Trump imposes a voluntary federal framework for AI

Washington steps back into the governance of advanced models

Donald Trump has signed a new executive order that puts the White House back at the center of the debate over oversight of the most powerful artificial intelligence systems. According to information reported by The Verge in its article on the decision, the text provides for a mechanism to review AI models before they are released, with a particular focus on frontier models, these cutting-edge models capable of advanced performance in generation, reasoning, coding, or multimodal use. The decisive point, however, lies elsewhere: this is not a binding regime in the traditional sense, but a voluntary federal framework.

This nuance is essential. Since the public explosion of ChatGPT at the end of 2022, Washington has oscillated between two lines: on one side, the idea that the most powerful models need closer monitoring because of risks tied to cybersecurity, disinformation, biology, or military uses; on the other, the desire not to slow a sector that the United States sees as strategic in relation to China. The executive order signed by Donald Trump fits precisely within this tension. It marks a return of the federal executive to AI governance, but in a more flexible form than the one advocated by some lawmakers, AI safety researchers, or federal agencies over the past two years.

The choice of a voluntary mechanism is not insignificant. It reflects the industry’s persistent resistance to any overly rigid ex ante oversight. The major U.S. labs, whether OpenAI, Anthropic, Google DeepMind, Meta, or xAI, have multiplied public commitments on safety, red teams, evaluations, and guardrails. By contrast, they have often been more reserved when it came to formal obligations, systematic transmission of technical information to the state, or regulatory blocking before deployment. The new text therefore attempts a compromise: reintroduce a form of prior review without triggering a head-on confrontation with the players that now dominate the global ecosystem.

This decision does not emerge in a vacuum. It comes after a period of several years in which generative AI became at once an industrial priority, a national security issue, and a diplomatic matter. The United States had already gone through a first phase of more active governance under the previous administration, with the 2023 executive order on AI, which notably used the Defense Production Act to require certain information reporting on the most powerful models. The text signed by Donald Trump does not mechanically reproduce that architecture. It shifts it. Where the previous approach relied more heavily on administrative obligations and more explicit reporting requests, the new approach returns to a logic of supervised self-regulation.

For European observers, the interest of the executive order goes far beyond U.S. domestic politics. The major labs operating from San Francisco, Seattle, New York, or London release their models worldwide. Their compliance practices, evaluation processes, and publication standards have a de facto extraterritorial effect. As a result, even a non-mandatory framework decided in Washington can weigh on the conditions for launching future models in Europe, including on how companies align their U.S. obligations with those of the European AI Act.

The signal being sent is therefore twofold. On one hand, Washington confirms that it sees frontier models as a distinct category deserving specific attention before being placed on the market. On the other, the Trump administration indicates that it favors light federal intervention, compatible with the sector’s competitiveness. This balancing act between safety and innovation is not new. But it now takes on a more concrete significance, because the latest models are no longer just technological demonstrations: they are integrated into office suites, search engines, cloud infrastructure, software development tools, and increasingly into decision-making chains within companies.

What the executive order provides for: review before release, focus on frontier models, and voluntariness

According to The Verge, the core of the executive order signed by Donald Trump is to organize a model review process before publication, with particular attention paid to the most advanced systems. The text follows the line of debates around so-called frontier models, meaning those whose capabilities are high enough to justify a specific review of risks. In the vocabulary of U.S. public policy, this notion generally refers to models trained with massive computing resources, deployed at large scale, and liable to be misused in sensitive areas.

The voluntary nature of the mechanism is the most commented-on element. Concretely, this is not a prior regulatory clearance comparable to a formal administrative authorization. The federal government is not, at this stage, giving itself a general power of systematic prohibition before commercial release. The mechanism relies instead on a strong incentive to submit models for review, within a framework coordinated with federal authorities. This architecture allows the administration to say it is acting on the safety of advanced models, while avoiding the image of a central watchdog blocking innovation.

This choice responds to a political and economic reality. Major AI companies have argued in recent months for rules targeted at high-risk uses rather than at the models themselves. They have also stressed the speed of iteration in the sector. In a market where a major new version can be released in a few months, or even a few weeks, heavy administrative control could be seen as a strategic handicap. U.S. labs put forward a simple argument: if domestic regulation slows launches too much, foreign competitors, whether Chinese, open source, or based in other jurisdictions, could gain the advantage.

The executive order therefore tries to get around this opposition by relying on voluntary commitments. This method is not new in Washington. In July 2023, several major AI players, including OpenAI, Google, Microsoft, Anthropic, Meta, and Inflection, had already announced voluntary commitments to the White House on watermarking, safety testing, and disclosure of certain system limitations. The novelty here lies in the fact that review before release is explicitly reintroduced as an element of a federal framework, at a time when the public debate had shifted toward the states, the courts, and sectoral regulators.

