explainx / blog
The G7 held a 110-minute working lunch with 12 AI CEOs on June 17, 2026. Here is everything that happened — the Trusted Partners scheme to restore frontier AI access, the child safety declaration, Modi's human-centric AI push, Japan's Hiroshima Process update, and what Trump said about Anthropic.
Jun 21, 2026
The White House's June 2, 2026 EO mandated a classified AI benchmarking process and a 30-day pre-release framework for "covered frontier models" within 60 days. Fable 5 launched June 9 — 7 days after the EO — with no government pre-brief. The June 12 ban was the mechanism for forcing cooperation the voluntary framework could not compel. Here is the full Section-by-Section breakdown and why it matters.
Jun 13, 2026
Ban lifted July 1, 2026. Commerce restored Fable 5 and Mythos 5 globally after 18 days. GPT-5.6 broad access expected next on same export-control framework.
Jun 11, 2026
"Treebeard and his forest are waking up." Anthropic CEO Dario Amodei argues that the window to shape AI policy is now open—and closes out five concrete policy areas where governments must act before the exponential outruns democratic institutions.
The 52nd G7 Summit in Évian-les-Bains, France concluded on June 17, 2026 with AI at the centre of its final day. Twelve technology CEOs sat down with G7 heads of state for a 110-minute working lunch titled "Ensuring Safe, Rapid, and Efficient AI Adoption." The backdrop was impossible to ignore: Anthropic's Fable 5 and Mythos 5 had been suspended globally for five days under a US export control directive, turning what might have been a forward-looking policy conversation into an immediate crisis management session.
This is the full picture of what happened — the proposals made, the commitments given, the things deliberately left unresolved, and what it all means for how frontier AI gets governed from here.
The working lunch was unprecedented in format: seven heads of state seated alongside the CEOs of the companies building the technology they were discussing. Confirmed attendees from the AI and tech sector:
| CEO | Company |
|---|---|
| Sam Altman | OpenAI |
| Dario Amodei | Anthropic |
| Demis Hassabis | Google DeepMind |
| Arthur Mensch | Mistral AI |
| Aidan Gomez | Cohere |
| Marc Benioff | Salesforce |
| Alex Wang | Meta |
| Victor Riparbelli | Synthesia |
| Robin Rombach | Black Forest Labs |
| Representatives | Sakana AI, Domyn, Salvum AI |
On the government side: US President Donald Trump, French President Emmanuel Macron, UK Prime Minister, Japanese Prime Minister Sanae Takaichi, German Chancellor, Italian Prime Minister, Canadian Prime Minister, and EU Council President António Costa. India's Prime Minister Narendra Modi attended as a partner country invitee.
The format — structured working lunch rather than plenary session — was a deliberate choice to enable direct conversation between leaders and technologists rather than formal statement exchange.
The most significant AI policy development from Évian was not in any official communiqué. It was a proposal taking shape in the margins.
G7 leaders discussed the creation of a Trusted Partners scheme — a mechanism that would grant allied nations access to advanced US AI models currently restricted by export control. The immediate context was the Fable 5 ban: when the US Commerce Department suspended foreign national access to Anthropic's Fable 5 and Mythos 5 on June 12, it locked out every non-US user — including citizens of the US's closest allies.
French President Macron made the Trusted Partners proposal his signature item at the summit. He met privately with Dario Amodei and Sam Altman on June 17 before the working lunch, and told reporters afterwards: "I expect progress in coming weeks on broadening access to leading US AI models." He framed it explicitly as a response to the Fable 5 ban and as a necessary condition for European AI sovereignty not to become entirely dependent on models European governments cannot access.
The scheme as discussed would operate on a tiered trust model: G7 members first, then broader allied nations — analogous in structure to the Five Eyes intelligence sharing arrangement. Countries would need to meet specified data governance and security standards to qualify. No formal criteria have been published; the scheme is still at the proposal stage.
What it is not: The Trusted Partners scheme is not the UK exemption proposal that died on June 17. That was a bilateral carve-out from the existing Commerce Department directive — a faster but narrower path. The Trusted Partners scheme is a broader multilateral framework being negotiated at the intergovernmental level, with a longer timeline and potentially more durable effect. Macron's expectation of "coming weeks" may be optimistic; building a multilateral AI access framework typically takes months.
What it means for international users: If the scheme advances, allied nation users could regain Fable 5 and equivalent frontier model access through a government-certified channel. If it stalls or collapses under US domestic political pressure, international users are back to the alternatives — GLM-5.2, Kimi K2.7, OpenRouter Fusion — indefinitely.
President Trump told reporters on June 17 that negotiations with Anthropic over restoring Fable 5 and Mythos 5 access are "going fine."
