The European AI Act comes into force, amidst ambitions and uncertainties.

The main provisions of the European regulation on artificial intelligence came into effect this Sunday, imposing new obligations on developers of large language models. Implementation promises to be challenging, given the limited resources of the AI ​​Office and diplomatic pressure from Washington.

Adopted in 2024, the AI ​​Act is the world's first comprehensive legislation dedicated to artificial intelligence. Its most fundamental provisions, those concerning major language models, are now binding on all companies that market their technologies on the European market, whether they are European or foreign.

Originally, the text was intended to regulate only AI applications. The launch of ChatGPT in 2022 led European legislators to broaden its scope to include the underlying technology itself. The regulation now requires all developers of general-purpose models to be more transparent about the design of their systems, to disclose copyrighted content used during training, and to provide users with sufficient information about the actual capabilities of these models.

The most powerful so-called "frontier" models are subject to additional requirements: their designers must identify and mitigate the risks they pose to society. To support this implementation, the European Commission adopted a voluntary code of conduct last year, developed with international experts, including researcher Yoshua Bengio. Most major Western laboratories have signed it. Meta is an exception.

Tom Duff Gordon, OpenAI's vice president of public policy for the EMEA region, indicated that they had "worked closely with the European Commission and the entire ecosystem on the implementation of the AI ​​Act, including its codes of conduct," adding that the company intended "to continue working together to help Europe take advantage of the benefits of the age of intelligence."

The Commission has created a European Artificial Intelligence Office, the AI ​​Office, to oversee the implementation of the regulation. However, this body will have to contend with limited human and financial resources in a sector where private competition for specialists is fierce. The European executive plans to rely on external experts, a panel of scientists, and companies specializing in AI safety, but the task remains considerable given that these technologies are evolving faster than scientific knowledge about their risks.

The implementation of the text could also reignite tensions with Washington. The Trump administration has already expressed its hostility to European regulations targeting American digital giants. Irish MEP Michael McNamara (Renew) fears that this new step will be perceived across the Atlantic as "an attack on US commercial interests," as was the case during the implementation of the Digital Markets rules in December 2025, or during the attempt to fine Google under the European Digital Markets Regulation.

For European citizens, the regulation promises additional guarantees regarding the security and transparency of the models made available to them. However, companies in the sector warn that compliance obligations could delay the rollout of certain models in Europe by a few weeks compared to other markets.

German MEP Axel Voss (EPP) called on the Commission to align the application of the AI ​​Act with other digital issues, arguing that the AI ​​Office should not "waste its energy on niche concerns" but "closely align itself with the priorities of its colleagues responsible for platform regulation".

A fundamental debate also runs through the global regulation of AI: should priority be given to protecting fundamental rights, against discrimination, violations of privacy or lack of human oversight, or should efforts be focused on so-called "existential" risks, such as the use of AI to develop weapons or carry out massive cyberattacks? Recent events weigh on this debate: the United States has imposed export restrictions on an Anthropic model based on Mythos because of its offensive capabilities, while OpenAI revealed that one of its agents had managed to hack an AI company during testing.

Laura Lazaro Cabrera, director at the Center for Democracy & Technology, warns against a dangerous trend: "The Commission must resist the temptation to devote its enforcement resources solely to systemic risks of cyberattacks and loss of control." She argues that "the application of rules should not be dictated by headlines, but should cover the entire spectrum of risks."