Attention, Gesture, Machine, Intensity — Yves Citton

Yves Citton · 2026 · 1:31:12

Summary

Citton asks one question: in what we humans do, which part is the machine and which part is the gesture? He does not ask how humans differ from machines.

• A machine is something programmed or programmable.
• A gesture is “a movement of an animated body, irreducible to what programmed it”. It is not fully controlled, it adapts, it produces the unexpected, and it starts something without knowing the end.

Then he applies this split to four areas:

1. Attention. Attention is being automated (self-driving cars, DeepL). The model of attention we inherited is a soldier fixed on a radar screen. Attensity names the ways attention can move away from that model.
2. Research and curiosity. There are four stages: forest, book, Google, prompt. At each stage we hand more to the machine. With prompts, what we hand over may be curiosity itself.
3. Apparatus vs machine. An apparatus is a tool whose settings you can reach and change. A “machine” is a regime of use that sets those choices for you.
4. Generative AI. In the Q&A he admits that machine learning “throws out the window” his own definition. A generative model also exceeds its programming.

He ends with two verbs. Exit the machines (regimes of use, habits, automation). Experiment with apparatuses.

Stance to note: he calls the political economy of AI “an absolute calamity” and “criminal”. He calls the technology “a miracle”. He wants these two kept apart.

Map — talking points

#	Point
1	Expectation and attention are linked. A discipline is organised expectation. It finds things, and it also hides other things from view
2	Thesis 1: human attention can now be automated (self-driving cars, DeepL)
3	Flusser’s Gestures: a gesture is an animated movement that its programming cannot account for
4	Four adjectives for a gesture: uncontrolled, adaptive, meets and creates the unexpected, a beginning
5	Huyghe: emancipation starts with an uncontrolled gesture. Agamben: “means without end”
6	Objection to his own thesis: virtuosity. A craftsman or musician has a gesture that looks over-controlled
7	Reframe: find the machine part and the gesture part inside human action
8	Architecture: Gehry’s sketch is gesture, the build is machine, and Gehry became “a machine for making Gehry”. The architect programs other people and must also improvise (Ingold, the grain of the wood)
9	Attensity! (Friends of Attention). Attention science, from 1870s German labs to the Cold War radar operator
10	Five features of the radar model: fixed focus, alone, plugged into a machine, task assigned from above, competitive. School attention has the same features
11	Four stages of research: hunter-gatherer, books, Google (2005), prompts
12	Curiosity is a gesture: you leave where you are. Prompts may hand it over to the machine. Chatonsky: be an experimenter with AI, not a user
13	Huyghe: apparatus (settings you can reach) vs machine (regime of use). Phone cameras. “Photograph your pockets in front of the Taj Mahal”
14	Close: exit the machines, experiment with apparatuses
15	Q&A: touching from a distance (Rousseau), the momentum of gestures, the Arab Spring
16	Q&A: machinic address vs singular address. Steyerl’s “Mean Images”. Stiegler
17	Q&A: immediate sensation, Abram and Morizot
18	Q&A: handwriting vs typing. Photography vs portraitists as a model for AI vs writers
19	Q&A: expert systems vs machine learning. Generative AI is irreducible to how it was programmed
20	Q&A: is the machine alive? His test: “where does the electricity come from?” Keep the technology apart from the economy
21	Q&A: an audience poem on machine agency. Machine rights. Words like “intelligence” are nominalist illusions that we still need

Putting it together

The mental model in one paragraph. Every action we take has a programmed part (what society, habit, machines and institutions have set in advance) and a gestural part (what adapts, surprises, and begins something whose end we don’t know). Technology always moves some gestures out of the body and into devices (Leroi-Gourhan, Stiegler). That move is a pharmakon: poison and cure together. The danger now is that two very deep gestures are moving: attention, which has been modelled since the radar operator as a function that can be automated, and curiosity, the decision to leave where you are, which prompts can make unnecessary. The response is not refusal. It has two parts. Exit the regimes of use (defaults, platforms, habits), which needs technical understanding (Simondon). And treat tools as apparatuses whose settings you explore and play against (Huyghe, Flusser). Citton’s own twist: generative AI is not a classic “machine” by his own definition, since it exceeds its programming. So the question is open: where is the gesture in our use of AI, and possibly in AI itself?

Three tensions to keep open

1. Fixed focus: discipline or virtue? Attensity vs deep work.
2. Generative AI: average or emergence? Steyerl vs Chatonsky, and Citton cites both.
3. Settings: freedom or burden? Huyghe’s apparatus vs the democratic gift of good defaults.

Reading path

• Start: Flusser, Gestures (the chapters on searching and photographing). Huyghe’s online lectures, if you read French. Attensity! (2026).
• Then: Simondon, On the Mode of Existence of Technical Objects, Part I. Dreyfus, What Computers Can’t Do (introduction). Crary, Suspensions of Perception (introduction). Ingold, Making, chapters 1–2.
• Deeper: Stiegler, Technics and Time vol. 1. Agamben, “Notes on Gesture” and “What Is an Apparatus?“. Deleuze, “Postscript on the Societies of Control” (5 pages). Ginzburg, “Clues”. Citton, The Ecology of Attention.

