Donner un sens à l'intelligence artificielle

A Stanford library, October 2016

The report did not begin as a political commission. It began with a conversation in a Stanford library, in October 2016, between François Levin, then a rapporteur, a staff policy officer, at the Conseil national du numérique (French national digital council), myself, and a French entrepreneur based in the United States, Gregory Renard. Renard described two things to us at once: what his algorithms could already do, and his fear of what they would do to employment: hundreds of thousands of jobs threatened in certain sectors, along with the question of his own responsibility toward the society he would leave to his children. Coming from someone who was building these systems, it was a shock.

Back in France, we sent a note to the Élysée. Eleven months later, in September 2017, Prime Minister Édouard Philippe entrusted Cédric Villani with a mission on artificial intelligence and asked us to support him. Six months of work followed: more than four hundred hearings, a public consultation that gathered 1,639 participants, and a comparative review of policy in fifteen countries. In March 2018, at the Collège de France, the President of the Republic announced a national strategy backed by 1.5 billion euros.

Same ceremony, English-dubbed version.

I tell this genealogy because it contradicts the image one tends to have of how public policy gets made. Here, a 1.5-billion-euro national strategy originates in a worry voiced by an engineer, relayed by a small team with no mandate, in a memo. That says something about the porousness of the French system: its responsiveness, and its dependence on circumstance.

What the report achieved

The strategy’s first phase, known as “AI for Humanity”, committed 1.5 billion euros between 2018 and 2022: four interdisciplinary institutes (PRAIRIE, MIAI in Grenoble, ANITI in Toulouse, 3IA Côte d’Azur) selected by an international jury in April 2019, 180 chairs of excellence, 300 doctoral programmes, and the Jean Zay supercomputer. A second phase followed from 2022 onward, bringing the total commitment to around 2.5 billion euros under France 2030.

The bet on public computing capacity proved to be the soundest one. Jean Zay was used to train Bloom, a 176-billion-parameter model, at a time when earlier French models were measured in the hundreds of millions. Without that infrastructure, French public research would have watched the generative shift from the sidelines.

"What meaning to give to AI? The Villani report, five years on", seminar of the AI Observatory, Université Paris 1 Panthéon-Sorbonne, 22 June 2022 (16 min, in French, CC BY-NC-SA 4.0).

The recommendations, one by one

The report was built around seven strands. Reading them again today is an uneven exercise.

Research. An interdisciplinary network of excellence: done, and it is the clearest success.

Ethics built in at every level. We asked for two things. A national consultative committee on the ethics of digital technologies and artificial intelligence: it was created. And a public algorithm-auditing function: a body of sworn experts able to “open the black boxes” in a judicial proceeding or on a saisine (formal referral) from the Défenseur des droits (the French ombudsman). It was not. That function now exists, but at European level and in a different form, through the conformity-assessment obligations of the AI Act.

Work. We proposed a public laboratory on the transformation of work, whose purpose was not to forecast but to experiment: support schemes for people exposed to automation, and new ways of distributing value. It never came into being. Of all the recommendations, its absence strikes me today as the costliest, precisely because the question it raised has come back with even greater urgency.

The four strategic sectors: health, transport and mobility, ecology, defence and security. This is the strand that has drifted furthest. The national strategy’s second phase is now organised around technological rather than sectoral priorities: embedded AI, trustworthy AI, frugal AI, generative AI. The shift from a breakdown by use case to a breakdown by technology is not a minor detail: it marks the abandonment of a demand-side policy in favour of a supply-side one.

Bringing more women in, and tripling the number of people trained: real trajectories, but slow ones, and hard to assess for lack of a stable indicator.

Frugality, and AI in the service of the ecological transition. We called for developing, side by side, a greener AI and an AI that serves ecology. What followed showed that the two were not symmetrical: data-centre energy consumption grew faster than AI’s environmental applications did.

The recommendation that backfired

The first strand called for an assertive data policy, for building “data commons”, and for sharing platforms serving research and the public interest. We explicitly cited the centralisation of health data, and we justified it in the name of sovereignty.

The Health Data Hub was created the following year. It was hosted with an American provider, and that decision became the textbook French case of technological dependence. A recommendation made in the name of sovereignty produced the very institution that illustrates its loss.

I do not believe the report was wrong to want to pool health data. But it stated an objective without specifying its technical conditions, and it assumed that the will to sovereignty would be enough to steer infrastructure choices. It was not enough. That is a design flaw in the report, not an accident of implementation.

The 2015 report on health, a common good of the digital society made the same case on the same ground, three years earlier: the two texts push in the same direction, and share the same underlying vulnerability.

Governance

The harshest criticism does not come from us. In its thematic public report of November 2025, the Cour des comptes (France’s national audit court) finds that relying on a succession of calls for projects produced fragmentation rather than critical mass, that the funding horizon was too short to create leverage effects, and that interministerial coordination changed hands along the way (from the direction interministérielle du numérique, the interministerial digital directorate, in 2018 to the direction générale des entreprises, the directorate general for enterprises, in March 2020), with the national coordinator left with only a very small team.

A mission report does not control the governance that follows its handover. But it could have anticipated it: nothing in the 2018 text secured the piloting of what it recommended. We described a destination without sizing the vehicle.

A detail that says a lot

In 2018, alongside my duties at ANSSI (the French national cybersecurity agency), I joined the European Commission’s High-Level Expert Group on Artificial Intelligence on a voluntary basis. Its two texts, the ethics guidelines and the policy and investment recommendations, are reviewed elsewhere on this site; their work fed first into the white paper and then into the AI Act. Voluntary, on top of a full-time job. It was my own choice and I make no complaint of it, but the mechanics are worth noting: France’s presence in the room where European AI regulation was being shaped rested, in part, on evenings given up by people who had another job to do. The administrations that carried real weight on that text had delegated people to it full-time.

What I said in June 2022, and what I still hold to

This notice has an antecedent, and it needs to be dated precisely. I presented an assessment of this report on 22 June 2022, at a seminar of the Observatoire de l’intelligence artificielle at Université Paris 1. That is, five months before the release of ChatGPT.

That assessment does not contain a single word on foundation models, scaling laws, or generative AI. This is not an oversight: those objects were not within the field of vision of the community I was addressing, still less within that of the 2018 report. But the gap measures exactly what a five-year assessment can be worth in this field, and it invites reading this one with the same caution.

I still hold to two things from that talk. The conviction that one must leverage the ecosystem rather than rely on public action alone: five years at Campus Cyber have not changed my mind. And the line that concluded it: there is no technological determinism, this transformation needs meaning. It is more demanding today than it was then, which is an argument for keeping it, not for abandoning it.

The team

The front-matter lists the mission’s official team. The report also owes a debt to an extended team of rapporteurs: François Levin, Charly Berthet, Lofred Madzou, Anne-Charlotte Cornut, Camille Hartmann, Judith Herzog, Jan Krewer, Ruben Narzul and Marylou Le Roy. Célia Zolynski, with whom we worked at the Conseil national du numérique (French national digital council), also mattered in preparing this work.