It was brief, casual, and notably unspecific. But its significance is contextual: it was the first direct comment from Trump himself on the ban, delivered from a summit where allied leaders were pressing him for clarity. "Going fine" is optimistic framing — not a committed timeline, not a policy announcement, not an acknowledgement that the export control was misapplied.
Read alongside the Trusted Partners discussion, it suggests the administration is comfortable with the talks proceeding but has not committed to the outcome those talks are pushing toward. Whether "going fine" means "we will lift the ban soon" or "we are managing the diplomatic pressure comfortably" is not clear from the statement itself.
Prime Minister Narendra Modi attended as a partner country invitee and spoke at the AI session on human-centric principles — making India's position among the most clearly articulated of any non-G7 participant.
His core argument: the true test of AI is not how powerful machines become, but how much it empowers ordinary people. He called out the risk of AI becoming a tool that concentrates capability in a handful of advanced nations while the Global South is left out or exposed to the downsides — misinformation, deepfakes, exploitation — without the governance infrastructure to manage them.
Modi's specific recommendations:
1. Safe-by-design AI systems — safety as a fundamental design principle, not a post-hoc addition. He called for this to be an international baseline standard rather than a voluntary corporate practice.
2. Common standards and testing frameworks — including regulatory sandboxes where AI can be deployed and tested under controlled conditions before broader release. The framing: innovation and governance should advance together, not sequentially.
3. Child safety as a non-negotiable — Modi was emphatic that without proper standards, AI that educates children can also expose them to exploitation. He supported the formal child safety declaration that emerged from the summit.
4. Inclusive access for all democratic countries — "All democratic countries should have access to such AI models so that they can protect their critical information infrastructure." This is a direct reference to the Fable 5 ban and the Trusted Partners discussion — India is not a G7 member and would need to be part of any scheme that extends beyond the core seven.
5. Global cooperation on deepfakes and cyber fraud — effective joint response to cross-border misinformation and synthetic media.
Modi also referenced Pope Leo XIV's recent remarks on AI — that human values, inclusivity, and meaningful human control must be the foundation for AI development. India's MANAV Vision encodes the same principle domestically; Modi framed the two as aligned expressions of a global consensus that technology must serve people, not the reverse.
The speech positioned India as a normative voice on AI governance — not as a developer of frontier models (India is not in that tier currently), but as a large democratic nation with standing to shape how access and safety standards are applied globally.
Japanese Prime Minister Sanae Takaichi used the working lunch to advance Japan's Hiroshima AI Process — the multilateral framework Japan launched at the 2023 G7 to develop international guiding principles for frontier AI, which has been running since.
Takaichi positioned AI as the "cornerstone of Japan's growth strategy" and argued for a framework built on "Data Free Flow with Trust" (DFFT) — the principle that data should move across borders freely within a governance architecture that ensures security and privacy. She proposed that AI ecosystems be tailored to different regional needs through international cooperation rather than a single global standard imposed from one country.
Japan also committed to hosting an AI Summit to advance the Hiroshima Process, and emphasised the importance of cybersecurity protections for AI-dependent critical infrastructure — specifically naming finance, telecommunications, and electricity sectors.
The Hiroshima Process had produced a set of guiding principles in 2023 and a code of conduct for AI developers in 2024. The 2026 G7 conversation was focused on whether those voluntary frameworks need to be strengthened — and whether a mechanism exists to apply them to models being pulled offline by export controls, which the original process was not designed to address.
The most concrete formal outcome from the G7 on AI was a Roadmap to Secure Digital Space for Minors, co-signed not just by G7 nations but by India, Brazil, Kenya, Egypt, and South Korea as partner countries.
The declaration's key commitments:
Non-negotiable prohibitions: A joint reaffirmation that CSAM, non-consensual intimate imagery, and deepfakes involving minors are categorically prohibited. The "non-negotiable" framing was deliberate — an explicit pushback against arguments that content moderation standards should be negotiated case-by-case.
Safety-by-design requirements: Digital service providers are called on to adopt safety-by-design approaches including privacy-preserving defaults, age-appropriate experiences, and parental control tools enabled by default — not opt-in. The distinction matters: default-on safety is a meaningfully different standard than opt-in safety.
Conversational AI specifically named: The declaration explicitly calls on providers to "make conversational artificial intelligence tools safer for children and youth, in a timely manner." This is the first G7-level declaration to specifically call out AI chatbots as a child safety concern alongside social media and content platforms.
Age assurance: The roadmap calls for development and deployment of age assurance solutions — technical mechanisms to verify user age before granting access to certain content or capabilities. This is a live regulatory debate in the UK, EU, and US; the G7 declaration adds multilateral political weight to those domestic proceedings.
Governance follow-through: G7 ministers are directed to assess progress on the child digital safety principles by end of 2026.
