Deleuze, desire and the machine: asking a better question about AI
The argument about artificial intelligence runs almost entirely on questions from the philosophy of mind. Deleuze and Guattari offer a different unit of analysis — desire as production, the machine as a cut into a flow, the assemblage as the thing that actually does something. What that buys, and what it cannot buy.
Women operating a Bell System switchboard for overseas calls, 22 December 1943; Records of the Women's Bureau, US National Archives. A switchboard is the plainest picture of what Deleuze and Guattari meant by a machine: an operator, a cord and a jack, with every connection a cut made into a flow · Wikimedia Commons
Almost every public argument about artificial intelligence is conducted in the vocabulary of a single philosophical sub-discipline. Does the system understand what it says, or is it shuffling symbols it cannot read? Is there something it is like to be it? Those questions come from the philosophy of mind, and they share a shape: each is asked of one object and answered by describing that object's interior.
There is another way of proceeding, from a corner of French philosophy that has never had much to say to analytic philosophy of mind and has mostly been ignored in return. Yesterday's piece on Deleuze covered the metaphysics of difference, the collaboration with Félix Guattari, and the objections that have stuck. This one goes into one seam of that work: desire, the machine, and whether the apparatus gives a better purchase on machine intelligence than the standard one does.
Two warnings. The prose is genuinely difficult, sometimes because the ideas are hard and sometimes because the authors were performing, and no summary that makes it sound lucid is honest. And Deleuze died in November 1995, so much of what follows about artificial intelligence is extension rather than exegesis. I will say each time which it is.
Desire produces; it does not shop for what it lacks#
The dominant tradition defines desire by absence. In Plato's Symposium, Socrates gets Agathon to concede that one desires only what one does not have, and Diotima's myth then makes Eros the child of Poros, resource, and Penia, poverty. Hegel makes self-consciousness a desire that works by negating its object; Schopenhauer is bleaker still, since satisfaction only clears the ground for the next want. Lacan gives the position its sharpest modern statement: desire opens in the gap between need and demand, and its object is structurally the object that is missing.
L'Anti-Œdipe, published by Minuit in 1972 and translated in 1977 by Robert Hurley, Mark Seem and Helen Lane, reverses the polarity. The formulation on page 26 of the English edition is flat enough to quote whole:
Richard Rorty asked philosophy to stop pretending it could underwrite science, morals and politics, and to settle instead for making cruelty harder to ignore. Thirty years on, with the strongman passage from Achieving Our Country quoted in every election, the question is whether his liberal ironist can still hold the line.
Desire does not lack anything; it does not lack its object. It is, rather, the subject that is missing in desire, or desire that lacks a fixed subject; there is no fixed subject unless there is repression.
Desire, on this account, is an activity that connects things and makes something out of the connection. An eye and light, a hand and a tool: neither is intelligible as a shortfall being made good.
The political payoff arrives two pages later. "Lack is created, planned, and organized in and through social production," they write, and in the same passage: "The deliberate creation of lack as a function of market economy is the art of a dominant class." Lack is real, then, but manufactured rather than given, and somebody manufactures it.
It is not obviously the better theory. Lack handles some things production handles badly — grief, which really is organised around an absence, and addiction, where the repetition is self-defeating in a way that "desire makes connections" does not illuminate. Lacanians have pressed that point for fifty years without a full answer. What production explains better is why desire attaches to wages, machinery, nations and interfaces, none of which has anything to do with anybody's family.
The machines are meant literally, which is where most readers stop#
Anti-Oedipus opens with a paragraph that has deterred more readers than almost any other in modern philosophy. "It is at work everywhere, functioning smoothly at times, at other times in fits and starts. It breathes, it heats, it eats. It shits and fucks. What a mistake to have ever said the id." A few lines on comes the sentence that matters: "Everywhere it is machines — real ones, not figurative ones: machines driving other machines, machines being driven by other machines, with all the necessary couplings and connections."
Not figurative. The technical definition arrives on page 36. A machine "may be defined as a system of interruptions or breaks", the French coupures, cuts. Every machine cuts into a flow produced by the machine it is attached to, and is itself a flow for the next machine along. The mouth cuts the flow of milk; the milk comes from a breast, itself a machine cutting another flow. Hence the formula: "every machine is a machine of a machine."
Three things follow. There is no first machine and no privileged unit, so neither the organism nor the person is the natural place to stop. A machine is identified by what it cuts and passes on rather than by what it is made of, which is how an institution and a nervous system can be machines in the same sense. And the analysis has no interior: a desiring-machine can be described completely without asking what it feels like to be one.
