United States: voluntary standards for AI models on the way — what this changes for regulation

United States: voluntary standards for AI models on the way — what it changes for regulation

🔎 Three days, three signals, one single message

Between June 29 and July 2, 2026, three events reshaped the global AI regulation map. The European Union approves its Digital Omnibus and pushes back its heaviest obligations by sixteen months. The UN publishes an expert report sounding the alarm on the gap between model capabilities and safeguards. Finally, the United States is accelerating on voluntary release standards, with an announcement expected for the week of July 7.

The contrast is striking. On one hand, Europe is backpedaling on the timeline. On the other, the UN says time is running out. In the middle, Washington is advancing an approach that rejects any binding framework in favor of a voluntary system. The problem: no one yet knows if "voluntary" will truly mean "optional".

What is at stake right now is not a theoretical debate. It is the definition of the rules of the game for models like GPT-5.5, Claude Opus 4.7 or Gemini 3.1 Pro — the systems currently sitting at the top of global rankings. And the companies developing them are at the negotiating table.

The Essentials

• The US government is negotiating voluntary AI model release standards with leading developers, with a possible announcement the week of July 7, 2026.
• Trump's Executive Order (June 2026) explicitly prohibits any mandatory licensing and relies on a voluntary 30-day pre-release framework, with a federal deadline of August 1, 2026.
• The EU approved on June 29, 2026, a postponement of the AI Act's high-risk obligations from August 2026 to December 2027, but maintains transparency (Article 50) by August 2.
• The UN scientific panel (Bengio + Ressa) warned on July 1 that safeguards are not keeping pace with capabilities, on the eve of the Geneva Global Dialogue on AI Governance.

Recommended Tools

Trump's Executive Order: a "voluntary" framework with very real consequences

The "Promoting Advanced AI Innovation and Security" Executive Order, signed in June 2026, is the founding text of the new American approach. Its central message: no licensing, no preclearance, no mandatory permitting to develop or release an AI model.

It's written in black and white in the text. Freshfields le confirme: the EO explicitly prohibits mandatory licensing, preclearance, or government permitting for the development, release, or distribution of models.

But behind this "no" to hard regulation, there is a "yes" to a mechanism that could prove just as structuring. Frontier model developers are invited — the keyword being "invited" — to voluntarily submit their most powerful models to the federal government up to 30 days before their public release. The stated goal: cybersecurity testing.

The detail that matters: Latham & Watkins précise that federal agencies must design a detailed framework for this voluntary system by August 1, 2026. We are therefore no longer in the realm of a declaration of intent, but in operational construction.

Voluntary on paper, mandatory in practice?

This is the entire paradox that Ropes Gray décortique: the June 2026 EO creates a "voluntary with mandatory implications" framework. According to the firm, combined with the March 2026 legislation, this voluntary approach generates effective obligations for developers.

In short: no company will be legally forced to submit its model. But the consequences of not doing so — in terms of relations with regulators, public procurement, and media treatment — could make the choice rather binary.

NPR résume la situation: the EO asks AI companies to voluntarily submit their most powerful models to the government for testing up to 30 days before their release. The verb "asks" does all the political work here.

Ongoing negotiations: what Reuters reveals

On July 2, 2026, Reuters rapporte that the US government is in advanced negotiations with AI companies to establish voluntary model release standards. The information, initially reported by the Financial Times, specifies that an announcement could come in the following week.

This timing is not coincidental. It coincides almost exactly with the Global Dialogue on AI Governance organized by the UN in Geneva on July 6 and 7. Washington has every interest in showing that it is making progress on the issue before the debate is captured by multilateral bodies.

The negotiations focus on the concrete form the voluntary framework will take. What criteria define a "frontier model"? What cybersecurity tests are required? What is the exact duration of the pre-review window? These are all questions that will determine whether the system will be a symbolic gesture or a genuine tool of control.

Reuters avait déjà anticipé le mouvement début juin: the Trump administration would ask leading AI developers to voluntarily submit their most capable models for federal cybersecurity testing. A month later, the discussions have moved from the announcement stage to the technical negotiation stage.

Which models would be affected?

