AI where it threatens children: a charter with OpenAI and Anthropic at the Paris Peace Forum

🔎 Why this week could mark a turning point

"If there is a risk to children from certain uses of AI, then it must be banned." That quote is from Emmanuel Macron, interviewed by Les Echos on the evening of November 9, 2026, on the eve of the Paris Peace Forum. It was picked up the very next day by AFP (Le Monde, Arab News). And it didn't come out of nowhere.

On Tuesday, November 10, a committee of experts was due to submit its recommendations on protecting minors from AI. Hot on its heels, a charter bringing together OpenAI, Anthropic, Mistral and Microsoft, among others, was expected to be adopted at the forum, following the signing of the Paris Call for AI safety by nearly 50 leaders the day before (Pollar). Around 50 heads of state and government took part in a video conference on the risks AI poses to collective security (The Star, AFP).

Why now? Because the pressure has been building: lawsuits in the United States following tragedies involving chatbots, New York's law against "unsafe" behavior involving minors, the European Digital Fairness Act proposal expected this week, and a summit of AI ministers on Friday in Dublin. The political window is open. What remains to be seen is whether it will produce regulation — or mere window dressing.

The Essentials

• Targeted ban: Macron is calling for AI to be "banned" wherever it threatens children's mental health (Les Echos, November 9, 2026).
• Unprecedented charter: pre-auditing and pre-certification of models before exposure to minors, independent evaluations, access to model data.
• Expected signatories: OpenAI, Anthropic, Mistral, Microsoft.
• Expert report: The Council for AI and Digital Technologies (CIANum) recommends tightening control over assistants "without delay," "from ChatGPT to Gemini or Grok" (Pollar).
• Three-pronged framework: child protection, democratic security, national security (Euronews).
• Major limitation: voluntary commitments, without binding force—for now.
• Next step: European AI ministers' summit, Friday, November 13 in Dublin.

Recommended tools

Until a credible "minors" certification exists, supervision remains operational: here are the tools I recommend to families, plus a hosting solution for schools and associations that want to publish their own AI education resources.

"It must be banned": what Macron actually said

Direct answer: the president is calling for a conditional ban on certain uses, not a blanket prohibition of AI. "It's not because it's AI that it has to be the Wild West," he added in the same interview, as reported by Le Monde and Euronews.

Two quotes frame the moment. The first, blunt: "This system is not tenable, it has to be changed, up to banning it if it's bad" (Pollar). The second, programmatic: "We are going to advance European regulation and push the Americans, the Canadians and the others to do the same. We are creating the groundwork for frontier AI regulation."

The report submitted on November 10 by the CIANum (the AI and Digital Council) points in the same direction: tightening, without delay, oversight of consumer AI assistants, from ChatGPT to Gemini or Grok. In other words, the target is not research, but products exposed to millions of teenagers.

My analysis: the choice of the word "ban" shifts the burden of proof. Until now, harm had to be proven before regulating. From now on, the industry would have to prove safety to reach underage audiences. This is a shift in philosophy, not a detail of vocabulary.

The three-pronged framing

Euronews sums up the strategy in three pillars: children, democratic security, national security. The "children" pillar is the furthest along. The European Commission was due to present its Digital Fairness Act proposal on Wednesday, which provides for the blacklisting of ten online design "traps". Brussels is also reportedly preparing a ban on social media and AI chatbots for under-15s. New York, for its part, has already passed a law against chatbots' "unsafe" behavior with children.

The Paris Peace Forum charter: what the labs are actually accepting

Direct answer: the charter contains three concrete commitments, considerably more technical than your usual statement of intent.

First commitment: the principle of pre-audit and pre-certification, with "a prior review of models before putting them in the hands of children" (Arab News, AFP). This is the logic of a market authorization, transposed to generative models.

Second commitment: independent evaluations of generative AI systems placed on the market. Third: access to model data for evaluators. Without that last clause, the first two would be dead letter — you cannot audit what you cannot inspect.

