Power Without a Voice
A reply to a comment, and what a French philosopher's concept of the ventriloquist state reveals about the directive that shut down Fable 5
A comment on a recent article asked a question I could not fully answer in comment-thread format. The reader, present at the original Fable 5 discussion, pushed the argument one level further: if the G7 cannot govern the source of frontier AI, does the burden shift entirely to firms to build their own internal governance, or is this a failure no amount of firm-level architecture can fix? The reply I gave closed on a single line: what is actually being tested, underneath the AI question, is whether the separation of powers the Western model depends on can survive a technology whose strategic value moves faster than any legislature can deliberate. That sentence deserved more room than a comment box allows.
Stanford's Institute for Human-Centered AI, a leading academic centre on AI policy, published an analysis in February that addresses a basic problem: everyone in AI policy uses the word sovereignty, but rarely means the same thing by it. Their answer is a taxonomy built on two questions. Who is making the sovereignty claim, a state or a firm? And which part of the technology does that claim actually cover? On the first question, states pursue sovereignty in two distinct registers. Some pursue what the paper calls "harder" self-sufficiency, building a fully domestic AI stack so that no foreign government can interrupt access. Others pursue a softer strategic autonomy, accepting some dependency on foreign technology while keeping enough leverage (alternative suppliers, domestic fallback capacity) to avoid being dictated to. On the second question, the paper separates the technology itself into distinct layers, from computing hardware and training data through to the models, the applications built on top, and the expertise required to build any of it. A country can hold genuine sovereignty over one layer while remaining entirely dependent on others; France has invested seriously in domestic compute while still relying heavily on foreign-trained foundation models. The result is a map worth using: it shows precisely who controls what, rather than treating AI sovereignty as one undifferentiated ambition.
Australia offers a live test of the softer version. In December 2025, Canberra joined a nine-nation framework, what one analysis has termed Pax Silica, alongside the United States, the United Kingdom, Japan, South Korea, Singapore, the Netherlands, Israel, and the UAE (Sweden and India joined in the following months), formalising access to allied AI infrastructure as conditional on political alignment rather than treating it as an open market good. It was built to function as a parallel supply chain partners could depend on rather than a market they competed in. Five months later, Fable 5 went dark for every signatory at once, Australia included, with no distinction made between treaty partners and anyone else.
What it does not tell you is how a given decision escapes democratic scrutiny in the first place. Sovereignty taxonomy maps actors and layers. It does not explain the mechanism by which a directive like the one that shut down Fable 5 arrives without ever needing to justify itself in terms a legislature, a court, or a citizen could contest. For that, a different vocabulary is required, and I found it, unexpectedly, in a French philosopher whose questions matter to me more than his conclusions do.
Eric Sadin, a French philosopher who has spent the past decade examining how digital technologies reshape political and social life, published a tribune in Le Figaro in February, reprinted in Il Corriere della Sera's La Lettura supplement in March, on what he calls "la technologie ventriloque." His argument concerns a specific and verifiable phenomenon: candidates increasingly use generative AI to structure, and at times draft outright, their campaign speeches. Sadin's claim is that this initiates what he names a "désubstantialisation de l'essence même du politique," a hollowing-out of the substance of the political itself. When the words a candidate speaks are no longer fully authored by anyone accountable for them, the speech retains its political form (a podium, a microphone, an audience) while losing its political content. The candidate becomes a vessel through which something else speaks. The same mechanism is already at work well beyond politics: a brand voice, built by a marketing team but increasingly drafted, tested, and optimised by AI, asks for a customer's trust in exactly the same borrowed register, a human signature attached to words that were never quite anyone's own. Sadin names the political version of this precisely: not metaphorical evasion, but the literal displacement of authorship from the person nominally responsible to the system generating the words.