The text primarily targets frontier models because they are the ones concentrating the strongest concerns. Since 2023, several reports, including some from the labs themselves, have highlighted risks of dual-use capability: assistance with malware development, automation of phishing campaigns, support for sensitive biological research, generation of misleading content at large scale, or optimization of offensive cyber operations. Even if some of these scenarios remain debated, they have been enough to produce a minimal consensus: not all models are equal from the standpoint of national security.

Voluntariness does not prevent a certain degree of political pressure. In the United States, when a federal framework is officially presented as the expected best practice, large groups have an interest in complying with it, if only for reasons of reputation, relations with agencies, and management of future legal risk. A lab that ostentatiously refused to participate in a review process backed by the White House would expose itself to immediate criticism in Congress, in the press, and among its business partners. In that sense, the voluntary nature does not amount to an absence of constraint. It instead creates a market norm likely to shape the behavior of dominant players.

One central question remains: who will evaluate what, according to what methodology, and with what level of transparency? This is where implementation details will be decisive. Over the past two years, the industry has developed its own tools: internal and external red teaming, dangerous capability evaluations, alignment testing, cybersecurity audits, data supply chain reviews, documentation of limitations, and publication of system cards. But these practices are heterogeneous. Anthropic, OpenAI, Google DeepMind, and Meta do not publish the same metrics, the same thresholds, or the same procedures. If the executive order merely endorses this diversity, its real impact could remain limited. If it pushes toward minimal standardization, it could instead weigh on the market for the long term.

The long U.S. history: from self-regulation to national security

To measure the scope of the executive order, it must be placed back in the recent history of U.S. AI governance. For years, the United States favored a relatively liberal approach. Under the Obama administration and then at the beginning of Donald Trump’s first term, AI was mainly seen as a driver of innovation, productivity, and competitiveness. Strategic documents emphasized research, cloud, semiconductors, and talent. The question of guardrails already existed, but it remained secondary compared with the technological race.

The landscape changed after 2022. The meteoric success of ChatGPT, which reached 100 million monthly users in about two months according to the widely cited UBS estimate at the time, turned generative AI into a major political issue. Congressional hearings multiplied. Sam Altman, Jensen Huang, Dario Amodei, Sundar Pichai, Satya Nadella, and other executives were called on to comment publicly on the risks and benefits of their technologies. At the same time, concrete incidents fueled concerns: hallucinations in professional contexts, generation of fake political content, guardrail bypasses, growing use of AI in cybercrime, and questions about market concentration around a handful of compute providers.

The Biden administration had then sought to structure a federal response. Its 2023 executive order on AI was more interventionist. It notably requested safety reports for certain advanced models, relying on existing legal instruments. It also encouraged the development of standards by NIST, the National Institute of Standards and Technology, which had already published its AI Risk Management Framework. This doctrine rested on the idea that a modern state cannot be satisfied with voluntary promises when potentially critical systems are deployed at large scale.

The executive order signed by Donald Trump does not amount to a blank page, but it recomposes that legacy. It acknowledges that advanced models must be reviewed before release, which constitutes substantive continuity. By contrast, it shifts the institutional cursor toward a formula more compatible with American market culture and with industry expectations. This shift is politically consistent with Donald Trump’s recurring deregulatory rhetoric, but it does not mean a complete withdrawal of the state. On the contrary, it shows that even an administration more favorable to business considers it impossible to leave aside entirely the question of frontier models.

This point is important because it reveals a deeper transformation: AI is no longer just a technology sector, it is a sovereignty issue. The United States now sees large models as strategic infrastructure, on the same level as advanced semiconductors, hyperscale data centers, or certain critical software. The debate over review before release is therefore not only a compliance debate. It is also a way of defining who controls the conditions of access to computational and cognitive capabilities deemed sensitive.

This national security logic explains why frontier models are treated differently from more ordinary applications. A marketing assistant, a document summarization tool, and a latest-generation general-purpose multimodal model do not carry the same geopolitical consequences. The more models gain in apparent autonomy, planning capability, scientific performance, or coding aptitude, the more the idea of specific oversight gains ground. Donald Trump’s executive order is part of this rise of strategic logic, even if it expresses it through a language of voluntary cooperation rather than formal control.

It should also be recalled that the United States is not alone in this evolution. The United Kingdom organized the first AI Safety Summit at Bletchley Park in November 2023. South Korea and France took part in the diplomatic sequence on AI safety. The G7 launched the Hiroshima process. The UN took up the issue. The idea of special oversight for advanced models has therefore become transnational. What Washington is changing today is the way of orchestrating that oversight on its own territory: less explicit coercion, more political coordination and normative pressure.