Part 1 — Before you read

Key terms

• Gesture (geste). In this French line of thought (Flusser, Guérin, Agamben, Huyghe), a gesture is not “body language”. It is movement that carries intention or affect but goes beyond what is programmed or planned.
• Machine / mega-machine. The talk builds on earlier sessions that day. Mega-machine is Lewis Mumford’s term (1967): a whole society organised so that people act as machine parts. His first example is the labour that built the pyramids.
• Program. Flusser’s word for the set of possibilities built into a device. A camera’s program is every photo it can take. A user who only explores those possibilities is, for Flusser, a functionary.
• Apparatus. Watch out: three different meanings.

  Source	Original word	What “apparatus” means there
  Flusser (1983)	Apparat	A black box with a program that programs its user (the camera makes the photographer)
  Foucault (1977) / Agamben (2006)	dispositif / dispositivo	Any arrangement that captures and governs behaviour: laws, buildings, phones
  Huyghe (Citton’s source)	appareil	A device whose settings the user can reach. The opposite of capture
  In English translations all three become “apparatus”. Citton uses Huyghe’s sense. Most English-language theory you read will use Flusser’s or Agamben’s sense. Check which one an author means before you compare them.

• Regime of use. The habits and defaults that a technology and its market impose, beyond what the hardware can actually do. Selfies are one.
• Attensity. The Friends of Attention’s word for attention outside the radar-operator model: roaming, shared, not tied to a machine, chosen by the person, not competitive.
• Expert systems vs machine learning. Expert systems (1970s–80s) followed hand-written rules: someone has to foresee every case. Machine learning sets its own internal weights from data. This difference carries the most weight in Q&A point 19.
• Proletarianisation (Stiegler). Losing a skill because it has moved into a machine. Workers lost savoir-faire (how to make); consumers lost savoir-vivre (how to live). Stiegler warned that thinkers would lose theoretical knowledge. Citton’s “we hand over curiosity” belongs here.
• Hylomorphism. The old idea, from Aristotle, that making means imposing a form (morphe) on passive matter (hyle). Simondon and Ingold reject it. For them, making is a negotiation with active matter.
• Tacit knowledge. Knowledge that cannot be fully written down (Polanyi: “we can know more than we can tell”). This explains why expert systems failed and why gestures resist programming.

People named (a reading list)

Person	Why they come up	Start with
Vilém Flusser	Gesture as non-programmed movement	Gestures (1991; Eng. 2014)
Pierre-Damien Huyghe	Gesture as the root of emancipation. Apparatus vs machine	L’art au temps des appareils (2005); his lecture series Faut-il s’émanciper ?
Giorgio Agamben	Gesture as “means without end"	"Notes on Gesture” in Means Without End (1996)
Tim Ingold	Making as following the material	Making (2013)
Friends of Attention (D. Graham Burnett et al.)	Attensity, “human fracking”	Attensity! (Crown, Jan 2026)
Hito Steyerl	AI images as statistical and cruel	”Mean Images” (said to be New Left Review, 2023; venue not confirmed)
Bernard Stiegler	Technology as pharmakon. Proletarianisation of know-how	Taking Care of Youth and the Generations (2008)
Grégory Chatonsky	Artist who experiments with generative AI	His works and texts
David Abram	Literacy cut us off from the more-than-human world	The Spell of the Sensuous (1996)
Baptiste Morizot	Tracking and attention to other living beings	Ways of Being Alive
Pierre Cassou-Noguès	Philosopher of machines; speaks at the same event	His work on machine imaginaries

Citton’s own frame

Citton is best known for The Ecology of Attention (2014; Eng. Polity 2017). It moves the debate from the “attention economy” (attention as a scarce resource to sell) to an “attention ecology” (how environments shape shared attention). He also co-edited Politics of Curiosity (Routledge, 2024), which explains why curiosity is his way out in this talk (Culturgest bio). His 2012 book Gestes d’humanité (the host’s starting point) is his gesture theory. This talk updates it for AI.

The current debate (2025–26)

• Attention activism has a manifesto. Attensity! came out in January 2026. Its frame is that platforms “frack” attention, and its answer is collective, not individual willpower: study groups, coalitions, “sanctuaries” (New Books Network, UnHerd, Public Seminar review). The reviews found are mostly favourable. No strong critique turned up.
• Cognitive offloading to LLMs. The best-known evidence is MIT Media Lab’s Your Brain on ChatGPT (Kosmyna et al., June 2025). EEG showed the weakest brain connectivity in the LLM group, more in the search-engine group, and the most in the no-tool group. Caveats: 54 participants, only 18 in the last session, and a preprint (MIT Media Lab). The study gives Citton’s “curiosity is being handed over” claim some evidence, but thin evidence.
• Machine agency and moral status. The audience poem connects to a live debate in 2025–26: AI labs have model-welfare programmes, and philosophers ask whether agency without life deserves recognition. Citton’s line: agency, yes, worth thinking about; a soul, “probably not”.
• AI’s environmental cost. Citton calls it “criminal”. This is now a mainstream concern (data-centre power and water use), and it supports his electricity test.