The expected signatories — OpenAI, Anthropic, Mistral, Microsoft — are therefore accepting ex ante control, having still fought this very principle during the AI Act negotiations in 2023-2024. The charter "aims to inform national and European legislative debates": it is a voluntary bridge toward legal constraint.

The open-model conundrum

Pre-certification assumes a controlled checkpoint. But open or self-hosted models — DeepSeek V4 Pro, Kimi K2.6 (Moonshot AI), GLM-5 (Z.AI) — circulate through a simple weight download. You cannot "withdraw from the market" a model file. The real pressure will fall on distribution: app stores, cloud providers, access to compute. On this point, the charter remains silent. That is its biggest technical shortcoming.

US, EU, China: does "frontier regulation" exist?

Direct answer: three regulatory philosophies coexist, but minors' protection is creating an unprecedented convergence — at least rhetorically.

Australia paved the way on social media: its ban for under-16s came into force in December 2025. If Brussels imposes a 15-year age limit for social media and chatbots, the West will have a de facto alignment on minors' protection — even if everything else diverges.

The Dublin summit on Friday, November 13 will serve as a test: can Europe's AI ministers coordinate 27 capitals on minimum age and age verification? My prediction: yes on principles, no on the timeline.

What the charter legally implies — and what it doesn't

Direct answer: the charter has no binding force. Its value is political and evidentiary, and that's already a lot.

The commitments rest on voluntary participation, not on a binding law (Pollar). A signatory can withdraw. No sanction is provided for. In short: soft law, the most flexible legal category there is.

But soft law hardens through three mechanisms. First, it becomes a de facto standard: the future codes of conduct under the AI Act and the Digital Fairness Act will build on these existing frameworks. Next, it carries weight in litigation: in the event of a tragedy involving a minor, the presence — or absence — of certification becomes part of the case file. The courts are already scrutinizing AI providers' promises: a judge thus found the "5× more usage" claim put forward by Anthropic for Claude Max misleading (our article). Finally, certification will become the entry ticket to public procurement, with education and healthcare at the top of the list.

The thorny question of age verification remains. The EU is testing an infrastructure backed by the European digital identity wallet, but the GDPR paradox is real: collecting sensitive data to protect sensitive data. The charter doesn't resolve this contradiction; it inherits it.

Key dates to remember:

Why labs are signing: the economic calculation

Direct answer: because, for them, the cost of compliance has become lower than the cost of uncertainty.

The enterprise market has become the financial engine of the sector. Anthropic and OpenAI are each launching their corporate joint ventures, with $10 billion to deploy AI in SMEs and large corporations (our analysis). However, a CIO does not buy a model suspected of harming minors: an "audited" label transforms a risk into a selling point, particularly in education.

Valuations play the same role. Anthropic aims for $900 billion with a $30 billion funding round (our article). At these levels, unaddressed regulatory risk is costly in terms of investor confidence. Signing a charter means buying predictability.

Top floor: control of the technical stack. Anthropic acquired Stainless for over $300 million, securing control of the SDK layer and API access (our analysis). Compliance—audit logs, guardrails, age filtering—is now integrated directly into the distribution infrastructure.

My opinion: certification will eventually act as an entry barrier. Established players will absorb the audit costs; open-weights challengers, without equivalent margins, will pay the same price for limited market access. The labs understood this before the regulators did.

The criticisms: three blind spots

Direct answer: the charter reassures the comms people more than it constrains the engineers. Three criticisms come up systematically.

First blind spot: self-regulation without a cop. No funded independent audit authority, no sanctions, free exit. Worse: the labs co-write the rules that will audit them — the classic risk of regulatory capture, AI edition.

Second: age verification. Between biometric solutions (intrusive), age databases (hackable), and age estimation (imprecise), no option is clean. And a VPN erases it all. Digital rights NGOs have been reminding us since 2023 that child protection must not come at the price of mass surveillance.