The framework targets "frontier models," meaning the most high-performing systems. Based on current rankings, this would directly target the top of the agentic leaderboard: GPT-5.5 (98.2), Gemini 3 Pro Deep Think (95.4), Claude Opus 4.7 Adaptive (94.3), or even GPT-5.4 Pro (91.8).

These models are the ones whose capabilities are advanced enough to justify a prior cybersecurity review according to the administration's criteria. Mid-tier models like Claude Sonnet 4.6 (81.4 in agentic) or Grok 4.1 (79) would likely remain outside the scope.

The European pullback: the Digital AI Omnibus

While the US builds, Europe pushes back. On June 29, 2026, the EU Council gives its final approval to the Digital AI Omnibus, a text that profoundly alters the timeline of the AI Act.

DLA Piper details the most significant change: the AI Act's high-risk obligations shift from August 2, 2026, to December 2027. That is a sixteen-month delay, an eternity in the AI development cycle.

Sidley clarifies the outlines of the provisional agreement reached on June 22. Article 50 on transparency remains at August 2, 2026, but the grace period to comply with it is reduced from six to three months. The European message is therefore twofold: it maintains pressure on transparency, but eases up on high-risk.

Why this delay changes the global dynamic

The AI Act was supposed to be the counter-model to the American approach. Strict regulation, based on risk classification, with proportionate obligations. By pushing the core of the mechanism to December 2027, Europe creates a temporal void that the voluntary American approach fills.

Companies operating globally — OpenAI, Google, Anthropic — find themselves in a paradoxical situation. In the US, they are offered a voluntary framework with concrete implications. In Europe, the mandatory framework is pushed back to a horizon where current models will already be obsolete.

The OpenAI GPT-5.6: Sol, Terra and Luna family arriving on the market will be subject to American voluntary standards before facing European high-risk requirements. It is a discreet but real geopolitical reversal.

The UN sounds the alarm: the Bengio-Ressa report

On July 1, 2026, the UN's independent scientific panel on AI published its preliminary report. Co-chaired by Yoshua Bengio (Turing Award winner, historic figure in deep learning) and Maria Ressa (Nobel Peace Prize laureate), this panel brings together more than 30 countries.

The findings are unequivocal: safeguards are not keeping pace with AI capabilities. Reuters cites the report: policymakers face a growing dilemma between innovation and safety, and current protection mechanisms are out of step.

Asanify summarizes the week: the UN report arrives exactly between the European approval of the Digital Omnibus (June 29) and the revelations regarding US negotiations (July 2). Three days that tell a coherent story — that of global governance in a state of permanent catch-up.

The Geneva Global Dialogue as a tipping point

This preliminary report directly feeds into the Global Dialogue on AI Governance scheduled in Geneva on July 6 and 7, 2026. The objective: to converge national approaches toward a multilateral framework. The problem is that the two main regulatory poles have taken opposite directions in 72 hours.

Europe is pushing back its obligations. The US is betting on voluntary measures. The UN says it is not moving fast enough. The report published by the UN clearly positions the panel as an independent third party that refuses to choose between innovation and precaution, but demands that both advance at the same speed.

What this concretely changes for developers

Beyond geopolitics, what does this voluntary framework change for teams training and deploying models? Several concrete elements emerge from the available sources.

The 30-day window: technical constraint or formality?

Submitting 30 days before release for federal cybersecurity testing might seem trivial. In practice, for a model like GPT-5.5 which dominates the agentic leaderboard with 98.2, this means freezing the model a month before its public availability. During that month, the competition — Gemini 3 Pro Deep Think at 95.4, Claude Opus 4.7 at 94.3 — could advance.

This is a real competitive cost that companies will factor into their release timing calculations. The question is no longer "should we submit the model?" but "at what point in the development cycle should we trigger the submission to minimize competitive impact?"

The definition of a "frontier model" as a central issue

The detailed framework expected for August 1, 2026, will need to precisely define what a frontier model is. Compute thresholds? Benchmark scores? Agentic capabilities? This definition will determine which models go through the pre-review and which are released directly.

A model like DeepSeek V4 Pro Max (88 overall, unranked in standard agentic) could find itself in a gray area. Too performant to be ignored, not "frontier" enough to trigger the full mechanism. These edge cases will be the subject of intense negotiations in the coming weeks.