Between the European AI Act and the U.S. model: two philosophies of compliance

For companies operating on both sides of the Atlantic, the contrast with Europe is immediate. The European Union’s AI Act, adopted after several years of negotiation, is based on a structured legal logic, with risk categories, defined obligations, and potential sanctions. The European text is not limited to end uses: it also introduces rules for general-purpose AI models, with reinforced requirements for models presenting systemic risk. Technical documentation, transparency, copyright compliance, evaluations, cybersecurity, reporting of serious incidents: Europe has chosen the architecture of the written rule.

Donald Trump’s executive order, by contrast, illustrates another philosophy. The federal government indicates the direction, sets a framework, identifies frontier models as a sensitive category, but leaves a central place to industry commitments. The gap could be summed up this way: Europe codifies, Washington incentivizes. This does not mean that either of the two systems is necessarily more effective in absolute terms. It does mean, however, that global players will have to build hybrid compliance chains.

For a lab like OpenAI or Anthropic, that probably implies a double effort. On one side, satisfying European requirements when a model is made available in the Union, directly or through integrators. On the other, demonstrating to Washington that a serious prior review process exists, even if it is not imposed in the same terms. In practice, companies may be tempted to align their internal processes with the most demanding standard, then adapt it by jurisdiction. That is often what happens in globalized sectors: the most structuring framework ends up becoming the internal operational reference.

But the opposite effect is also possible. If the U.S. framework remains more flexible and faster, some labs could reserve more documented, more cautious, or more limited versions of certain models for Europe, while continuing to launch faster in the United States. This kind of gap has already been observed in other digital fields, whether data protection, targeted advertising, or certain platform features. AI could follow a comparable trajectory, with differentiated launch timelines depending on compliance constraints.

For French and European companies, this divergence raises a competitiveness question. Large U.S. groups have the legal teams, audit capabilities, and compute resources needed to absorb complex compliance. European startups, by contrast, risk being caught in a squeeze effect: on one side, a more prescriptive framework with the AI Act; on the other, U.S. competitors benefiting from a more flexible domestic environment while also having the financial capacity to comply with Europe when necessary. Donald Trump’s executive order, even though voluntary, potentially reinforces this asymmetry, because it gives U.S. giants room for direct negotiation with the federal state.

That said, this should be qualified. Europe can also benefit from this situation. If Washington officially recognizes the importance of review before release for frontier models, even on a voluntary basis, that partly validates the European intuition that advanced models cannot be treated like simple consumer software. The U.S. discourse is therefore moving closer to the European diagnosis, even if it diverges on the instruments. For Brussels, Paris, or Berlin, this is a useful argument: oversight of cutting-edge models is no longer a uniquely European regulatory singularity, it is a concern shared by the major technological powers.

In the French context, this partial convergence is particularly interesting. France has for several years sought to reconcile industrial ambition and normative influence. With Mistral AI, Hugging Face, LightOn, or players such as OVHcloud and Scaleway on infrastructure, the national ecosystem wants to exist in global competition. But Paris also supports a governance approach that does not leave the market entirely to self-regulate. The U.S. executive order could therefore be read, from the French side, as confirmation that review of advanced models is becoming a central practice, including in a country historically more favorable to innovation without constraints.

What this changes for major labs, cloud providers, and user companies

Beyond the political signal, the executive order can very concretely change corporate behavior. For major model labs, the first consequence is organizational. Review before release, even if voluntary, implies more formalized processes: definition of risk thresholds, testing protocols, external red teaming, documentation of sensitive capabilities, internal decision-making mechanisms on full or partial publication of a model, and coordination with legal and government affairs teams. Many of these building blocks already exist, but the executive order makes them harder to treat as mere internal options.

Labs that have built a public narrative centered on safety could find a competitive advantage in it. Anthropic, for example, has communicated extensively about its Responsible Scaling Policies. OpenAI regularly publishes system cards and safety analyses. Google DeepMind emphasizes its advanced evaluations and security teams. In a world where Washington values prior review, these players can monetize their compliance maturity as a strategic asset. Conversely, smaller structures or open-source projects could fear rising governance costs, even without a strict legal obligation.

Cloud providers are also concerned. Amazon Web Services, Microsoft Azure, and Google Cloud no longer merely host models: they have become structuring partners in their development and release. If the federal government encourages review of models before launch, cloud providers could be led to integrate more contractual controls, usage logs, testing mechanisms, and safety guarantees into their offerings. The precedent already exists in other sensitive sectors, where infrastructure becomes a checkpoint for compliance.

For user companies, the effect will be more indirect but real. Large French and European enterprises that integrate U.S. models into their business tools are increasingly asking for guarantees on safety evaluations, resilience, traceability, and governance. If U.S. labs adopt more standardized review procedures under the effect of the executive order, that could improve the quality of information available to customers. In other words, even a voluntary framework can produce better risk readability in the market.

