What can the State requisition?

The instrument of last resort

An artificial intelligence capability becomes indispensable to an essential function: intelligence, cyberdefence, the operation of a critical infrastructure, the continuity of a vital service. It belongs to a private operator. A crisis arises. The State must be able to dispose of it immediately.

For exactly this situation, the law possesses an instrument of last resort. It is rarely mobilised, but it is what makes all the others credible: the State negotiates, frames, conditions, because in the last instance it can requisition. Requisition is not one instrument among others; it is the silent horizon of all the others.

French law has just reaffirmed it in the most universal terms a text can employ. Article L. 2212-1 of the Defence Code, rewritten by the military programming law of 1 August 2023 and made applicable by Decree No 2024-895 of 1 October 2024, allows a decree in the Council of Ministers to order, in the event of a present or foreseeable threat to the activities essential to the life of the Nation, "the requisition of any person, natural or legal, and of all the goods and services necessary to counter it". Any person. All goods. All services. The legislator renovated this regime precisely because the former one was judged obsolete and difficult to mobilise in a crisis. The lock was changed in 2023.

This power is nothing like an opportunistic confiscation. It is reserved for exceptional circumstances, subject to proportionality and to the absence of any other adequate means available in time, and the arrangement organises even the temporary placement of assets or activities under control. It is an instrument of necessity, built to dispose, for a time, of what is lacking · not to appropriate.

The same horizon exists elsewhere. The ninth paragraph of the 1946 Preamble provides that any good, any undertaking whose exploitation has acquired the character of a national public service or a de facto monopoly must become the property of the community. The American Defense Production Act of 1950 makes it possible to order a company to serve defence needs as a priority. Three different legal traditions, one same certainty: what matters to the life of the Nation can, in the end, be taken.

This tribune does not contest the extent of that power. It is real, and the 2023 text carried it as far as a text can go. It questions something else: what the taking, once exercised, actually delivers to the one who takes.

For requisition is not, here, an object of study: it is a yardstick. The strongest power of disposal the law knows, confronted with the newest object it encounters · what that power delivers, or fails to deliver, measures what every lesser title can promise.

The material presuppositions of the taking

No requisition text defines what a seizable object is, and none had to. The effectiveness of the taking rested on presuppositions that matter itself guaranteed in place of the law. They must be reconstituted, for it is their solidity, not the text's, that has always made the value of the instrument.

They are four. The object of the taking was locatable: the requisition designated a place, a territory, an address where the public power could present itself. It was detachable from its holder: the taking dispossessed one in order to serve the other, and the compensation of direct and certain material losses, which the regime organises, presupposes precisely a measurable dispossession. It was operable under another direction: the requisitioned factory produced for the State, the requisitioned ship sailed under a new command. It persisted, finally, through the taking: what was seized today was still worth something tomorrow.

A semiconductor foundry verifies these four presuppositions. A stock of masks verifies them, and the health crisis showed the instrument in operation. A power plant, a rail network, a mineral deposit verify them. The requisition of persons itself, which the renovated regime founds on profession and skills, rests on a variant of the same base: the art required is individual and is performed on order · that of the doctor, the pilot, the mechanic.

These four presuppositions are not legal conditions. They are the implicit physics on which the instrument prospered. The law of requisition was written for things that stay in place.

The composed capability

The capability that now organises decision-making, production and soon the conduct of conflicts defeats these four presuppositions. Not because it escapes the text: it is, on the contrary, entirely within it. A frontier artificial intelligence model decomposes into elements each of which, taken in isolation, presumably falls within the categories the text enumerates. The numerical weights presumably fall within intangible goods. The servers are goods. Hosting, maintenance, the supply of compute are services. The engineers are persons, requisitionable by reason of their skills. The requisition decree can name each of these elements.

That is precisely where the question changes in nature. For the capability is not to be confused with any of the elements the decree can name. It results from their composition: weights, a compute infrastructure, data chains, a continuous flow of adjustments, and an organisation whose knowledge is written nowhere. Let us submit that composition to the four presuppositions.

It is not locatable as a whole: its elements are distributed among operators and jurisdictions the decree does not all reach, and a national prerogative does not execute itself beyond the territory where it commands. It is not detachable: copying weights dispossesses no one, and a taking that leaves the power intact in the hands of its holder operates no transfer · it produces a duplicate. It is not operable by mere substitution of direction: seizing the servers delivers a state of the system, not the system, for part of what makes it work resides in the organisation that maintains, adjusts and understands it. It does not persist of itself: a model separated from that flow of upkeep degrades, and the State would find itself holding an earlier state of the power, not the power.