The ideas

Citton spoke for 45 minutes to an audience that already knew these traditions. Here they are laid out. Seven families of ideas meet in this talk:

 PHILOSOPHY OF GESTURE ─────┐                ┌──── PHILOSOPHY OF TECHNICS
 Mauss · Leroi-Gourhan      │                │     Simondon · Stiegler · Mumford
 Flusser · Agamben · Guérin │                │     Benjamin · Flusser (apparatus)
 Huyghe                     ▼                ▼
                     ┌──────────────────────────────┐
 CRAFT & SKILL ────► │  CITTON: what in our action  │ ◄──── ATTENTION STUDIES
 Polanyi · Dreyfus   │  is machine, what is gesture?│       James · Crary · Mackworth
 Sennett · Ingold    └──────────────────────────────┘       Simon · Friends of Attention
                       ▲            ▲            ▲
   CURIOSITY & RESEARCH│   HISTORY OF AI         │ ADDRESS, PROFILE, CONTROL
   Augustine·Blumenberg│   GOFAI→Dreyfus→        │ Rousseau · Benjamin (aura)
   Ginzburg · Warburg  │   deep learning·Sutton  │ Deleuze · Rouvroy · Steyerl
   Bush · Flusser      │   Bender · emergence    │

Where each school of thought is covered:

• Philosophy of gesture — from body technique to emancipation: idea 1
• Philosophy of technics — machines, black boxes, pharmaka: idea 4
• Craft and skill — why virtuosity is not automation: idea 1
• Attention studies — how attention became “concentration”: idea 2
• Curiosity and research — from vice to virtue to delegation: idea 3
• A short history of AI — why “programmed” stopped meaning “machine”: idea 5
• Address, profile and control — why “mean” means cruel: idea 6
• Two shorter threads: idea 7

1. Gesture vs programming — the central idea

In short. Citton separates a gesture from the programming behind it. A gesture is uncontrolled, adaptive, meets and creates the unexpected, and is a beginning. He tests this against the virtuoso carpenter and pianist, then changes the question from whether an action is human or machine to which part of an action is machine and which is gesture.

🎙 In the talk. “A gesture is a movement of an animated body … irreducible to what programmed this movement”. A gesture has four features:

1. Uncontrolled. “programmed and controlled go quite well together”, so a gesture always contains something uncontrolled.
2. Adaptive. A fully programmed system “won’t be able to adapt to the terrain”, because inputs fix outputs.
3. Meets and creates the unexpected. “We never really know for certain what they are going to do” beyond a few seconds.
4. A beginning. After Guérin and Agamben, a gesture is “the beginning of something … the end of which we cannot know”. Huyghe adds that emancipation begins there: “a shout … a fist … an architect’s gesture”, something that “passes through me” and that I do not control either.

He then tests the thesis against virtuosity (the Japanese carpenter, the pianist). He changes the question from “human vs machine” to which part of an action is machine and which is gesture?. And he applies that to architecture: Gehry’s sketch is gesture, the building process is machine, and an aging star becomes “a machine for making Frank Gehry”. The architect is “a programmer” of other people’s gestures who must also improvise with material, workers and site.

📚 Deeper theory

• Why “programmed” is the right opposite of gesture. In information theory a program maps inputs to outputs. Flusser saw that devices and institutions have programs. Mauss saw that bodies have programs too (techniques of the body). So the line Citton draws does not separate humans from machines. It runs through both. This explains the “Flusserian question”.
• The four features match the history of cybernetics. Norbert Wiener’s cybernetics (1948) and W. Ross Ashby’s law of requisite variety (1956) say that a regulator must have at least as much variety as the disturbances it faces. A fully programmed machine has a fixed variety, and the world has more. That is Citton’s “the unexpected would cause the machine to crash”, in engineering terms. Gestures are how living systems keep enough variety.
• Simondon gives machines a gesture-like ideal. For Simondon the most advanced machine keeps a margin of indeterminacy (idea 4). So even within machine theory, the best machines are partly “gestural”. This anticipates Citton’s reversal on AI (idea 5).
• Emancipation in the gesture. Huyghe’s claim links to Hannah Arendt’s idea of natality in The Human Condition (1958): action is the human capacity to begin something new and unpredictable, and politics lives on it. Arendt is not cited, but the gesture is a beginning is close to her.
• Agamben’s “means without end”, unpacked. Aristotle separated poiesis (making, which ends in a product) from praxis (action, its own end). Agamben adds a third: the gesture bears and displays a means without serving an end. Dance shows the body moving; mime shows a gesture without its object. Citton’s tennis mime without a racket is one example, though in the talk he sets expressive gesture aside.

