Matignon buries an AI report. Because it isn't scary enough.

Matignon buries an AI report. Because it isn't scary enough.

You usually picture a state secret as a way to hide a disaster. Here it's the opposite. Matignon has just locked down a report on the impact of artificial intelligence in the civil service, and the reason making the rounds is that its conclusions aren't alarming enough to be shown.

The story comes from Acteurs Publics, on June 18. A cross-inspectorate report has been sitting in Prime Minister Sébastien Lecornu's drawers for several weeks. A decision was made, "at the highest level," not to release it. And the detail that changes everything: the inspectors who wrote it were themselves asking for it to be published.

A report commissioned, delivered, then locked away

The mission was anything but improvised. It brought together the General Inspectorate of Social Affairs and the General Inspectorate of Finance, two of the administration's most serious corps, later joined by the General Inspectorate of the Administration. The work had been commissioned under the Bayrou government, well before the subject turned sensitive.

The inspectors did their job. They handed in their copy. And according to Acteurs Publics, they found themselves "subject to strict confidentiality instructions," even as they wished to see their conclusions made public.

So here's the situation: experienced civil servants produce a quantified analysis at the state's request, recommend publishing it, and the state imposes silence on them. The party that commissioned it has every right to keep a report to itself. That right exists. The question is why it exercises that right when the content is, by one source's own admission, "not alarmist."

Why hide good news

This is the counterintuitive part. A report saying "AI will devastate public employment," you'd understand the temptation to bury it. But a reassuring one?

The answer lies in a calendar. We're less than a year from the presidential election, and a few months from the civil-service professional elections. In that context, a number is no longer a number. It's ammunition.

Too low, it feeds those who want to slash headcount ("see, AI lets you do the same with fewer people"). Too nuanced, it serves those who denounce a hidden cuts plan. Whatever the result, someone seizes on it.

A public report is like a card laid face-up on a poker table: everyone sees your hand and adjusts theirs. Matignon prefers to keep its cards face down until the table is less dangerous. The caution makes sense politically. It poses a democratic problem: the information exists, it's funded by public money, and the citizen has no access to it because the timing is inconvenient.

The gap that makes the silence awkward

To grasp why these figures bother people, you have to look at the chasm in the official line on headcount.

On one side, the Lecornu government advertised a net cut of 3,000 jobs for 2026. On the other, when you read the budget documents stripped of accounting effects, the real balance across ministries is firmly positive: more than 8,000 additional full-time equivalents, including 5,400 in Education. Acteurs Publics calls the gap between the advertised figure and reality a fudge.

In the middle of that fog, a report that puts in black and white what AI can or can't do to headcount becomes explosive, in either direction. All the more so since orders of magnitude exist elsewhere: the Roland Berger consultancy estimated that "more than a third of public jobs would be exposed, over time, to significant changes" tied to generative AI. Exposed to changes is not the same as eliminated. But on a campaign poster, the nuance doesn't survive.

The public line, for its part, is well rehearsed. The executive's message repeats that AI isn't meant to cut headcount, that it should "free up time spent on paperwork" to reinvest in human relationships. The sovereign ChatGPT rolled out for agents since 2025 is framed this way. A quantified report, whether it confirms or qualifies that line, forces officials out of their talking points. That's precisely what you avoid a year before an election.

Meanwhile, in Brussels

The contrast jumps out when you look across the administrative border. The European Commission deployed its own large language model, GPT@EC, open to its entire staff, with the code set to go open source in 2026. The institution that regulates AI with the AI Act publicly owns its in-house tool and talks about it.

Paris, by contrast, is negotiating a framework for AI use in the civil service with the unions by autumn, while refusing to put on the table the study that would measure its effects on employment. Union representatives in fact discovered a draft agreement already wrapped up at their June 18 meeting. Hard to negotiate blind when the state holds figures it keeps to itself.

What the gesture reveals

The real subject is the reflex more than the report itself. Faced with data on AI and public employment, the machinery of state treated the information as a risk to neutralize rather than an element of debate to hand to the public. Transparency has become an adjustment variable of the electoral calendar.

It's exactly the blind spot we found the same day in Jeff Bezos's speech on the labor shortage: refusing to put the figures on the table. Bezos does it out of commercial interest, because he sells AI and a narrative of abundance serves his startup. Matignon does it out of political caution, because a number becomes uncontrollable in an election year. The motivations are opposite, the result is the same.

Everyone now has an opinion on AI and employment. The billionaire promises abundance, the state promises to support. But when the time comes to show the books, the drawers close. And that may be the only truly telling figure in this file: zero pages made public.