At the G7's Evian Summit, AI Chiefs and World Leaders Strike a Voluntary Deal Centered on Youth Safety

When the 52nd G7 summit closed on June 17 in Evian-les-Bains, the most consequential photograph was not of seven heads of state. It was of a lunch table. Seated among the leaders of the world's largest democracies were Sam Altman of OpenAI, Dario Amodei of Anthropic, and Demis Hassabis of Google DeepMind — the three executives whose labs sit closest to the frontier of artificial intelligence, gathered in one room for the first time at a summit of this rank. The working lunch, themed "ensuring a safe, rapid and effective deployment of artificial intelligence," produced a deal. But it was a deal defined as much by what it was not as by what it was: voluntary, not binding; centered on children, not on the existential risks the labs themselves keep warning about.

What the leaders agreed

The headline outcome was a "Leaders' call on a safer digital space for minors," issued jointly by the G7 and a striking roster of partner nations — India, Brazil, Egypt, Kenya, and South Korea among them. The declaration commits signatories to a "safety-by-design" standard: platforms would be expected to ship protective default settings, age-assurance mechanisms, and privacy-preserving defaults that empower parents and guardians, rather than bolting on safety features after harm occurs.

The language was unusually firm for a multilateral text. The prohibition of child sexual abuse material, online grooming, and sexual extortion, the leaders wrote, "remains a non-negotiable principle in the development and deployment of artificial intelligence systems and digital service." For a document otherwise built on voluntary cooperation, that phrasing was the closest Evian came to drawing a hard line.

That youth safety topped the agenda was no accident. Chris Lehane, OpenAI's chief global affairs officer, had signaled before the summit that child protection sat at the top of Altman's personal priorities, with frontier cyber and bio risks a secondary concern. By the close, the companies had agreed to a package of voluntary commitments rather than submit to any new legal regime.

The room within the room

The marquee attendees were the three frontier-lab chiefs, but the broader guest list read like a census of the global AI industry: Mistral's Arthur Mensch, Cohere's Aidan Gomez, Black Forest Labs' Robin Rombach, Sarvam AI's Pratyush Kumar, Synthesia's Victor Riparbelli, Meta's Alexandr Wang, Salesforce's Marc Benioff, and Sakana AI's Ren Ito. Altman attended at the personal invitation of French President Emmanuel Macron, who used the summit's margins to press his industrial-policy case. That courtship paid a tangible dividend: SoftBank's Masayoshi Son committed roughly €45 billion to French AI data centers after Macron's lobbying.

Inside the lunch, the executives did not merely listen. Amodei and Hassabis used the platform to call for a U.S.-led coalition to shape international AI rules and standards — Amodei arguing that cooperation should encompass structured access to frontier models and a China-excluding trade regime for chips and critical components, alongside joint defenses against AI-enabled cyber operations and bioterrorism. Altman, for his part, pitched an international testing forum for advanced models. None of it produced binding text. Canada, notably, signaled openness to the coalition idea.

Where it fell short, and why

The summit's other AI workstream — data-center sustainability — came out conspicuously thin. A more ambitious environmental position was watered down, attendees said, because of U.S. resistance. That same dynamic shaped the entire AI outcome. Washington's opposition to mandatory regulation is the reason the Evian deal arrived as a set of voluntary commitments rather than enforceable obligations, and it is the clearest expression yet of the transatlantic split now defining AI governance. Europe, having already passed the binding AI Act, wanted checks on what it sees as American dominance; the United States wanted room to run.

The over-reliance risk got an unexpected champion in Canadian Prime Minister Mark Carney, who pointed to Anthropic's recent forced shutdown of its Fable 5 and Mythos 5 models — disabled for all customers after a U.S. export order — as a live warning. "The situation we're in collectively right now with Mythos and Fable is something that can happen with over-reliance on certain models," Carney said, comparing the systemic fragility to the interbank linkages that amplified the 2008 financial crisis. The blackout was still unresolved as the summit ended.

Analysis: the gap between the table and the text

The image of all three frontier CEOs at one G7 table is the story's real significance. It marks a shift in how AI governance is being negotiated — not in regulators' chambers but across a lunch table, with the regulated parties as co-authors. That access cuts two ways. It gives leaders direct insight into capabilities moving faster than legislation. It also lets the labs steer the conversation toward the rules they prefer, and away from the binding constraints they resist.

The voluntary-versus-binding tension is now the central fault line of multilateral AI policy. Europe brings the legal apparatus; the U.S. brings the companies and the leverage to keep commitments soft. Youth safety became the consensus deliverable precisely because it is the rare issue where both camps, and the public, agree — making it the safe ground on which to declare progress while the harder questions of liability, model concentration, and enforcement go unanswered.

What to watch

Three threads will test whether Evian was substance or stagecraft. First, the Fable 5 and Mythos 5 blackout: a U.S.-Anthropic deal to restore access would validate Carney's model-risk warning even as it resolves the immediate crisis. Second, the proposed AI standards forum — whether Altman's testing body or the Amodei-Hassabis coalition gains real institutional form, or evaporates as so many summit ideas do. Third, the minors roadmap itself: voluntary "safety-by-design" pledges have a long history of slipping, and the gap between Evian's "non-negotiable" language and the absence of any enforcement mechanism is where this deal will be judged.