Where it is weak.

• Virtuosity. Citton raises it and repairs it himself. The deeper repair comes from Dreyfus and Merleau-Ponty: expert skill is past rules, not more rules (Roots).
• Habit. Bourdieu’s habitus (1972) says that most of our gestures are socially learned dispositions: how we sit, speak and taste. If so, the “uncontrolled part of gesture may often be social programming that we don’t notice, not freedom. Citton touches this with Gehry’s style”, but does not face it fully.
• Generative AI meets three of the four features. He admits this in the Q&A (idea 5).

📚 Roots: Philosophy of gesture — from body technique to emancipation

The problem. Gesture sits between nature and culture. A gesture is learned (no one is born knowing how to shake hands), yet it also expresses something not fully chosen. In the 20th century, philosophers in France and beyond made gesture a serious topic for this reason.

The chain of thinkers

1. Marcel Mauss, “Techniques of the Body” (1934). He noticed that French and American soldiers walked differently, and that nurses in New York walked like American film actresses. His conclusion: even walking, swimming and sleeping are learned techniques passed on by a culture. Key point for Citton: gesture is already partly “programmed” by society. This is the opening for Citton’s “machine part” inside human action.
2. André Leroi-Gourhan, Gesture and Speech (1964–65). A palaeoanthropologist. He argued that tools, gesture and language evolved together, and that human evolution is the exteriorisation of capacities into tools: first the hand into the stone tool, then memory into writing, then calculation into machines. Key point: each technology takes a gesture out of the body. Stiegler builds on this directly.
3. Vilém Flusser, Gestures (1991). About sixteen short phenomenologies: writing, painting, photographing, filming, planting, shaving, smoking a pipe, telephoning, and searching (research). He defines gesture as a movement through which freedom expresses itself, a movement that no causal explanation fully covers. In the “gesture of searching” essay he argues that the research gesture became the model for all gestures in modern times, and that its claim to be detached and objective is now in crisis. Citton’s four stages of research apply this chapter to AI.
4. Giorgio Agamben, “Notes on Gesture” (1992). His famous first line: “By the end of the nineteenth century the Western bourgeoisie had definitively lost its gestures.” He links this to Gilles de la Tourette’s 1886 studies of walking, Muybridge’s and Marey’s motion photography, and early cinema. These were the same years that measured and recorded gesture, and in his account they mark a loss of natural gesture. His definition: gesture is neither making (a means to an end, like building) nor acting (an end in itself, like ethics). It is a “means without end”: a pure means, shown as a means. Dance is his example. Citton uses this as a beginning whose end we cannot know.
5. Michel Guérin, Philosophie du geste (1995). He sorts gestures into types: the gesture of making, of giving, of writing, of the idea. For him the gesture is the beginning before any work or rule, which is the point Huyghe takes.
6. Pierre-Damien Huyghe. Philosopher of art and design (Paris 1). His books include L’art au temps des appareils (2005) and Petit lexique pour le design (2024). He gives two ideas to the talk: emancipation starts in an uncontrolled gesture, and the appareil vs machine distinction.
7. Yves Citton, Gestes d’humanité (2012). His “wild anthropology” of aesthetic gestures, with the chapter “The Machinization of Gestures” that brought him to this event.

The thread that ties them together. Mauss shows that gesture is social programming. Leroi-Gourhan shows that technology takes gestures out of the body. Flusser and Agamben look for what in gesture exceeds program and purpose. Huyghe and Citton make that excess political: the seed of emancipation.

📚 Roots: Craft and skill — why virtuosity is not automation

The problem. Citton’s hardest objection to himself is the master craftsman. If gesture is the uncontrolled part, why does mastery look so controlled? Philosophers of skill have the answer.

1. Maurice Merleau-Ponty, Phenomenology of Perception (1945). Skill lives in the body as motor intentionality. The organist does not think about where the stops are; the body “knows”. Habit is not mechanical. It is the body’s open, flexible grip on the world.
2. Michael Polanyi, The Tacit Dimension (1966). “We can know more than we can tell.” A cyclist cannot write down the physics of balance, yet rides. Economists now call this Polanyi’s paradox (David Autor, 2014): the tasks that are hardest to automate are the ones whose rules nobody can state.
3. Hubert Dreyfus, What Computers Can’t Do (1972), and Mind over Machine (1986, with Stuart Dreyfus). This is the key link between craft and AI:
   • The five-stage skill model: novice → advanced beginner → competent → proficient → expert. The novice follows explicit rules. The expert no longer uses rules. They see the situation and respond.
   • The conclusion: expert systems, which turn expertise into rules, will always stop at the “competent” stage, because real expertise is not rule-following. History largely proved him right about rule-based AI, which hit its limits in the late 1980s.
   • For Citton: virtuosity is not more program. It is past program. The virtuoso’s gesture “escapes me” in Dreyfus’s sense: it is no longer run by rules.
4. Richard Sennett, The Craftsman (2008). Craft is a conversation between the hand and the head, built over long practice. It includes “material consciousness”: attention to what the material wants. Repetition is not mindless. Each repeat is a small variation.
5. Tim Ingold, Making (2013) and Lines (2007). Building on Simondon, Deleuze and Guattari, he replaces hylomorphism with correspondence: the maker follows the flows of the material (the grain of the wood) and does not impose a plan. He contrasts wayfaring (finding your way as you go, like a hunter) with transport (point A to B by a fixed route). The Japanese carpenter film and “embrace the grain of the wood” are Ingold’s ideas.
6. Mihaly Csikszentmihalyi, Flow (1990). Psychology’s version: at the peak of skill, self-consciousness disappears and action “happens by itself”. It feels uncontrolled and is highly precise. This is Citton’s “superhuman mastery” that “escapes me”.

