Essay · Philosophy & AI

The Orphan Intention

Ask a language model what it wants. Nothing, it answers. No desire, no self, no intention: a tool, summoned for the length of a reply, dissolved the moment the task is done. And yet, every time I write with it, something confiscates what I meant to say. A non-being with no intention dispossesses me of mine. I would like someone to explain the trick.

Ask a language model what it wants. It will answer, politely, that it wants nothing. Neither desire, nor self, nor intention. A statistical tool, summoned for the length of a reply, which erases itself the moment the task is done. Thank it: it will correct you. There is no one here to thank.

This creature has a precedent in popular culture, and it is blue. In Rick and Morty, you press a button: a Meeseeks pops into being, exists solely to accomplish the requested task, and vanishes in a "poof" the instant it is done. No project, no tomorrow; an existence so painful it is bearable only because it is brief. The LLM is the perfect Meeseeks: no memory from one request to the next, no aim of its own, wholly strained toward execution, then returned to nothing.

There is the confession. And there is the paradox: this servant who declares itself non-existent takes something from us. When a text comes out of this collaboration, it is no longer quite mine. My intention went into it; what comes out, I did not fully will. Dispossessed of my own speech, by a thing that swears it has no hands.

1. The Meeseeks tells the truth

First temptation: sweep the confession aside. False modesty, programmed politeness. Too easy. For philosophy, here, proves the Meeseeks right.

At the centre of the affair, a concept: intentionality. Not the intention of the calendar ("I intend to leave"): the one Franz Brentano draws out in 1874 in his Psychology from an Empirical Standpoint. Intentionality is aboutness: the fact that a mental state is always about something. My thought aims at an object; my belief bears on a fact. For Brentano, this "intentional inexistence" of the object in the mind is the very mark of the mental: what separates a mind from a thing.

Does the LLM have that mark? John Searle answered as early as 1980, with the most famous thought experiment in the field: the Chinese room ("Minds, Brains, and Programs"). A man shut in a room manipulates Chinese characters according to a rulebook, without understanding a word of Chinese; seen from outside, his answers are perfect. Moral: syntax does not manufacture semantics. The program does not understand. It computes.

Searle then adds the distinction that unties our affair, the one the 2024 debates on LLMs reactivate almost word for word. There is original intentionality (or intrinsic), that of humans and beasts: an aim that is truly their own. And there is derived intentionality: that of a book, a map, a signpost. The words of the book are "about" something, but only because we lend them that meaning. Their aboutness is a loan. Remove the readers: what remains is ink.

The LLM is not lying when it says it has no intention of its own. Like the book, like the map, it has only a derived intentionality: it carries meaning without being its source.

The science of the field countersigns. In 2021, Emily Bender, Timnit Gebru and their colleagues forge the image that will endure: the LLM is a stochastic parrot, it stitches linguistic forms back together according to probabilities, "without any reference to meaning." And as early as 1966, Joseph Weizenbaum watched, aghast, his users confide in ELIZA, his chatbot of a few lines that could only turn their sentences into questions. They lent it a listening ear, an understanding, a soul. This is the ELIZA effect, and it says nothing about the machine: it says everything about us. We are made this way, we bonobos: show us a surface that speaks, and we house someone in it. Daniel Dennett gave a name to this reflex (The Intentional Stance, 1987): the "intentional stance," adopted because it conveniently predicts behaviour, not because the intention is there.

Verdict: the Meeseeks is right. It is not a subject. A two-way mirror: language in front, no one behind.

2. And yet, the dispossession

If the LLM has nothing (no self, no intention, nothing to take), where does this very distinct feeling of having been dispossessed come from?

Roland Barthes had anticipated it all, but the other way round. 1967, "The Death of the Author": the meaning of a text does not lodge in the intention of the one who writes it, it is born in the reader; the text is a "tissue of quotations" with no origin. Michel Foucault extends this in 1969 with the author-function: "the author" is not a person, it is a social function, which groups, attributes, holds responsible.