The US strategy in the global context

The US voluntary approach is not improvised. It is part of a coherent strategy that aims at three simultaneous objectives.

Avoiding the European trap

The AI Act has been criticized — including by European companies — for its complexity and its ambitious timeline. The December 2027 delay for high-risk obligations is an implicit recognition that the initial timeline was not feasible. Washington observed this difficulty and drew a lesson from it: a framework that is too rigid breaks.

It is in this context that one must understand the Great American AI Act bill, a 269-page proposal circulating in Congress that could freeze AI regulation in the United States for three years. Trump's Executive Order and the voluntary standards are partly a response to this legislative threat: showing that it is possible to regulate without legislating.

Maintaining pressure without creating a legal precedent

The Executive Order signed by Trump is a balancing act. By refusing mandatory licensing, the administration avoids creating a precedent that could be used by future administrations to tighten the framework. By establishing a voluntary mechanism, it maintains a channel of access to models without codifying it into law.

Positioning the United States ahead of Geneva

The possible announcement of voluntary standards in the week of July 7 is not a calendar coincidence. The Global Dialogue on AI Governance opens on July 6. Arriving in Geneva with an already negotiated and soon-to-be-operational framework gives the United States a position of strength in multilateral discussions.

Comparison table of the three approaches (July 2026)

❌ Common mistakes

Mistake 1: Confusing "voluntary" and "without consequence"

The Executive Order is presented as voluntary, but Ropes Gray shows that the implications are de facto mandatory. A company that refuses to submit a frontier model exposes itself to political and regulatory consequences that are disproportionate to the "choice" it makes.

Mistake 2: Thinking that the European delay cancels the AI Act

The December 2027 delay concerns high-risk obligations. Article 50 on transparency remains scheduled for August 2, 2026, with a grace period reduced to three months. Companies that think they have 18 months of respite risk missing this deadline.

Mistake 3: Dismissing the UN report as a mere diplomatic exercise

The panel co-chaired by Bengio and Ressa is not a discussion group. Its preliminary report directly feeds into the Geneva Global Dialogue and will serve as a scientific reference in national regulatory debates. The figures and warnings it contains will be cited in every upcoming piece of legislation.

Mistake 4: Assuming that the US framework only concerns US companies

The criterion is not the nationality of the company but the US market. A model developed in China (DeepSeek V4 Pro) or in France (Mistral, if it reaches the frontier threshold) that targets the US market would logically be subject to the same voluntary mechanism.

❓ Frequently Asked Questions

Who decides if a model is a "frontier model"?

The detailed framework expected by August 1, 2026, from federal agencies should define objective criteria (compute thresholds, benchmarks). So far, the EO does not specify a numerical threshold, leaving the field open to ongoing negotiations with companies.

What happens if a company refuses voluntary pre-release?

Legally, nothing. The EO explicitly prohibits mandatory licensing according to Freshfields. In practice, the relational and media consequences could be significant, even though no sanctions are codified.

Is Europe lagging behind the United States on AI regulation?

Not exactly. Europe has a law in force (the AI Act), which the United States does not. But the postponement of high-risk obligations to December 2027 creates an operational gap that the US voluntary framework fills temporally. The two approaches are at different stages, not in direct competition.

Will the UN report lead to an international treaty?

The scientific panel is an advisory body, not a negotiating one. Its role is to inform the Global Dialogue on AI Governance in Geneva. A treaty would require a separate intergovernmental process, which is not on the immediate agenda according to available sources.

Are open source models covered by the US framework?

The EO targets "developers of frontier models." An open source model like Kimi K2.6 in its self-host version (88.1 in agentic) raises a boundary question: is the initial developer responsible for all variations? The August 1 framework will have to settle this question.

✅ Conclusion

In 72 hours, from June 29 to July 2, 2026, global AI regulation took a clear direction: the US voluntary approach is establishing itself as the default operational benchmark, Europe is buying itself time, and the UN is documenting the urgency. The detailed framework expected by August 1 and the announcement of voluntary standards in the coming days will determine whether this architecture holds up in the face of models that continue to gain in capabilities.