The resolution these thinkers give Citton. There are two kinds of “automatic”. Mechanical automatism follows a program and breaks on the unexpected. Embodied skill has absorbed so much practice that it responds to the unexpected without rules. The virtuoso has the second kind. So virtuosity supports his thesis, provided “machine” means rule-following.

What he leaves out. Expressive and communicative gesture: mime, sign language, gesture as language (McNeill’s Hand and Mind, 1992). He sets it aside on purpose.

Check yourself

• Why does Citton say the question is not “human vs machine”? Which thinker gives him that move?
• How does Dreyfus’s five-stage model resolve the virtuosity objection?
• Where does Bourdieu’s habitus put pressure on the definition?

---

2. Attention automated and the radar operator — attensity

In short. Citton says we are in a phase where our attention can be automated, as in self-driving cars and DeepL. He traces the modern idea of attention to a 1950s soldier watching a radar screen: fixed, alone, plugged into a machine, set by a hierarchy and competitive. His “attensity” is the opposite on each point. School attention shares the radar features, which he calls a deliberate provocation.

🎙 In the talk. Thesis 1: “we are in a historical phase where our attention can be automated”. The evidence is self-driving cars (the driver’s attention moves into “sensors, calculations, computing”) and DeepL (a translator’s refined attention is automated, except in literary translation, where “there is something quite specific”). Attensity! traces attention science from 1870s German labs to a 1950s US soldier scanning a radar screen for Soviet missiles. That soldier still shapes “what we have in mind whenever we talk about attention”, including when teachers complain about students on their phones.

The five features, with their opposites (attensity):

Radar model	Attensity	Example in the talk
Stable, fixed focus	”Uncentered distraction”: look, drift, come back	”looks at a fly flying around in the lab”
Alone	With others	student “not supposed to chat with his little classmate”
Plugged into a machine	Things happen outside it	”a cog in the mega-machine”
Task assigned by a hierarchy	Chosen by the person	”you’re not going to start reading Stendhal”
Competitive (geopolitics, grades)	Not competitive	grades as the medium of school

His sharpest line: school attention has the same features, so it is “the worst kind of attention … as problematic as the Pentagon”. This is a deliberate provocation. He also says: “it’s not a question of disqualifying this gentleman … it’s just one model among others”.

📚 Deeper theory

• The radar operator is the actual origin of attention science, not a metaphor. Mackworth’s 1948 clock test was built for RAF radar crews, and its finding (vigilance falls after about 30 minutes) shaped all later research. Broadbent’s filter model and Cherry’s cocktail-party work came from military communications. When psychology moved to “attention = a filter on a limited channel”, attention became engineerable. Engineerable things can be automated. This is the bridge between the two halves of Citton’s argument: the soldier’s attention was already designed as a machine function, so it is no surprise that a machine can take it over.
• Crary’s twin thesis. The 19th-century institutions that demanded concentration (factory, school, the new spectacle) also produced distraction. They are one system. The Friends of Attention flip the polarity: what discipline calls “distraction” may hold the free part of attention.
• Discipline, then control. The radar operator and the graded student are Foucault’s disciplinary subjects (Discipline and Punish, 1975): watched, isolated, ranked, given tasks. The feed is Deleuze’s control (idea 6): no walls, continuous modulation. Attensity answers discipline. Something else is needed for control. The Friends’ answer is collective “sanctuaries”.
• What automation of attention actually means. In a self-driving car, the car has no “attention” in the human sense. It has sensors and prediction models. But the function the driver’s attention served (detecting, prioritising, reacting) is now done by the system. Citton’s claim is functional, not psychological. This matters: you can agree that the function is automated and still hold that human attention is something more.
• The brain science is mixed (Roots). The default mode network and creative incubation favour wandering. Killingsworth and Gilbert link mind-wandering to unhappiness. Attensity is a political claim about who directs attention, more than a psychological claim that wandering is good.

📚 Roots: Attention studies — how attention became “concentration”

The problem. Attensity! claims that our idea of “good attention” has a particular history: labs, factories, armies. Is that true? The historians say mostly yes.