The obvious complaint is that a concept applying to everything discriminates nothing. Deleuze and Guattari accept the charge and convert it into a method. "The question posed by desire is not 'What does it mean?' but rather 'How does it work?'", they write on page 109. That is a real standard, and also loose enough to license a great deal of unfalsifiable writing.
The term the later work hangs on is agencement, and English readers have been mildly misled about it since Brian Massumi's 1987 translation of Mille plateaux settled on "assemblage".
John Phillips, then at the National University of Singapore, made the most-cited objection in two pages: "Agencement/Assemblage", published in a glossary issue of Theory, Culture & Society in 2006, volume 23, numbers 2 and 3, pages 108 to 109. His complaint concerns connotation. Agencement is an ordinary French word for arrangement or layout, the fitting-out of a space so that it works; a shop has one, so does a kitchen. It carries fit and purpose, and leans towards a process. "Assemblage" in English arrives with a gallery accent, meaning a collage of found objects, and suggests a heap. The translation, Phillips warns, "can give rise to connotations based on analogical impressions" that cut a term loose from the argument which made it.
What survives the translation problem is the claim this essay hangs on: an assemblage is composed by function, not by kind. The example on page 89 is feudalism.
We would have to consider the interminglings of bodies defining feudalism: the body of the earth and the social body; the body of the overlord, vassal, and serf; the body of the knight and the horse and their new relation to the stirrup; the weapons and tools assuring a symbiosis of bodies — a whole machinic assemblage.
An oath and a piece of leather in the same list, because they are components of one arrangement that does something. Medievalists have argued for seventy years over how much the stirrup explains — the dispute has its own name, the Great Stirrup Controversy — but the assemblage point holds whichever way it goes, being a claim about the unit of analysis rather than about causal weight.
Manuel DeLanda gave the concept a harder edge in A New Philosophy of Society (Continuum, 2006), arguing that assemblage components stand in relations of exteriority: a part keeps its own identity, can be detached, and can be plugged into a different arrangement. He is explicit that this extends Deleuze rather than reports him. It also makes the claim checkable — if the components cannot be unplugged, you were describing an organism — which matters later.
Ask what the arrangement produces, not what is in the box#
Here the reporting stops and the extrapolation begins.
A Deleuzian reading changes the unit. It asks what the system is connected to, what it cuts into, and what the whole arrangement produces. Take a recommendation engine. The model is one component; the assemblage is the model, a ranked feed, a person with fifteen spare minutes, an auction market pricing that person's attention in real time, and a supply of content whose producers adapt to whatever the ranking rewards. What it produces is a distribution of attention, a set of incentives on the supply side and a revenue line. Whether the ranker understands anybody is a question about the wrong component, and answering it either way changes none of that.
Consider a sharper case: the same model weights, two arrangements. In one, the model drafts a reply and a human agent reads, edits and signs it. In the other, it closes tickets directly against an average-handle-time target. Nothing has changed inside the box, and what the arrangement produces has changed completely — the error profile, the liability, the skill content of the job, who gets blamed. Labour-process research on AI-mediated work documents exactly this without needing a theory of machine understanding, as does postphenomenology from a Husserlian starting point.
The nearest analytic cousin is the extended mind thesis of Andy Clark and David Chalmers, which also moved a boundary that had looked fixed. But the extended mind is still an argument about where the mind is, and the assemblage framing does not need a mind to be anywhere. Yuk Hui's work on the recursive machine shows how much can be built on that side of the boundary without raising the question of inner experience at all. That is the framing's economy, and it is also its bill.
Declining a question is not settling it, and there is a hole here that no amount of reframing closes.
If there is something it is like to be some system, now or in thirty years, the assemblage account has nowhere to write that fact down and no instrument that would detect it. A method whose unit is the arrangement is structurally blind to a property of one component. That is not an academic worry. Work on AI welfare and machine-consciousness research proceed on the assumption that moral patiency turns on inner states, and the New York Declaration on Animal Consciousness rests on the same assumption for octopuses and crows. If they are right, assemblage analysis can only tell you what to do after somebody else has told you who counts. The honest Deleuzian reply — that the hard problem may never be answerable while the arrangement questions are answerable today — is a division of labour rather than an answer.
A second objection is methodological. Nothing in the concept says where an assemblage stops, so any collection of things whatever can be described as an arrangement that produces something, and much published work using the word does exactly that. DeLanda's exteriority condition is the best available repair, because a claim that components survive detachment can fail; most assemblage-talk does not trouble to make it.
Escape is not automatically good news, and they said so#
The word for an assemblage coming apart is deterritorialisation, and the line along which it comes apart is a ligne de fuite, a line of flight. The accelerationist reading, running from Nick Land in the 1990s through much of what is now written about technology, takes Deleuze and Guattari as cheerleaders for dissolution.