Kenya's inclusion as a co-signatory was notable — representing Sub-Saharan Africa's first formal participation in a G7 child AI safety commitment. The broad coalition (G7 + India + Brazil + Egypt + South Korea + Kenya) gives the declaration more normative weight than a G7-only statement.
No full transcript of the working lunch has been published, but on-the-record statements before and after provide a picture of where the industry positioned itself.
Sam Altman (OpenAI): Told leaders that "the debate over whether AI is useful is over" and urged the G7 to take control of AI governance before it becomes ungovernable. He framed this not as a call for restriction but as a call for structured engagement — governments building the expertise and institutional capacity to understand what they are regulating. Altman's message: governance is now inevitable; the question is whether governments do it well or badly.
Dario Amodei (Anthropic): The most scrutinised attendee given the Fable 5 situation. His presence at the table — the CEO of the company whose flagship product is currently suspended by his own government — was a live illustration of the tensions the summit was trying to resolve. His specific statements were not publicly reported in detail, but he met privately with Macron prior to the lunch, and the Trusted Partners proposal that emerged aligns closely with Anthropic's stated interest in reaching a negotiated access framework.
Demis Hassabis (Google DeepMind): Engaged primarily on the scientific and infrastructure dimensions of AI safety — consistent with DeepMind's public positioning as a research organisation with safety as a core research agenda.
Arthur Mensch (Mistral): France's domestic frontier AI champion, with an obvious stake in the Trusted Partners discussion. Mistral's models are European-built and do not face the same export control exposure as US models, giving Mensch a somewhat different posture: he benefits from the Fable 5 gap but also benefits from a stable multilateral framework that doesn't fragment AI access along nationalist lines.
For all the significance of the room — twelve AI CEOs, seven heads of state, 110 minutes — the one thing the G7 did not produce was any binding framework for frontier AI development or deployment.
No agreement on compute thresholds that trigger international review. No binding safety evaluation requirements before release. No shared standards for what constitutes a "jailbreak" serious enough to justify regulatory action. No mechanism to prevent a repeat of the Fable 5 scenario — where a single government issues an unilateral export control that effectively severs global access to a major AI product.
This gap is not a failure of will at Évian specifically. It reflects a genuine unsolved problem: frontier AI governance is moving faster than international institutions can build consensus on it. The Hiroshima AI Process produced voluntary principles; the EU AI Act produces binding rules for EU jurisdiction; the US operates through export controls and national security authorities that are not coordinated with anyone. These frameworks do not interlock.
The Trusted Partners scheme is an attempt to solve one piece of this — restoring access for allies — without addressing the underlying governance architecture that made the Fable 5 ban possible. If it succeeds, allied nations get their models back. The next time a US export control directive hits a frontier model, the same scenario plays out again.
The harder version of what Évian failed to produce: a shared standard, with multilateral enforcement, for when a jailbreak is serious enough to justify access suspension — and a process that involves the affected parties before action is taken. That standard does not exist. As of June 17, 2026, it is not being negotiated.
Trusted Partners scheme: Macron's "coming weeks" target puts the scheme's initial framework at roughly early-to-mid July. The first question is whether the US Commerce Department, which issued the Fable 5 directive, is prepared to accept an allied-nation carve-out — or whether that requires a political decision above the Commerce Secretary level.
Fable 5 specifically: Trump's "going fine" and Macron's optimism about the Trusted Partners timeline both point toward some form of resolution in July. Whether that resolution is domestic-only (via Anthropic's ID verification mechanism), or also covers allied nations via the Trusted Partners scheme, or is a full lifting of the directive, determines who actually gets Fable 5 back.
Child safety follow-through: G7 ministers are tasked with progress assessment by end of 2026. The inclusion of "conversational AI tools" in the declaration will feed into domestic regulatory proceedings in the UK (Online Safety Act expansion), EU (AI Act implementation), and US (ongoing platform liability debates).
The Hiroshima Process: Japan's planned AI Summit will be the next multilateral checkpoint for the governance frameworks discussed at Évian. No date announced yet.
The G7 AI conversation in Évian was substantive in ways that previous G7 AI discussions were not. Twelve real decision-makers sat in the same room. A concrete proposal — Trusted Partners — emerged with a named champion and a timeline. A formal declaration on child safety was signed by twelve nations. The backdrop of a live AI ban added urgency that abstract policy discussions rarely carry.
What it was not: a resolution. The fundamental tension — between US national security prerogative over its AI companies' products and the global nature of those products — left the room in the same unresolved state it entered. The Trusted Partners scheme is an attempt to manage that tension, not dissolve it.
For the full story of the Fable 5 ban that dominated the G7 AI discussions, see Why Did the US Government Ban Fable 5?. For what the ban means for users outside the US, see Will Fable 5 Only Be Available in the USA?. For restoration timelines, see When Will Fable 5 Be Available Again?.