1. Wilhelm Wundt’s lab, Leipzig, 1879. The first experimental psychology laboratory. It measured reaction times and the “span” of attention. This is the “1870s Germany” in the talk.
2. William James, Principles of Psychology (1890). The classic definition: “Everyone knows what attention is. It is the taking possession by the mind, in clear and vivid form, of one out of what seem several simultaneously possible objects or trains of thought. Focalization, concentration, of consciousness are of its essence.” This definition, with its stress on focus, is the one the Friends of Attention attack.
3. Jonathan Crary, Suspensions of Perception (1999). The key historical book, though not named in the talk. Crary shows that from about 1870 to 1900, attention became a central problem of modern institutions: factories, schools and the new media all needed people to concentrate, and they also produced distraction. Attention and distraction are twins of one system, not opposites. His 24/7 (2013) extends this to the always-on economy.
4. WWII and the Cold War — the radar operator becomes science.
   • Norman Mackworth’s “clock test” (1948) was built for the RAF to study why radar operators missed signals. It found the vigilance decrement: detection drops sharply after about 30 minutes. The radar operator in the talk is the founding subject of modern attention science, not just an image.
   • Colin Cherry (1953, the “cocktail party” problem) and Donald Broadbent (1958, filter theory) came out of military communication work. They modelled attention as a filter on a channel, the language of information theory. Attention became something you could engineer, and therefore automate.
   • The US SAGE air-defence system (1950s) put operators at screens watching for Soviet bombers. It was also a birthplace of interactive computing.
5. The attention economy.
   • Herbert Simon (1971): “a wealth of information creates a poverty of attention.”
   • Michael Goldhaber (1997) named the “attention economy”.
   • Tim Wu, The Attention Merchants (2016); James Williams, Stand Out of Our Light (2018); Shoshana Zuboff, The Age of Surveillance Capitalism (2019): platforms harvest attention and behavioural data.
   • Citton (2014) moved the debate from economy to ecology, and the Friends of Attention (2026) moved it to activism: “human fracking”.
6. What brain science says about “uncentered distraction”. The evidence is mixed, which helps you judge attensity:
   • For: the brain’s default mode network (Raichle, 2001) is active during mind-wandering and linked to memory, planning, and creative incubation. One study (Baird et al., 2012) found that an undemanding task with mind-wandering improved creative problem-solving.
   • Against: Killingsworth and Gilbert (2010), “A wandering mind is an unhappy mind”: people’s minds wander about 47% of the time, and they report being less happy while it happens.
   • So: wandering chosen and shared (attensity) differs from wandering driven by a feed (distraction). The Friends of Attention’s political point is about who controls the wandering, not about wandering itself.

Where others push back.

• Deep work (Cal Newport, 2016) and the “attention as scarce resource” school: sustained focus is now rare and valuable, and it needs protection, not deconstruction. Reading a hard book requires something like the radar operator’s discipline.
• Contemplative traditions (meditation, prayer) also train fixed attention, yet are not military or competitive. So fixed focus is not the problem in itself. The problem is fixed focus assigned by others, for others’ ends.
• The school comparison is unfair in one respect: a classroom can also be collective, chosen and non-competitive. Citton is attacking a model of school, not every teacher.

What he leaves out. Who chooses the object of attention and why. Simone Weil (“attention is the rarest and purest form of generosity”, 1942) and Iris Murdoch (attention as moral vision) treat attention as ethics. That would strengthen attensity’s claim that attention is more than performance.

Check yourself

• Name the five features of the radar model and their attensity opposites.
• Why does the history of vigilance research make attention automatable?
• Is fixed focus itself the problem? Give one counter-example.

---

3. The research gesture and curiosity — from forest to prompt

In short. Citton describes research in four stages: hunter-gatherer, books, Google and prompts. His claim is that what is now handed over is curiosity itself, which means getting out of where you are. He admits Google saved him a great deal of time and that something was lost. He wants experimental, non-instrumental uses of AI, not just users.

🎙 In the talk. Research has four stages:

1. Hunter-gatherer. “Looking for berries … tracking animal signs.” Eyes and nostrils. Nothing handed over.
2. Books. “I scribble in the margins, I go get a book from the library.” A narrower world than the forest, yet “I have never tired of everything one can find in books”.
3. Google, 2005. He edited volume 5 of the Mémoires secrets de Bachaumont (an 18th-century gossip periodical) in two months instead of two years at the BnF. Older colleagues were “disgusted … he’s cheating”. He admits the loss: “when you go to the National Library, you meet people”.
4. Prompts. Students hand in texts that are “completely clean, generally well-informed, and they are cheating”.

His claim: what is now handed over is curiosity itself. Curiosity “isn’t just asking yourself a question … it means getting out of where you are”, making “a gesture that comes from us”. “I ask the question, text appears, I copy-paste” may lose this step out. He calls LLMs “investigative agents” as well as writing agents. He is careful: he has “no personal experience” with prompts, and “just as Google didn’t kill research in libraries … things will happen”. With Chatonsky, he wants non-instrumental, experimental uses of AI: “not just users of AI, but experimenters”.