What it has to ignore sits in the ninth plateau of A Thousand Plateaus, "1933: Micropolitics and Segmentarity", where four dangers are set out — fear, clarity, power, and a fourth belonging to lines of flight themselves. The fourth is "the line of flight crossing the wall, getting out of the black holes, but instead of connecting with other lines … turning to destruction, abolition pure and simple, the passion of abolition." Such lines, the same passage says, can "emanate a strange despair, like an odor of death and immolation, a state of war from which one returns broken." The plateau is dated 1933 because the case under discussion is fascism, treated not as an excess of order but as an escape that turned suicidal.
The two-handed structure is easy to see in technology. Recorded music, taxi dispatch and classified advertising were all decoded at speed, with real gains for some people, then recoded onto a rent taken at a single chokepoint. Deterritorialisation did its work each time; whether it was emancipatory is a separate question with a separate answer.
What they do not supply is a criterion for telling a productive line of flight from a destructive one before the fact. Their practical advice amounts to going slowly and keeping enough of the old organisation to survive the change, which is temperament rather than method. It is still more than the accelerationists credit them with.
Becoming is not imitation, and identity is a slowing-down#
The tenth plateau, "1730: Becoming-Intense, Becoming-Animal, Becoming-Imperceptible…", is the material most often reduced to slogan, and the slogan — identity is fluid — is duller than the claim.
Deleuze and Guattari are explicit about what becoming is not:
Becoming is certainly not imitating, or identifying with something; neither is it regressing-progressing; neither is it corresponding, establishing corresponding relations; neither is it producing, producing a filiation or producing through filiation.
Becoming-animal is therefore not dressing up, not totemism, not metaphor. It means entering a relation in which your capacities change: what you can affect and what can affect you. Their test case is Freud's Little Hans, the five-year-old with the horse phobia, whose becoming-horse they call "a truly political option" rather than a symbol standing in for his father. And it is never solitary: "A becoming-animal always involves a pack, a band, a population, a peopling, in short, a multiplicity."
Underneath sits an austere account of what a body is: relations of speed and slowness among its parts, plus a capacity to affect and be affected. On that definition a stable identity is a region where the rates have settled, a slowing-down rather than a foundation. Becoming-imperceptible, the limit case, means ceasing to present a recognisable profile at all — which reads differently now that presenting a profile is what most systems require.
It is tempting to say that a language model becomes: that it shifts persona mid-conversation, that fine-tuning is a becoming, that a model in a new deployment has joined a new pack. The temptation should be resisted, and this is my judgement rather than a reading of the text. Becoming in the tenth plateau requires a real alteration in what a body can affect and be affected by. A model prompted into a different register has not changed its capacities; the arrangement around it has, which the previous section covers without this vocabulary.
It picks the right unit. Most of what anyone actually cares about in this technology — who is harmed, what gets cheaper, which jobs change shape, what a feed does to a public — is a property of the arrangement and invisible at the level of the model. It turns the central question into one with answers, since what a system produces can be measured while whether it understands has had half a century without adjudication. And it blocks two symmetrical errors: that a model is merely a tool, which ignores the arrangement, and that a model is a mind, which ignores everything but the box.
What it does not buy is a theory of moral status, and that gap is not a detail. It does not explain mechanism either, since knowing what an arrangement produces tells you nothing about why a particular model produces what it does — the business of interpretability research, which has no Deleuzian content whatever. And it is silent on what ought to be done. What a thing can do does not entail what it should do, and the ethics Deleuze took from Spinoza is badly placed to supply the missing step.
None of this settles the question most people want settled. It changes which questions have answers. An arrangement can be described, its components named, its outputs counted; the interior of a model, on current evidence, cannot. That is a smaller claim than is usually made on Deleuze's behalf, and the one that survives contact with the books.
Confucius is a name most readers know and few can place. This profile introduces the philosopher behind it: his ideas of ren, ritual, the exemplary person and the rectification of names; his claim that good government begins in character; and the compiled, argued-over text we call the Analects.
Analytic philosophy tells its origin story as an escape from F. H. Bradley, the Merton recluse who died on 18 September 1924. But the regress argument about relations in Appearance and Reality was never answered, only bypassed, and the ethics of Ethical Studies anticipated the communitarian turn by a century.
.jpg?width=500 "Kant")
%2520Family%2520Mansion%252C%2520Qufu%2520(13044361055).jpg%3Fwidth%3D1600&w=3840&q=75)
Comments
0Loading comments…