His argument, as written, concerns campaign speech specifically, and the brand-voice parallel above is my own extension, not his. What follows is a further extension, also mine. If generative AI can hollow out political substance by displacing authorship from the candidate to the system, the same mechanism operates, in a different register, when administrative power displaces its own authorship from political choice to technical necessity. The Lutnick directive that shut down Fable 5 on the 12th of June was not generated by an AI system. But it performed the identical structural move Sadin describes: a decision with unmistakably political content (whose AI capability gets restricted, on what national security logic, with what consequence for a hundred and thirty countries) was spoken in the grammar of administrative compliance rather than political choice. No criteria were disclosed against which the decision could be tested. No legislature voted on it. The directive contained no voice that could be held accountable the way political speech is held accountable, because nothing about its form ever presented itself as political at all. It arrived as a technical instrument, deemed export doctrine extended to cloud infrastructure, and technical instruments do not get debated; they get complied with.
The G7 communiqué that followed at Évian performs a softer version of the same evasion. "Voluntary." "Non-binding." "Guiding principles." These phrases were not chosen because political actors declined to act. They were chosen because they let no one in particular be said to have decided anything. Macron's plea for democratic cooperation and Trump's refusal of binding language are both, in their own ways, political positions, openly held and openly contested, which is precisely why they are not the problem. The problem is the document that emerges between them, engineered in its very phrasing to avoid sounding like a decision anyone made. That is ventriloquism at the multilateral scale: many voices, contributing to a text, none of them ultimately answerable for what it says.
Sadin's broader work is deliberately totalising; he has described AI's expansion as an "antihumanisme radical," and I do not entirely share that conclusion. What I take from him instead is his discipline, the willingness to ask what happens to political substance once its authorship can no longer be traced. Most governance writing, mine included, stops at process and rarely pushes the question down to the level of voice; he does, and that is worth borrowing even where his conclusions are not.
This is the answer to my reader's question, stated plainly. The burden does not shift to firms. A company can build excellent internal governance and still hold no power over what is actually missing, which is not architecture but accountability, a requirement that whoever makes a decision says so in terms that can be examined. The sovereign compute and domestic models I argued for in the Fable 5 piece remain necessary, and nothing here changes that. But building them would not, on its own, have stopped the Lutnick directive from being written in the language it was written in, and it would not stop the next G7 communiqué from being drafted to avoid naming who decided what.
What is also required is something closer to what the EU AI Act already attempts for deployed systems: named criteria, disclosed reasoning, a right of challenge before harm occurs, extended to the kind of administrative and security action that switched off Fable 5. That extension does not yet exist in any serious form, in the EU framework or elsewhere; national security exceptions remain the one place where this vocabulary still gives way to discretion.
So: can it survive? Only on the condition that someone builds what is currently missing; nothing about the technology guarantees it on its own. What Fable 5 and Évian both demonstrate is that the separation of powers does not fail because legislatures are too slow, though they often are. It fails when decisions are built, deliberately or not, to never reach the legislature at all, dressed in language that excuses them from being heard as political. Ten days before the summit, the same administration issued a further executive order directing federal agencies to harden national security systems against AI-related threats and coordinate vulnerability response across the AI industry, again entirely through agency directive rather than legislation. By Évian, that grammar was already becoming routine rather than exceptional. Survival depends on closing exactly that gap: extending disclosure and appeal to the security and administrative actions currently exempted from them, building the vocabulary this moment requires before the next directive arrives in the grammar of necessity rather than the grammar of choice. That gap, more than any other, is where I have chosen to spend my attention and where I intend to keep it.
I did not expect a comment thread to lead here, but that is what a serious exchange is supposed to do. The letter that shut down Fable 5 was addressed to Anthropic, signed on a Friday afternoon, and offered no criteria a recipient could contest. That hour has an older use, one civilisations have relied on long before any of this: people gathered around a glass of wine or a beer, working out together, openly and without hurry, what kind of world they actually want to live in. Nothing about artificial intelligence exempts us from that obligation, the one place our alterity still shows itself as the intelligence no machine has yet matched.