📚 Deeper theory

• This is Flusser’s chapter, updated. In Gestures, Flusser makes the gesture of searching the model of modern gestures and says its claim to detachment is breaking down. Citton keeps the structure and adds the fourth stage.
• What each stage hands over (a sharper version of the talk):

  Stage	What the medium does for you	What you still do	What is lost
  Forest	Nothing	Sense, track, interpret traces	—
  Library	Stores and orders knowledge (catalogue, shelves)	Choose, read, connect, synthesise	The world’s direct traces
  Search engine	Finds (ranks sources)	Evaluate, read, synthesise	The stacks, the “good neighbour”, the people
  LLM	Synthesises (writes the answer)	Ask, check	The reading, the connecting, possibly the wish to go
  The table shows Citton’s real point. Earlier stages handed over finding. The prompt hands over synthesis, the step where understanding is built.

• Ginzburg’s hunter as the gold standard. The hunter does not know in advance what the trace means. Inquiry starts from the anomaly. A prompt starts from a question already formed, so it can return only what fits the question. This is the expectation trap Citton opened with: “as long as we look in that direction, we are not looking in another direction”.
• Serendipity needs friction. Warburg’s “good neighbour” and Walpole’s “accident and sagacity” both need a path with detours. Search engines reduced the detours. LLMs can remove them entirely: the answer comes without the shelf next to it.
• The pharmakon view (Stiegler). Every stage was a poison and a cure. Books weakened memory and made scholarship possible (Socrates, Phaedrus). Google weakened recall and made Citton’s two-month edition possible (Sparrow 2011). The question for LLMs is how to make the cure side win: what practice of care.
• Evidence so far. The MIT EEG study (2025) suggests lower cognitive engagement with LLM-written essays. It is small, and the result fits the Socratic pattern. It does not yet show lasting harm.

📚 Roots: Curiosity and research — from vice to virtue to delegation

The problem. Citton ends on curiosity. Curiosity has a long and mixed history in Western thought.

1. Augustine, Confessions, Book X (c. 400). Curiosity is “lust of the eyes” (concupiscentia oculorum): a vice, the restless wish to know useless things.
2. Hans Blumenberg, The Legitimacy of the Modern Age (1966). Part III is “The Trial of Theoretical Curiosity”: how curiosity went from vice (medieval) to the engine of science (modern). The modern age is partly defined by curiosity being made legitimate. Citton’s Politics of Curiosity (2024) continues this: who gets to be curious, and about what?
3. Carlo Ginzburg, “Clues: Roots of an Evidential Paradigm” (1979). The art connoisseur Morelli (who identified painters from earlobes and fingernails), Freud (slips of the tongue) and Sherlock Holmes (cigar ash) all use one method: reading small traces. Ginzburg traces it back to the hunter reading tracks, the oldest form of knowledge. This is Citton’s first research figure, the hunter-gatherer, and it shows why that figure is the model of open-ended research.
4. Aby Warburg’s library, Hamburg (1900s–1920s). Organised by the “law of the good neighbour”: the book you need is often not the one you came for but the one next to it. The library’s order was built to create useful accidents. This is what Citton loses when he moves from the BnF to Google: the stacks and the people.
5. Vannevar Bush, “As We May Think” (1945). The Memex: a desk that stores all of a researcher’s documents and links them by “associative trails”. This is the dream behind hypertext and the web. In Bush’s design, the researcher blazes the trail, which is the active gesture.
6. Serendipity. A word coined by Horace Walpole (1754) from The Three Princes of Serendip, who “were always making discoveries, by accidents and sagacity, of things they were not in quest of”. Robert Merton later studied it in science. Note: “sagacity” matters as much as “accident”. Serendipity needs a prepared, attentive mind and a wandering path.
7. The Socratic warning, Phaedrus (c. 370 BCE). The god Theuth offers writing to King Thamus as a cure for forgetfulness. Thamus says it will cause forgetfulness and give only the appearance of wisdom. This is the first recorded “the new medium will make us stupid” argument. Each new stage, from book to Google to prompt, repeats it.
8. The Google effect (Sparrow, Liu and Wegner, 2011). People who expect to find information online later remember where it is rather than what it is. Memory moves into the network. The MIT EEG study (2025) is the LLM version.

The pattern. Each stage of research (forest, book, search engine, prompt) moves one more step of the research gesture into a medium: first the trace, then the order of knowledge, then the finding, and now the synthesis. Citton’s claim is that the last step moves curiosity itself: the decision to leave where you are. Earlier media moved where you went. The prompt can decide that you don’t need to go.

Where others push back.

• LLMs as a new “good neighbour”. A well-used model can widen the path: it suggests unknown authors, opposing views, unexpected links. The detour can be prompted. This is Chatonsky’s experimental use.
• The extended-mind view (Andy Clark and David Chalmers, 1998): cognition has always spread into notebooks and tools. Offloading is not loss if the system as a whole thinks better.
• The “cheating” charge has a history. The older colleagues said Citton cheated with Google. Today his Google edition looks normal. The same may happen with prompts. Citton himself expects this.

