All summer long, the Grand Continent will stay in motion. Every day, wherever you are, we will bring you ideas you won’t find anywhere else (but which will be everywhere come autumn), texts that are scarce or refreshing with our Sundays. To receive them directly in your inbox and support this momentum, consider subscribing to the magazine
The AI Act is neither a Brussels folly nor a fortunate intuition. It fits into the long narrative of the information revolution, which has run since 1945. Artificial intelligence is not a rupture, but the fourth pillar anticipated, prepared for, and foreseen in this radical transformation.
An Evolution, Not a Revolution
The first pillar is the computer: the ENIAC, completed in 1945 to calculate artillery firing tables for the American army, whose first major task was precisely to study the feasibility of thermonuclear weaponry, followed by the great computers that succeeded it. The second, around 1970, is the mini-computer that moves computing out of the central computation facility into the enterprise, the laboratory, and the university. The third, around 1995, is the connected PC, Internet, the Web and, as a consequence, the accumulation, over twenty-five years, of humanity’s first global information heritage: our personal data, those of our companies, our administrations and scientific knowledge. The fourth pillar, which generative AI has operationally entered since 2022, is the large-scale exploitation of this heritage: mining the accumulated data for patterns, knowledge, and regularities that allow us to anticipate solutions to current and future problems.
This fourth pillar could well be the last of the information revolution. It is by no means a surprise; it is the mechanical consequence of the three preceding pillars. It is in this spirit, that of an expected evolution rather than an unpredictable revolution, that the AI Act was designed. It rests on a clear pyramid architecture: we prohibit what is plainly contrary to human dignity (social scoring, subliminal manipulation, recognition of emotions at school or work); we tightly regulate what touches fundamental freedoms (biometrics, education, employment, justice, essential services); we impose transparency where illusion threatens public truth (chatbots, deepfakes, synthetic content); and we leave free the rest, that is, the vast majority of uses.
Four Years in the Making
The preparatory work lasted four years, from 2020 to 2024. The final trilogue concluded in thirty-six hours in Brussels, during the night of December 8–9, 2023. Those thirty-six hours crowned a preparation dense with foresight: hundreds of hearings, dozens of researchers mobilized, technical exchanges with model providers, discussions with interested governments, sector authorities, industry, the academic world, creators, trade unions, and NGOs.
Adopted by the European Parliament on March 13, 2024 by 523 votes for, 46 against and 49 abstentions, i.e., nearly 85% of the votes cast across all parties, from the right to the left, including liberals, Greens, and up to seventeen MEPs of the National Rally, with only the deputies of France Insoumise voting against. It earned unanimous approval from the Twenty-Seven States at the Council on May 21, 2024. No text regulating technology has ever achieved such broad consensus among its co-legislators in any democracy in the world.
According to the original schedule, August 2, 2026 was to mark the entry into force of obligations for high-risk systems. The fight waged for eighteen months decided otherwise. A team of American lobbyists, a few European executives, and several capitals, including Paris, directly weighed to attempt to weaken the obligations applicable to foundation models. On June 29, 2026, the Council finally adopted a set of rectifying amendments: the new “Digital Omnibus on AI.” The obligations of Annex III are thus postponed to December 2, 2027 and those of Annex I to August 2, 2028. People are still wondering why such concessions were made to the American giants, two years after voting these rules unanimously.
This postponement is a retreat. But what does come into force on August 2, 2026 is nonetheless significant, in another domain. It must be measured carefully before we can judge what we have to defend overall.
Article 50 Is Not the Heart of the AI Act
After carrying this regulation for four years, I think it useful to provide the following clarifications.
Article 50, which is now in force, concentrates now all the attention. It is useful, even essential: it protects the informational space from disinformation, it prevents opaque chatbots, deepfakes, and synthetic texts from passing off as human speech.
