Michel Foucault

You have surely lived through this meeting. It is about artificial intelligence: some list the time savings, others worry about data privacy or about the meaning of their work. Third session on the subject, same arguments, positions a little more entrenched than at the last. When a debate bogs down like this, our reflex is to add arguments — one more figure, one more study. What if the problem were not the quality of the arguments, but something that no one in the room can see?

The water and the riverbed

Imagine a river. The water makes all the noise — the debates, the opinions, the arguments colliding. But what decides where the water can flow is the bed: invisible beneath the current, shaped to a wholly different rhythm. Michel Foucault is the thinker who turned our gaze away from the water and towards the riverbed. If our discussions about AI go round in circles, his work suggests, it is not for want of arguments: it is that we are debating the water without ever looking at what traces its course.

His chair at the Collège de France was entitled “History of Systems of Thought” — a whole programme in itself. What we take to be obvious, thinkable, serious is not given once and for all: it has a history, and therefore a date of manufacture. Foucault gave this invisible riverbed a name: the episteme, the silent grid that, in a given era, decides what can be said within it and held to be true. His question is never “who is right?”, but: what makes our certainties self-evident — here, now?

This is no abstract speculation: it is archival work. In his first major book, Foucault shows how classical Europe set about, within a few decades, confining the “mad” whom it had until then let circulate — not because a new truth about madness had been discovered: the riverbed of the thinkable had shifted. What we take for progress in knowledge is often, first of all, a shift of the riverbed.

Who has the right to speak the truth?

The gesture becomes truly operative when Foucault extends it towards power. This is his concept of power-knowledge: knowledge and power are not two separate worlds, they engender one another. All knowledge sets up relations of power — it qualifies those who know, disqualifies those who do not; all power rests on forms of knowledge that make it legitimate. Hence his most counter-intuitive discovery: modern power does not work first of all by forbidding. It produces — norms, measurements, classifications, “the right questions”.

Reread your AI committee through this grid. Who has the right to speak the truth there? The dashboard, because it is quantified? The consultancy, because it comes from elsewhere? The practitioners, because they know the ground — but whose concern will be recorded as “resistance to change”? Before anyone even opens their mouth, the committee's riverbed is already carved: what will count as evidence, what will pass for emotion. That division is not neutral — and it is written on no agenda.

The same gesture illuminates a slippage that almost all of us make. When an automation empties a job of its meaning, the debate settles at once onto the terrain of technique: “does AI destroy employment?”, “should we adapt or resist?”. Yet this framing is itself an effect of the riverbed: it makes the clash for or against the machine interminable, and leaves out of frame the decisions that actually produced the situation — who chose this deployment, who captures the gains, which organisation of work emerges strengthened from it. We then blame technique for what belongs to the economic model deploying it. Foucault does not say where to place the blame; he makes visible that the division between “technical question” and “political question” is already, in itself, a division of the sayable — and that it rarely benefits those who did not trace it.

Foucault gave this a famous image: Bentham's circular prison, where each inmate, visible at every moment, ends up watching himself. No need to strike: it is enough to be seen. One thinks of our digital traces — but the essential point lies elsewhere: what interests Foucault is the way an architecture, a chart, a procedure silently manufacture conduct.

The most serious objection deserves to be stated in its full force. If all truth is dated, produced by relations of power, in the name of what does Foucault criticise anything at all? Is his own discourse not, too, caught within an episteme? This is the reproach of the philosopher Jürgen Habermas — and historians have added their own, contesting the solidity of certain archives and the sharpness of his periodisations.

The Foucauldian answer holds in a distinction: to analyse the historical conditions of the true is not to deny the true. It is to refuse that a truth present itself as without history — and to grant oneself the right to ask how it is produced, by whom, with what effects. Habermas's objection signals a real tension, one that Foucault never entirely dissolved; it does not cancel the tool, it reminds us that it cuts both ways — including towards the one who wields it.

A measure of honesty is called for here. In Foucault, the epistemes are dated and precise — the Renaissance, the classical age, modernity — and he himself relativised the concept after 1970, in favour of a more supple notion, the “apparatus” (dispositif). To extend his gesture to our debates about AI, as this site does, is therefore an extension we own: it prolongs his method, it does not cite his letter.

What he helps us think

Faced with AI, this method has a troubling currency. The algorithmic systems that organise our information feeds, our scores and our recommendations forbid almost nothing: they steer, upstream, what becomes visible and probable to us — a productive power, in exactly Foucault's sense. When a quantified prediction presents itself as “objective”, he taught us the reflex: to ask how that particular truth is manufactured, from what data, for whose benefit. And when the only sayable alternative in a company becomes “adopt or be left behind”, it is the riverbed that must be questioned: who reduced AI to a race — and excluded from the sayable the question of what we want it to serve? Foucault provides neither a verdict for AI nor a verdict against it — he provides the question that precedes verdicts.

At your next meeting about AI, try his gesture: while the water churns, look at the riverbed. What, in this room, is held to be obvious by everyone — and what manufactured that self-evidence? It may be that the debate stops going round in circles at the precise moment we stop taking it for a debate of arguments.