It will be objected that the State can requisition the teams, since the regime allows it and sanctions refusal. The objection deserves better than a quick answer, for it touches the core. A requisitioned person cannot make the obligation disappear; refusal to comply exposes to sanction. But the law can impose a presence and the performance of a service. It cannot, by the requisition order alone, guarantee the continuous performance of a collective whose mastery rests on largely tacit coordination. Refusal to serve can be sanctioned; the quality of an ensemble's work cannot be decreed. Add geography: part of the teams, providers and dependencies stands outside the territory where the decree can command.

It will be objected, finally, that the arrangement organises, in certain cases, the temporary placement of assets or activities under control · in appearance, the taking of the organisation itself. But what is transferred then is a direction, not an adhesion: command changes hands, and the tacit coordination in which mastery resides does not follow the organisation chart.

The thesis then holds in one sentence, and it is more troubling than that of a bounded power: the law can requisition each of the components without their union transferring the capability. The text reasons in categories; the capability lives between the categories. The legal taking bears on elements; operational mastery resided in their composition. Between the two, the bridge remains to be built.

The upstream compensations

This dissociation is named nowhere, but everything proceeds as if States had understood it. The instruments proliferating around artificial intelligence capabilities share one trait: they all operate upstream of the taking. Export controls, which regulate the circulation of what cannot be retained. Tiered access, which negotiates the use of what one does not produce. Draft treaties, declaratory red lines. Four different arrangements, one same structure: organising access to a power whose taking, downstream, would no longer guarantee mastery.

The American debate on nationalising the frontier laboratories is no exception: it poses the political question, should one take, while assuming that the taking would deliver the thing. Yet nationalising transfers titles; titles carry the capital, not the capability. An earlier position of this chamber, "What does a state own when it owns AI?", established it for public equity in the laboratories: owning is not disposing. Requisition promised precisely the reverse · disposing without owning. It is that promise the composition puts in default.

The contemporary accumulation of artificial intelligence regulations must then be reread for what it perhaps is: not the deployment of a sovereignty asserting itself, but the compensation system of a prerogative whose effect has dissociated from its exercise. Upstream instruments multiply because the downstream instrument, intact in its letter, no longer guarantees its result.

The missing qualification

The law knows neighbouring figures, and none covers the object. The obligations on operators of vital importance impose duties of continuity on the operator; they transfer nothing to the State. The golden share confers a right of veto over disposals; it does not make a capability function. The defence contract buys a service; it presupposes a consenting counterparty, the exact opposite of a requisition. Each of these instruments seizes a fragment. None seizes the composition.

Artificial intelligence, moreover, is only the extreme case of a wider category. The difficulty appears whenever an essential capability resides in no asset taken in isolation, but in the continuously operated composition of assets, services and skills.

One must then distinguish what normal functioning confuses: having access to a capability, contracting for that capability, being able to dispose of it in the last resort. Three different situations, which the ordinary life of organisations superimposes, and which crisis separates. The contract organises the first two. Nothing, as the law stands, guarantees the third. Access is not disposal, any more than ownership was disposal.

And the lesson exceeds the State. If the most powerful instrument of the law does not, by itself, guarantee mastery of a composed capability, no lesser title can, by itself, establish it: neither the contract, nor the shareholding, nor the continuity clause. The legal title of access does not suffice to qualify effective mastery of the capability.

What the law lacks is therefore not an additional power. Powers abound, and Article L. 2212-1 is written as broadly as a text can be. What is lacking is a qualification: the composed capability as an object of public prerogative. As long as it does not exist, a requisition exercised over an artificial intelligence capability would order, in law, everything it can order, and risk obtaining, in fact, something other than what mattered: servers, copies and presences, while the capability came undone in the hands of the one who had seized them, because it was in none of the elements seized. It was in their composition, and none of the instruments examined qualifies the composition as an object of public prerogative.

The question, then, is not whether the State should give itself the power to take artificial intelligence capabilities: that power is written. It is what the taking would deliver to it. As long as it is not examined, States will hold instruments universal in their letter and uncertain in their effect, and will negotiate, as the requesting party, access to what their own laws claim to be able to seize.

A legal order can live for a long time with instruments whose effect has dissociated from their exercise. It discovers it only at the moment it mobilises them.

Paris, October 2026

Sources · French Defence Code, Articles L. 2212-1 et seq. (as amended by Law No 2023-703 of 1 August 2023, Military Programming Law 2024-2030) · Decree No 2024-895 of 1 October 2024 · Preamble to the Constitution of 27 October 1946, paragraph 9 · Defense Production Act of 1950 (US) · DELEX Consortium, "What does a state own when it owns AI?", 2026.