But the heart of the AI Act lies rather in the absolute prohibitions set out in Article 5 and in the regulation of high-risk systems. The objective is to protect individuals against the most dangerous AI applications, such as decisions about admission, credit, hiring, care, policing, or justice. It also aims to protect rights holders, creators, artists, writers, composers, and scientists, whose intellectual property forms the raw material of the major models. This is why this regulation was designed, voted, and adopted, and also why it is called AI Act and not a regulation on transparency of content generated by LLMs.
Contrary to what one might think from the General Data Protection Regulation, the AI Act was not designed to address plagiarism issues, which fall under other bodies of law (copyright, intellectual property, academic and editorial ethics). Confounding it with an anti-plagiarism tool would therefore be a double error: in terms of text and in terms of purpose.
The two-year delay achieved with the Digital Omnibus is a derailment. It is not merely an administrative delay, but a shift of the regulation’s center of gravity. By keeping Article 50 as the sole public face of the AI Act, while the provisions concerning risk protection and rights protection would not be applied before 2027–2028, the Commission — by weakness, by concession to American pressure — presents the peripheral and buries the essential, reducing a civilization regulation to a labeling regulation.
Article 50 also has a technological limit. Automatic tagging of texts (statistical watermarks, cryptographic fingerprints, signatures) is technically circumventable. Tomorrow, free tools that erase marks and rewriting services to escape detection will flourish. This battle, technology alone will lose. It is not the machine that will guarantee the truth of the informational space, but identifiable human responsibility. That is precisely why the editorial exception in Article 50, paragraph 4, was included, not as an escape hatch, but as the legal recognition of what alone endures: professional ethics that must prevail.
The lawyer, the architect, the investment banker, the physician, the notary, the professor, the editor-in-chief, the publisher, and the journalist are first and foremost bound by their order, by their charter, by their code, by their own law: medical ethics, Hippocratic oath, Ventôse Law for notaries, the order of architects and ten-year liability, duty of advice, pedagogical authority and academic freedom, the 1881 law, editorial responsibility. In all these professions with very high intellectual added value, where ethics is an indispensable guarantee, the machine is a tool; the responsibility remains human, full, entire, and personal.
This point is all the more important as it is necessary to encourage these professions to use AI, because it will make them more productive, free up their time, improve their services, their thinking and their creativity, and give them unprecedented analytical capabilities. The AI Act was never designed to prevent this usage. It was designed so that it could unfold under the ethical rules each profession will set for itself. Letting Article 50 dominate the scene for two years, while the heart of the regulation is relegated to the background, risks making AI appear as a mere labeling tool to our fellow citizens before it has produced its real effects.
The Misinformation from the Anti-AI Act Lobby
This detour in perspective cannot be understood without placing it in a broader narrative, that of a lobby meticulously sustaining itself for eighteen months to weaken the AI Act as a whole. According to this narrative, the AI Act would have become a bureaucratic straitjacket, a machine to kill European innovation, and a gift to American giants, because we would be the only ones imposing regulation. This narrative rests on four claims that must be debunked:
- First, “The AI Act stifled European innovation.” This claim is absurd because the AI Act only took effect ten days ago. It has therefore materially had no chance to stifle anything. The major foundational models (OpenAI’s GPT, Anthropic’s Claude, Google’s Gemini, DeepSeek, and Mistral in Europe) all developed between 2020 and 2025, a period during which no binding obligations of the AI Act applied. If Mistral is not as large as OpenAI, it is not the Commission’s fault but the result of a fragmented European capital market, risk aversion, and the absence of deep funding to train models at the highest level. Confounding these two elements, regulation and capital shortfall, is either bad faith, incompetence, or lobbying possibly remunerated.
- Then, “Europe regulates alone; it penalizes itself.” Europe is indeed the only region with a democratically adopted horizontal binding framework. This is precisely its advantage. In the United States, Joe Biden’s October 30, 2023 decree was repealed by Donald Trump on January 20, 2025. There is no equivalent federal framework in the US. China has scattered sectoral regulations oriented toward political control of the population and mass surveillance, not toward protecting individual rights. No major democracy today grants its citizens the right to obtain an explanation for an automated decision (Article 86) or the right to file a generalized complaint (Article 85), as we do.
