Who Speaks? Foucault and the Problem of Beginning
Michel Foucault, The Discourse on Language
1. The Desire Not to Begin
At the beginning of The Order of Discourse, Foucault imagines a different way of entering speech. He would have preferred not to be the one who begins:
“I should have preferred to become aware that a nameless voice was already speaking long before me.”
He imagines discovering that discourse was already underway, so that he would not have to produce its first sentence. He would only have needed to “join in, to continue the sentence it had started.”
The fantasy is easy to overlook. Foucault is not merely expressing modesty about his own authority. He is disturbing a familiar picture of how discourse begins.
We normally imagine a speaker first and discourse second:
speaker → discourse
Someone has something to say and puts it into words. Foucault reverses the direction:
discourse → speaker
The speaker may arrive late. What he says is already shaped by a field of words, concepts, distinctions, conventions, and possibilities that preceded him. His task is not to create discourse from nothing but to enter something already in motion.
The reversal can be stated simply: We tend to think that discourse begins with a speaker. Foucault asks us to consider the possibility that the opposite is true: we become speakers by entering a discourse that has already begun.
The point is not that individuals are irrelevant or that nobody ever speaks deliberately. It is that the speaker may not be the beginning we imagine him to be.
That leaves a more difficult question: who counts as a speaker?
2. Speaking Is Not Enough
The question is more difficult than it first appears. We usually treat speaking as sufficient evidence of a speaker: if a person produces words, that person is speaking. But Foucault's concern is not simply with the physical act of utterance. He is interested in the conditions under which an utterance can enter discourse as something that can be heard, recognized, and taken seriously.
A person can utter words without occupying the position of a speaker.
Discourse does not give every utterance the same status. What is said can be accepted, ignored, authorized, or excluded depending on who says it, where it is said, and the position from which it is spoken. A statement made in one setting may carry authority; the same words elsewhere may be dismissed as irrelevant or unintelligible.
The issue, then, is not simply: Who is speaking? but: Who counts as a speaker?
In The Order of Discourse, Foucault describes procedures that regulate the circulation and authority of speech. Certain people are recognized as qualified to speak about particular subjects. Certain settings confer legitimacy. Other forms of speech are excluded from serious discourse.
This is where Foucault's madman enters.
3. Nothing or Too Much
Foucault begins with a curious account of the madman's speech. For centuries in Europe, he says, the words of a madman could be treated in two apparently opposite ways. They could be ignored completely, falling “into a void.” Or they could be received as words of truth: “a naive or cunning reason, perhaps even a rationality more rational than that of a rational man”.
At first, these responses seem to have little in common.
Nothing.
There is no meaningful speech here. The words can be discarded as noise.
Too much.
There is extraordinary truth here. The words deserve attention precisely because they come from someone outside ordinary reason.
Yet Foucault draws them together: “whether excluded or secretly invested with reason, the madman’s speech did not strictly exist.”
The point is not that the madman literally said nothing. He spoke. The problem was the position from which his words could be heard.
In the first case, his words disappear. They are dismissed before they can become an object of interpretation.
In the second, they become something exceptional: prophecy, revelation, hidden wisdom, divine warning. We may listen for the truth that speaks through him rather than listen to him as a speaker among other speakers.
The two responses therefore converge in an unexpected way. One reduces the speech to nothing; the other makes it too extraordinary to function as ordinary speech: Nothing or too much.
In both cases, the madman is denied the position of an ordinary speaker. And that raises a more unsettling question: What if the speaker is not the source of what he says?
4. Who Speaks Through Him?
Plato’s Ion offers a useful parallel. Socrates complicates the idea that the rhapsode is the autonomous source of the words he delivers. The Muse inspires the poet; the inspiration then passes through the poet to the rhapsode, and from the rhapsode to the audience:
Muse → poet → rhapsode → audience
Socrates compares this chain to iron rings suspended from a magnet. The power does not originate in the individual ring. Each ring receives a power that allows it to draw another ring into the chain. The rhapsode is an intermediate link, while the spectator is the last.
The point here is not Plato’s theory of poetic inspiration. It is the position of the person who speaks. The rhapsode is clearly speaking, yet what he says does not begin entirely with him. He is a link in a transmission.
That possibility sharpens the question raised by Foucault’s madman. Speech can come through someone without making that person its ultimate source.
The question can now be put more broadly: Who speaks when someone speaks?
This becomes less abstract when a machine begins completing our sentences.
5. The Machine That Continues the Sentence
“I should only have needed to join in, to continue the sentence it had started.”
That sentence from Foucault sounds strangely familiar today.
We can now type a few words and ask a machine to continue the sentence. The response may arrive immediately, grammatically complete and stylistically coherent. It can sound as though someone has taken over where we stopped.
But who is speaking?
The user wrote the prompt. The model produced the continuation. Yet the continuation is also shaped by the data on which the model was trained and by the technical and institutional system in which it operates. The words on the screen therefore do not lead easily back to a single, identifiable speaker.
And yet the text has a voice. The text arrives with a voice even when its speaker is difficult to identify.
This is what makes the LLM relevant to Foucault's question. The issue is not simply that a machine produces language. It is that fluent discourse can appear while the position of its speaker remains uncertain.
We can attribute the response to the user, the model, the training data, or the system that governs the interaction. Each attribution captures part of what happened, but none settles the question completely.
That uncertainty is not something we have to eliminate before the phenomenon can be understood. It is part of the phenomenon.
The LLM makes Foucault’s question harder to ignore.
The machine can continue the sentence but what it cannot settle is whose sentence it is.
6. The Speaker Who Arrives Late
And this brings us back to Foucault's opening image. He imagines that he would not have to be “the one from whom discourse proceeded.” He would discover that a nameless voice was already speaking and simply join in.
At first, this can sound like a modest wish: a desire to avoid claiming too much originality. But the fantasy goes further. It challenges the idea of the speaker as the point from which discourse begins.
We tend to think that discourse begins with a speaker. Foucault asks us to consider the possibility that the opposite is true: We become speakers by entering a discourse that has already begun.
The speaker, in this sense, may always arrive late.
We built machines that can continue the sentence before we had settled the question of who speaks when we speak.
Bibliography
Foucault, M. (1971). L’ordre du discours: Leçon inaugurale au Collège de France prononcée le 2 décembre 1970. Gallimard.
Foucault, M. (1972). The archaeology of knowledge and the discourse on language (A. M. Sheridan Smith, Trans.). Harper & Row.
Plato. (1997). Ion (P. Woodruff & A. Nehamas, Trans.). In J. M. Cooper (Ed.), Plato: Complete works (pp. 937–952). Hackett Publishing Company.
OpenAI. (2025). GPT-5 system card. OpenAI.
Comments
Post a Comment