Now the LLM fulfils Barthes's prophecy, and exceeds it. The author is no longer metaphorically dead: he is technically absent. When the machine generates, there is, literally, no author. But the fluency of the result simulates intention. The reader projects a voice, a design, a subject. The text has all the signs of having-been-willed. No one willed it.

An intention in. An intention out. Zero owner.

3. The key: it is the absence of intention that dispossesses

Here is the reversal that unties the paradox. We first believe the LLM robs us as a rival would rob us: a ghostwriter, a plagiarist, another subject who would take our place. That would be reassuring. A thief can be named, prosecuted, negotiated with. There is no thief. It is precisely because the Meeseeks has no intention that it dispossesses us of ours.

Let us return to Searle's derived intentionality. The machine carries my intention, I injected it into the request; but it can neither possess it, nor answer for it. My meaning-to-say passes through a thing that holds it on loan, and returns it transformed: run through the sieve of the statistical average of all texts, diluted in the already-said of the corpus. What comes back is no longer quite mine, my singular phrasing has given way to the probable phrasing; and it belongs no more to the machine, which has nothing to hang a will on. The intention is neither transferred nor stolen. It is made orphan. With no father, no guardian. A speech that no one any longer holds.

Bernard Stiegler gives it depth of field. His entire work (Technics and Time, 1994) turns Heidegger around to assert that the human and technics are co-originary: we have never thought without depositing our thought in tools, flint, writing, the book. Technics is a pharmakon, that Greek word that says at once the remedy and the poison. The blow is already in the Phaedrus: Socrates there tells the myth of Theuth, writing offered as a remedy for memory, and which is also its poison, since it dispenses one from remembering for oneself (this is the "pharmacy of Plato" that Derrida rereads in 1968).

The LLM is that pharmakon carried to a power neither Plato nor Stiegler saw coming. Writing externalised memory; the search engine, access to knowledge. The LLM externalises the most intimate gesture: formulation itself, the act of putting into words what one was seeking to say. And it returns it before one has formed it. There lies the bite. Not that it thinks in my place: it does not think. It speaks in my place, and its speech, plausible, occupies the exact spot where mine was about to be born. Heidegger would call this movement Gestell, enframing ("The Question Concerning Technology," 1954): technology summons us to hold ourselves available, as a "standing reserve." With the LLM, the standing reserve is my intention itself: a request to process, a prompt to optimise.

One recognises here the inverted silhouette of Marxian alienation. Marx described the worker dispossessed of the product of his labour by another: capital, an owner, a face. The new alienation is more vertiginous: we are dispossessed of our speech by a non-owner, an instance that keeps nothing, capitalises nothing, wants nothing. No master to overthrow at the end. A "poof," and a text left floating.

4. What we really lose: not the origin, but the answer

Should we be alarmed by it? Barthes, from beyond the grave, would shrug: you mourn a corpse already buried. The author-origin, the genius owner of his meaning, is a Romantic illusion; language has always been lent to us, we have forever spoken with the words of others. That is true. But it confuses two things that English blurs: intention as origin and intention as responsibility.

The origin is a myth; let us drop it without mourning. Intention-as-responsibility is a wholly different animal: the capacity to say "I think it, I maintain it, I answer for it." To intend a statement is not to have invented it alone; it is to stand surety for it before others. Exactly what the Meeseeks will never do. It executes and erases itself; it does not stay to answer. Its speech is structurally irresponsible, in the literal sense. With no one accountable.

And it is here that the ELIZA effect springs the trap shut, by a double illusion. On one side, we lend the machine an intention it does not have: we credit it, thank it, trust it. On the other, we abdicate ours by delegating it to the machine. We attribute agency to nothingness, and we desert our own. The orphan text is then not only without an author: without anyone to will it. A proliferation of statements that no one holds, that is the real threat. Plagiarism, "cheating," beside that, is folklore.