There is, however, a risk of fragmentation. If each lab continues to define its own thresholds and publish heterogeneous elements, client companies will have to compare documents that are difficult to compare. One player may highlight its cyber tests, another its biological evaluations, a third its progressive deployment policy. Without a robust common standard, the market could end up with an inflation of documentation but little true comparability. This is one of the historical limits of self-regulation: it sometimes produces a lot of communication and not always a shared metric.

The debate is particularly sensitive for open source. Meta, with the Llama family, has contributed significantly to spreading open or semi-open models. Mistral AI has also established itself in Europe by defending a strategy combining openness and commercial offerings. Yet review before release raises a delicate question: how should a model whose weights are published and can be reused freely or quasi-freely by other players be treated? Once the model is released into the ecosystem, control over uses becomes much more difficult. The U.S. executive order, even though voluntary, could therefore increase pressure on publishers of open models to document their publication choices more extensively.

For investors, finally, the message is clear: model safety and compliance are becoming valuation variables. In 2024 and 2025, the amounts committed to AI reached considerable levels, with funding rounds of several billion dollars for some players and colossal infrastructure spending at Microsoft, Alphabet, Amazon, and Meta. In such a context, a lab’s ability to demonstrate that it can launch powerful models without triggering a regulatory or reputational crisis directly affects its attractiveness. Donald Trump’s executive order reinforces this logic by institutionalizing, even lightly, the idea that an advanced model should go through some form of review before release.

What impact for France and Europe, and what the U.S. trajectory reveals in the long term

For the French-speaking market, the main interest of this decision lies in its knock-on effect. Major French companies, administrations, software publishers, and integrators work predominantly with U.S. technologies, directly or indirectly. When Washington changes its expectations regarding frontier models, the shockwave spreads into tenders, security questionnaires, contractual clauses, and deployment trade-offs in Europe. Even if the framework is voluntary, it can become a de facto standard for any player wanting to appear credible in the advanced-model segment.

In France, this effect will be closely watched by players trying to build a European alternative. Mistral AI, valued at several billion euros after major fundraising rounds, is seeking to reconcile speed to market, technical excellence, and regulatory acceptability. If the United States institutionalizes prior review, even in a flexible form, the pressure to demonstrate equivalent procedures will also increase on European champions. This could paradoxically favor already structured players, to the detriment of smaller teams unable to absorb the costs of testing, documentation, and governance.

For French and European authorities, the U.S. executive order also offers a diplomatic foothold. For several years, the recurring argument against more demanding regulation in Europe has been that it would isolate the continent in the face of a more permissive world. Yet if the United States itself recognizes the need for review before release of advanced models, even in a voluntary form, it becomes harder to argue that any oversight of frontier models would by nature be anti-innovation. The debate then shifts to the quality of the mechanism: what thresholds, what evidence, what agencies, what transparency, what articulation with the market?

In the long term, the most likely outcome is the emergence of a mixed regime. The most powerful models will be subject to a combination of legal obligations, voluntary commitments, technical standards, and commercial pressures. The United States seems, for now, to be choosing a gradual ramp-up: first commitments, then voluntary federal frameworks, then possibly more targeted obligations if a major incident occurs or if Congress takes up the issue more directly. The history of U.S. technology regulation often follows this trajectory. Markets experiment, the executive incentivizes, then constraint strengthens when risks become politically impossible to ignore.

That is no doubt the main lesson of the executive order signed by Donald Trump. Voluntariness should not be read as a simple renunciation. It can also be interpreted as a pre-normalization phase. By establishing the idea that a frontier model deserves review before release, Washington is creating a lasting expectation. Labs that comply with it today will more easily prepare for a future mandatory framework. Those that diverge from it will take the risk of appearing out of step with the emerging norm. In strategic industries, today’s voluntary standards often become tomorrow’s minimum obligations.

For Europe, and France in particular, the question is therefore not only whether the U.S. framework is more flexible than the AI Act. It is to understand how that flexibility will influence global competition. If the United States manages to combine speed of innovation, federal coordination, and credible safety standards, its model could appeal to part of the international ecosystem. If, on the contrary, voluntariness produces practices that are too disparate or too opaque, the European approach will gain legitimacy as a more stable reference. The regulatory balance of power will depend as much on execution as on the texts.

From this perspective, Donald Trump’s executive order marks less an endpoint than a strategic repositioning. Washington is not withdrawing from AI governance; it is returning with a method that seeks to preserve industrial initiative while responding to national security concerns. For global labs, this means that the race toward ever more powerful models will no longer be decided only by the number of parameters, GPUs, or benchmarks, but also by the ability to prove, before release, that a system can be understood, tested, and governed. And if this logic takes hold for the long term, the decisive boundary in the market will no longer separate only the best models from the rest, but the players capable of turning safety into a competitive advantage from those that continue to treat it as a mere communication variable.