- Next, “The Draghi report calls for abandoning the AI Act.” The September 2024 report calls for an additional €300 billion per year in investment and better regulatory coordination. It does not advocate dismantling fundamental safeguards at any point. The instrumentalization that has been made of it is a misreading.
- Finally, “The AI Act is a new GDPR that will kill European businesses, as GDPR killed them.” Contrary to the legend spread from Palo Alto, GDPR did not kill any European business. It actually imposed Europe’s data protection standard on the entire world: GAFAM, Chinese, Brazilian, Japanese, and Korean companies complied with it and often adopted it for their own markets. It created a European compliance market, formed by tens of thousands of data protection officers, and gave Europeans control over their data and information heritage, with an unprecedented right to recourse. Those who today repeat almost word for word that GDPR would be a handicap for Europe are echoing a narrative constructed across the Atlantic precisely because it disrupts. The AI Act is destined to become a global standard in the same way. This is why it disturbs, for the same reasons, the same actors.
To these four pieces of misinformation circulating in public debates, two more arguments must be addressed.
The first is a one-sentence claim: “It is absurd to regulate a technology that has just been born and whose uses are not yet known.” This sentence, so frequently repeated, betrays a deep misunderstanding of the information revolution. Indeed, generative AI is not a technology that sprang from nowhere. It leverages an information heritage consisting of data with origins, contexts, owners, and initial intent. The rights of artists and writers, patient health data, media archives, scientific publications, and the images of our faces all had a status. Regulating AI is first and foremost about establishing the rules for reusing this collective heritage.
The second claim also comes in one sentence: “This isn’t about regulating the unknown. It’s about protecting the known,” a refrain heard with unwavering consistency: “Europe’s habit is to regulate before innovating. We will innovate before we regulate.” At VivaTech in June 2025, the French head of state himself stated: “The risk is regulating things we will not do or cannot do.” All these statements stem from the same misunderstanding.
Waiting for uses to regulate an exponential technology is regulating when it is already too late. That was the error with social networks, due to the absence of a DSA a decade earlier — and we pay the price today, both for our democracies and for the mental health of children. The AI Act is precisely the opposite. It is anticipatory regulation of an announced tsunami. This political foresight of an almost instantaneous surge is unprecedented in European legislative history and deserves to be hailed as such.
It is precisely this capacity for anticipation that Europe should have highlighted at the opening of the World Conference on AI, on July 17, 2026 in Shanghai. Xi Jinping delivered a speech around an explicit proposal, in the presence of the United Nations secretary-general: “to build a just and fair system for the governance of AI worldwide.” China, which has no horizontal framework, no right to explanation, and no right to individual complaint, is now positioning itself as a world governance architect for AI. Meanwhile, Europe, which has the only framework in the world that meets this ambition, was underrepresented, mute, and without an articulated counter-proposal.
This silence is incomprehensible. We should take the AI Act to Shanghai, to New Delhi, to the United Nations, to the halls of the G7 and G20, as France once carried the Paris Agreement on climate. Instead, at the Berlin summit on November 18, France officially asked for a postponement of regulation, and at VivaTech, the President of the Republic allowed Jensen Huang, Nvidia’s chief, to criticize European regulation without responding. The von der Leyen Commission appeared unable to defend its own achievement in the face of the joint pressure from Washington and major American suppliers.
Why the United States Wants to Undo the AI Act
Behind these pieces of misinformation lies a reality. The Americans are embarrassed — and “jealous” — of this democratic regulation. The AI Act is near becoming what the GDPR was yesterday, and to a lesser extent what the DSA is today: a tool of global soft power. A standard exported, imitated, and transposed. It translates into positive law what no other democratic system has been able to articulate with such clarity: the possibility to regulate the information space intelligently without harming the freedom to do business, while applying our republican, democratic, societal, and living-together values, as well as the rule of law principles.