Third: the announcement effect. France still applies its own digital majority law (2023) imperfectly. The AI Act suffers from a resource shortfall on the Commission's side. A highly publicized charter can serve as an alibi for legislative inaction: "look, the industry is self-disciplining."

Honest counterpoint: an imperfect common standard is better than the status quo. Independent evaluations create a measurement infrastructure that didn't exist before. New York and California show that hard law is moving forward, state by state. The charter is not the destination; it's the first mile.

What changes concretely for parents and educators

Direct answer: nothing changes in your apps this week, and a lot will change within 12 to 24 months.

The scale of the issue, in a few figures: 72% of American teens have already used an "AI companion," and 31% rate these exchanges as satisfying as — or more satisfying than — those with real friends (Common Sense Media, 2025). Still in the United States, 26% of teens had used ChatGPT for schoolwork as early as 2024, up from 13% in 2023 (Pew Research Center).

In practice, without waiting for certifications: turn on the native parental controls (ChatGPT has offered them since 2025), link accounts, and check the minimum age requirements of the services you use. With "reasoning" models such as o1-preview or GPT-5.4 Pro, keep in mind that the displayed chains of thought are not a safety filter — they're a display, not a guarantee.

To watch for in the coming months: the list of the first certified models, the final shape of the Digital Fairness Act, and whatever comes out of Dublin. If pre-certification becomes an app store requirement, then everything will change overnight.

❌ Common Mistakes

Mistake 1: believing AI will be "banned for children"

No. This is a targeted ban on certain uses plus a prior certification requirement, not a general prohibition. Solution: follow the exact scopes — uses, products, age brackets — rather than the headlines.

Mistake 2: confusing a voluntary charter with the law

The charter creates no obligations or sanctions. Solution: watch the legislative transposition and the Digital Fairness Act — the only things that can actually be binding.

Mistake 3: relying solely on age verification

It can be circumvented (VPN) and raises personal data issues. Solution: combine safe-by-default design, parental controls, and dialogue — technology is no substitute for education.

Mistake 4: ignoring self-hosted models

DeepSeek V4 Pro, Kimi K2.6, or GLM-5 escape market controls if freely downloadable. Solution: regulation will need to target distribution (stores, cloud, compute) — otherwise it protects no one.

❓ Frequently Asked Questions

Is the charter legally binding?

No. It relies on the voluntary participation of labs, with no sanction provided in case of non-compliance (Pollar, November 2026). Its strength lies elsewhere: it creates a technical and political standard that French and European legislators can turn into enforceable obligations in upcoming legislation.

Which models are affected by pre-certification?

On the front line, the frontier assistants: GPT-5.5 (OpenAI), Gemini 3.1 Pro (Google), Claude Opus 4.7 (Anthropic), or Grok 4.1 (xAI). Open models like DeepSeek V4 Pro, Kimi K2.6, or GLM-5.1 raise an unresolved question: how do you audit freely downloadable weights? The answer will likely go through distribution channels.

When will we see concrete effects?

Not for several months. A pool of independent auditors, methodologies, and trade-offs still needs to be built. The AI Act's "high-risk" obligations have been applicable since August 2026, and the Digital Fairness Act is expected to shape 2027. Visible effects on consumer products: during 2027, at the earliest.

Can the Dublin summit change the game?

It tests the EU's ability to coordinate 27 capitals on minimum age and age verification. A political agreement would give the charter a legislative foundation; a mere exchange of views would confirm that "frontier regulation" remains, for now, mere rhetoric.

Will the United States follow suit?

Not at the federal level in the short term: Congress remains gridlocked, and the American approach runs through states and courts. New York ("unsafe" chatbots, 2026) and California (SB 243, 2025) illustrate a fragmented regulatory approach — effective locally, incoherent at the national level.

✅ Conclusion

The Paris Peace Forum's charter legally binds no one, but it shifts the protection of minors from the realm of opinion debates to the technical ground of audits — and that is precisely where norms turn into laws.

To measure what these commitments are worth in the face of the courts, continue with Claude Max in court: a judge finds that 5x more usage was misleading.