Washington negotiates voluntary AI rules, a necessary gamble

Introduction: frontier AI enters the era of negotiated rules

A voluntary framework taking shape

According to the Financial Times, the White House is currently negotiating with OpenAI, Google, and Anthropic a voluntary framework meant to govern the release of artificial intelligence models deemed "frontier," with an announcement possible as early as next week according to information relayed by Reuters. This project builds on the presidential executive order of June 2, 2026, titled "Promoting Advanced Artificial Intelligence Innovation and Security," through which Donald Trump had already directed federal agencies to build an evaluation process for models before public release.

The new framework under negotiation would set precise testing criteria and review timelines, one more step toward concretely structuring a process that had remained vague in its application, despite the order having been published a month earlier.

Why this moment matters

This project comes as the American administration has already, in practice, asked OpenAI to delay the public release of its GPT-5.6 model, and as Anthropic saw its most advanced models, Claude Mythos and Fable 5, temporarily blocked by export controls before being partially restored in late June, according to CNN.

I think this long-awaited clarification is good news in itself: a framework that remained vague, applied case by case with no public criteria, risked producing an arbitrary system where every company had to guess the rules of the game.

The June 2 order, foundation of the new framework

What the presidential text already provides

The order signed by Donald Trump on June 2, 2026 establishes a voluntary mechanism by which developers of frontier models can give the federal government secure access for 30 days before sharing their systems with trusted partners, according to the text published by the White House. The order explicitly specifies that it creates no mandatory licensing regime for the development or release of new models.

A classified referral process, under the responsibility of the NSA director, is meant to determine which models reach the threshold of "covered frontier model," a threshold that remains publicly undefined as of this writing, according to legal analysis from the firm Crowell & Moring.

An implementation several observers call confusing

The outlet The Hill reports that the request made to OpenAI and Anthropic to delay their models has created confusion, with some critics noting that no one, including within the administration itself, seems to know precisely what safety standards a company must meet to get a green light for wide release.

I believe this gray zone, however frustrating for the companies involved, remains preferable to the total absence of oversight that previously prevailed over technologies whose strategic reach now extends well beyond the private sector alone.

What the new voluntary framework could change

Clearer testing criteria

According to the Financial Times, the new standards under negotiation would set technical benchmarks for advanced models along with precise timelines, while clarifying who can access these systems in the United States and abroad. This clarification would directly answer criticism voiced in recent weeks over the process's opacity.

Google is also reportedly in talks with the government ahead of releasing its own advanced coding models, whose cybersecurity capabilities are said to exceed those of previous generations, according to the British newspaper.

A delicate balance between security and competitiveness

The challenge for Washington is to establish safeguards sufficient to prevent the cybersecurity and biosecurity risks associated with the most powerful models, without slowing American innovation amid increasingly fierce international competition, notably from China.

I think the West doesn't have the luxury of choosing between security and speed: it must deliver both at once, since every month of poorly calibrated regulatory delay directly benefits Chinese labs, which feel no such scruples about racing ahead unencumbered by these precautions.

Tensions between labs and regulators

OpenAI and the GPT-5.6 case

OpenAI announced last week that it would present its new GPT-5.6 model, including the Sol, Terra, and Luna variants, only to a "small group of trusted partners" before any public rollout, at the request of the American administration and for cybersecurity reasons, according to TechCrunch. CEO Sam Altman reportedly told employees that the government would "approve access client by client" during the preview period.

This situation concretely illustrates how the presidential order is already playing out in practice, well before the new voluntary standards framework currently under negotiation is even finalized.

Anthropic and the export-control precedent

The Anthropic case illustrates the limits of the current system: its Claude Mythos and Fable 5 models were hit by export controls for several weeks before the Commerce Department lifted the restrictions in late June, following joint work on risk assessment, according to CNN and Commerce Secretary Howard Lutnick.

I see in this Anthropic episode proof that the system works, even imperfectly: a problem was identified, addressed, and resolved through cooperation rather than confrontation, exactly the kind of process a more formalized framework should be able to reproduce faster in the future.

The comparison with the Biden era

A voluntary precedent that already existed in 2023

The idea of voluntary AI standards isn't entirely new: in July 2023, the Biden administration had already secured non-binding commitments from seven major companies, including Amazon, Google, Meta, Microsoft, and OpenAI, covering in particular external safety testing before release, according to Politico and the American National Standards Institute.

This precedent shows that the logic of voluntarism runs through successive American administrations, regardless of political affiliation, proof that the question extends well beyond the usual partisan divides.

What sets the current approach apart

Unlike the largely symbolic 2023 framework, the 2026 order comes with concrete enforcement mechanisms, notably through the NIST's Center for AI Standards and Innovation, which is already conducting pre-deployment evaluations with major American labs, according to The GPU Trade.