The United States can no longer do so. The fragmentation and polarization of their political class, the decisive power of large private actors in public debate, and the weakness or the interest of an executive that has mediated everything to a few groups defending only their profits, now prevent them from producing a general-interest regulation of the information space.
That is precisely what J. D. Vance came to denounce in Paris, on February 11, 2025, at the height of the AI Action Summit, when he stated that the United States “dominate” AI and “intend to continue doing so” by implying, in effect, that Europe should abandon its regulatory framework. It is important to measure what this means. That the number two of a foreign government publicly demands the unraveling of such a text constitutes unacceptable interference in the European democratic process. We would not have tolerated it from any other power; there is no reason to tolerate it from this one either.
That is precisely what Meta sought to weaken by publicly refusing to sign the GPAI Code of Good Practice.
It also explains the personal attacks I endured throughout my tenure, up to my departure from the Commission in September 2024, and even after (with the ban on entering U.S. territory), because defending the AI Act and the European information space was defending a doctrine that Washington could no longer tolerate.
Nevertheless, we must go further. We are entering the monetization phase of artificial intelligence. After the euphoria of technological demonstration comes the time of the business model. The dominant American players have invested hundreds of billions of dollars to train their models, build their data centers, store and exploit data; they must now find profitability. Facing them, Chinese players, largely backed by the Beijing administration, propose open, low-cost or even free models, precisely to capture global usages. The European market is one of the most lucrative in the world: it is solvent, urban, educated, digitalized, and governed by law. It is easy to see what is really at stake. It is not the simplification that is at issue. It is our ability to defend our market, our data, and our citizens against a wave that would finance itself at the expense of Europeans. It is precisely for this reason that Europe must maintain its democratic regulation.
A Universal Progress, Aligned with the Magisterium of Pope Leo XIV
Against these concessions, it is important to recall the principles defended by the AI Act. This is not a technical text. It is an architecture of values translated into positive law: the dignity of the human person, fundamental freedoms, democratic integrity, and the truth of the informational space above economic optimization. Responsibility must remain human, traceable, and identifiable.
This architecture converges strikingly with the papal encyclical Magnifica Humanitas, by Pope Leo XIV, which is entirely devoted to artificial intelligence and human dignity. The Pope does not condemn AI, but calls for measuring its environmental effects, denounces the invisible work of annotators and moderators, rejects the idea that an algorithm could render war morally acceptable, and recalls, in the continuation of Antiqua and Nova, that “only the human being is truly a moral agent.”
This convergence is not by chance. It expresses a simple truth: safeguarding the human person against machines is not a European cultural preference. It is a universal moral requirement.
Next Steps. What No Omnibus Should Remove
The schedule of upcoming milestones spans two years.
Thus, the absolute ban on non-consensual sexual deepfakes will take effect on December 2, 2026, before December 2, 2027 marks the new deadline applicable to autonomous high-risk systems listed in Annex III. We will then await August 2, 2028 for high-risk systems integrated into products already regulated under Annex I, as well as general-use AI models placed on the market before August 2025, to come under regulation. Concurrently, the standardization work of CEN-CENELEC JTC 21 will continue, as will the annual revision of Annex III provided for in Article 112.
To these timelines, we must add — and not remove — several workstreams: specific protection for minors; a framework for autonomous agents that the regulation did not address; obligations for click workers and annotators, which Leo XIV spoke about better than any regulator; quantified environmental reporting of training infrastructures; a clear regime for political deepfakes.
What is at stake is not merely a regulatory deadline. It is proof that Europe is the only region in the world capable of both protecting and innovating, of guiding and liberating, of regulating without stifling. To betray this achievement in the name of competitiveness would be to renounce Europe’s ability to define alone, for ourselves and for those who wish to draw inspiration, what we accept and what we refuse in our relationship to the digital world and in our own informational space.
This sovereignty is not negotiable; it must be exercised.