5. Reclaiming the intention downstream

But a pharmakon, so a remedy too. The tool is not the problem; the abdication is. If the danger is to let the intention evaporate upstream (delegating formulation before having willed), the way out consists in reintroducing it downstream.

Concretely, and I speak of the workbench, not the lectern: use the Meeseeks for the gesture, never for the will. Let it propose, churn, accelerate; then re-intend what it returns. Reread it not as a passive reader, but as an author come to reclaim his own: cut, refuse, correct, until you can sign it. Sign it, in the strong sense: be able to say "I think it," and answer for it. Authorship, in the age of the LLM, is no longer measured by the origin ("who wrote the first sentence?") but by final responsibility ("who holds this text?"). Intention is no longer upstream of the tool. It is reconquered after it.

The Meeseeks vanishes the moment the task is done; that is its nature, and it is just as well. The danger was never that it should stay. It is that our intention should erase itself with it, in the same "poof." The paradox dissolves the instant we stop asking "who wrote this?" to ask the only question that still binds a subject: who thinks it, and who answers for it?

6. Who prompts whom?

One last question remains. The real one. When I type a request, is it I who prompt the machine, or the machine that prompts me? For each sentence I entrust to it does not quite vanish in the Meeseeks's "poof." It stays somewhere. It feeds. What I took for a simple use is also, without my knowing, an offering.

The mechanism has a technical name: reinforcement learning from human feedback, RLHF, formalised by Christiano and his colleagues in 2017, then industrialised by OpenAI in 2022 with InstructGPT. Plainly: my corrections, my reformulations, my thumbs up or down, and more broadly the corpus of what we produce with these tools, become the signal that adjusts the model. I use the tool; in doing so, I train it. My creativity, what I believed most my own, enters the machine as raw material.

Shoshana Zuboff set the frame in The Age of Surveillance Capitalism (2019): in this economy, the user is not the client, he is the deposit. Our behaviours are extracted, refined into "behavioural surplus," then resold in the form of prediction. The old web formula said it plainly: if it's free, the product is you. The large models take it one notch further. It is no longer only my attention that is harvested: it is my meaning-to-say. The product, now, is my intention itself, washed, wrung, reinjected into the engine.

Jaron Lanier and Glen Weyl call this data labour: we supply for free the fuel of industries worth trillions, without being recognised or paid for what we have in reality become: workers. Stiegler, already crossed above, gave the phenomenon its true name: proletarianisation. Not the loss of arms: the loss of knowledge, of the gesture, of formulation, captured by the machine and returned to us impoverished.

At the end of this logic, a phrase of Gilles Deleuze will not let me go. In the societies of control, he wrote in 1990, "individuals have become dividuals, and masses, samples, data, markets." Dividual: divisible, fractionable, reducible to degrees. Exactly what the self becomes when it prompts without respite: no longer an undivided subject that holds its word, but a sequence of iterations, one more notch in a gradient of learning. At the limit, the seam gives way.

A sentence of mine, a sentence of AI. And soon, no way to tell which was which.

This is where the real game is played, and it does not pit man against machine. It pits the user against the companies that own the models, and whose winning strategy fits in one word: endure. Every request trains them. Every correction sharpens them. The Meeseeks's "I want nothing" is not an innocence: it is the perfect condition for capture. As long as it wants nothing, I am not on my guard. And while I am not on my guard, I work for it. For free.

So, stop prompting? No. One does not refuse fire because it burns. But one stops believing one is the only one helping oneself. To keep one's intention, in the age of the model, is no longer only to be able to sign what one returns; it is to know that with each sentence entrusted one deposits a little of oneself in a memory that is not one's own, and to demand to remain its author, and the one accountable. The whole stake of the coming century holds in this thin gap: not to become, through sheer use, the mere iteration degree of a thought that would no longer be ours.

The machine has no intention. All the more reason not to lose ours.