I think this difference in rigor, between a token commitment in 2023 and a real technical evaluation mechanism in 2026, marks a welcome maturation in how Washington approaches AI governance.

The geopolitical stakes behind the regulation

The AI race as a field of strategic rivalry

The June order was initially delayed by concerns over its potential impact on American competitiveness against China, according to The Hill, a reminder that AI regulation cannot be considered independently of the broader technological rivalry between major powers.

The order also explicitly plans to strengthen national cyberdefense capabilities and ease access to defensive AI tools for critical infrastructure, from rural hospitals to community banks.

Why the West must stay ahead

This technological race is not merely a commercial matter: it directly shapes the balance of power between the West and its strategic rivals, chief among them China, which is investing massively in its own artificial intelligence capabilities without the same transparency constraints.

I remain convinced that America's advantage in artificial intelligence is one of the most valuable pillars of Western power today, and that a well-calibrated standards framework, neither too lax nor too rigid, is precisely what will preserve it over the long run.

Persistent criticism and gray areas

Vagueness over decision criteria

Several analysts, including commentator Ball cited by The Hill, note that "literally no one knows" what safety standards a company must meet to get authorization for wide release of its models, a vagueness that, in practice, amounts to systematic refusal until criteria are formalized.

This uncertainty weighs directly on the planning of AI labs, which must navigate shifting requirements with no clear visibility into timelines or final approval criteria.

Transparency still insufficient, according to some experts

Critical voices, including within the companies concerned, are calling for the publication of testing criteria, the list of trusted partners, and some form of recourse in case of disagreement, all elements still absent from the current setup according to several specialized legal analyses.

I think these criticisms are well-founded and deserve to be heard: a voluntary framework can only function durably if it rests on public, predictable rules — otherwise it risks turning into a mere informal power struggle between Washington and the major labs.

What tech companies expect next

A shared call for predictability

OpenAI and Anthropic have both advocated for a federal framework establishing clear safeguards around the highest AI risks, while diverging on how large a role government should play in the development process itself, according to The Hill.

This shared demand for predictability reflects a simple industrial reality: labs prefer binding but known-in-advance rules over an unpredictable approval process that delays their launch timelines without public justification.

The promise of a repeatable process

According to TechCrunch, OpenAI has already said it wants to work with the administration to establish a repeatable process for future model releases, a way of turning a one-off, constrained experience into a standardized mechanism applicable across the entire industry.

I think this stated desire to build a repeatable process, rather than endure case-by-case decisions, is the most encouraging sign in this entire sequence: exactly the kind of structured cooperation the West needs to stay competitive without sacrificing safety.

Conclusion: an imperfect framework, but a necessary direction

A welcome clarification despite the gray areas

The voluntary framework under negotiation between the White House, OpenAI, Google, and Anthropic represents a serious attempt to structure a technological field whose rapid evolution far outpaces the ability of traditional institutions to regulate it effectively and on time.

It remains to be seen whether the expected announcement in coming days will finally bring the clarity companies are demanding, or whether it will simply formalize a status quo already widely criticized for its opacity.

What the West stands to lose or gain

As China continues to invest massively in its own artificial intelligence capabilities, America's ability to establish a framework that is both safe and flexible enough to preserve innovation will prove decisive for the global technological balance of the years ahead.

I close this editorial convinced that an imperfect but transparent framework is always better than permanent improvisation, and that the West's future credibility in the AI race will largely depend on its ability to get this clarification done quickly and well.

By Maxime Marquette, columnist

Columnist's transparency note

Who I am and how I work

I am a columnist and analyst specializing in technology and geopolitics: I rely exclusively on the journalistic and official sources cited below, without privileged access to negotiations between the White House and the AI labs.

I have received no compensation from OpenAI, Google, Anthropic, or any government entity mentioned in this piece.

My limits and my acknowledged biases

The voluntary standards framework discussed here has not yet been officially announced at the time of writing; some details could change before its formal publication by the White House.

I hold an editorial stance favorable to Western technological leadership over China, which shapes the tone of this piece without altering the accuracy of the facts reported.

Sources

Primary sources

The White House, "Promoting Advanced Artificial Intelligence Innovation and Security" order — June 2, 2026

U.S. News/Reuters, Washington in talks with AI companies over voluntary standards — July 1, 2026

TechCrunch, White House asks OpenAI to slow the rollout of its new model — June 25, 2026

Secondary sources

The Hill, confusion reigns over the White House's AI policy — July 2, 2026

CNN, White House lifts export controls on Anthropic's models — June 30, 2026

Crowell & Moring, legal analysis of the frontier AI models order — June 3, 2026