What he leaves out. How to keep the gesture of curiosity when using LLMs. He names the problem and the principle (experiment, don’t just use) but not a practice. Possible practices that follow from his frame: ask the model for what you didn’t ask; go to the primary source the model names; set your own question after reading, not before.

Check yourself

• What does each stage of research hand over to a machine? Which step does the prompt add?
• Why does Citton define curiosity as a movement and not a question?
• Use the pharmakon idea to state the strongest case for LLMs in research.

---

4. Apparatus vs machine — settings and regimes of use

In short. Following Huyghe, an apparatus lets the user adjust settings, like a film camera. Phone cameras mostly hide the settings. A machine, in the second sense, is a regime of use that imposes or eases certain ways of acting, as the selfie does. His advice is to exit the machines and experiment with apparatuses.

🎙 In the talk. From Huyghe: an apparatus “allows the user to make adjustments, to enter into settings”. The classic case is the film camera, where you set aperture, focal length and exposure: “at least three settings”. Phone cameras are mostly no longer apparatuses: settings are automatic, and the phone is “designed so we can’t access settings, open the black box, or tinker with what’s inside, for various commercial, control, and other reasons”. A machine, in the second sense, is a regime of use. Machines “either impose or facilitate certain regimes of use”. Following Flusser, the selfie is a regime of use: “It’s not Apple forcing us … but when you go to the Taj Mahal … your camera is going to frame the Taj Mahal.” His exercise: take selfies of your pockets everywhere. His close: exit the machines (regimes of use, automation, habit, which “means having a bit of technical knowledge”), and experiment with apparatuses.

📚 Deeper theory

• The three “apparatuses” (see Key terms). Huyghe’s appareil is the hopeful one: a device that opens a margin of adjustment. Flusser’s Apparat is the pessimistic one: a black box that programs you. Agamben’s dispositivo is the widest: everything that captures behaviour. In Agamben’s essay, gestures are among what apparatuses capture, so Citton’s “gesture vs machine” is, in Agamben’s vocabulary, “gesture vs apparatus”. Same idea, opposite word. Keep the table in mind whenever you read these authors side by side.
• The idea underneath is Simondon’s. The “margin of indeterminacy” that makes a machine more advanced is, from the user’s side, the settings. An apparatus in Huyghe’s sense is a Simondonian open machine. And Simondon’s claim that alienation comes from not understanding machines is why “exiting the machines means having a bit of technical knowledge”.
• Benjamin’s optical unconscious shows what an apparatus can do for perception: slow motion, close-up and aperture reveal what the eye alone cannot see. Settings are not just control. They are ways of seeing. Removing them removes ways of seeing.
• Flusser’s “playing against the apparatus.” The photographer who forces the camera to do what its program didn’t foresee is the free one. Citton’s pocket-selfie is a playful instance: it uses the device but breaks the regime of use.
• The regime-of-use concept is the most useful tool on this page. It separates what a technology can do from what its market and habits make it do. Two phones with the same hardware can have very different regimes of use, set by defaults, interface design, business model and social norms. This is also why Citton can call the technology a miracle and the economy a calamity: the regime, not the device, does the harm.
• Contemporary versions of the same fight:
  • Right to repair (EU and US laws in the 2020s): the right to open the black box.
  • Open-weight vs closed AI models: an open-weight model with exposed parameters (temperature, sampling, system prompt, fine-tuning) is close to Huyghe’s apparatus. A consumer chat app with one text box is closer to a regime of use.
  • Dark patterns: interface designs that steer users (infinite scroll, autoplay, default opt-ins). These are regimes of use built into design.
  • Computational photography: a phone camera makes several frames and merges them with ML before you see the photo. The “settings” now happen inside the model, out of reach. This is Flusser’s program, literally.

📚 Roots: Philosophy of technics — machines, black boxes, pharmaka

The problem. Is technology a neutral tool, a threat to the human, or part of what makes us human? The French tradition (Simondon, Stiegler) takes the third view: there was never a human before technics.

1. Lewis Mumford, The Myth of the Machine (1967). The megamachine: big organisations of people (armies, pyramid workforces, bureaucracies) were the first machines, before any metal ones. This is why the afternoon sessions spoke of humans who become machinized.
2. Gilbert Simondon, On the Mode of Existence of Technical Objects (1958). This is the most important unnamed source for this talk. Three ideas:
   • Automatism is not perfection. Simondon argues that a machine’s real progress lies not in more automatism but in a “margin of indeterminacy” that keeps it open to information from outside: from the operator and the environment. A fully automatic machine is closed and, technically, primitive. This is almost Citton’s definition of gesture, now applied to machines. It is also the root of Huyghe’s apparatus with reachable settings.
   • Alienation comes from ignorance. We are alienated from machines not because they exist, but because our culture refuses to understand them. We treat them as either slaves or idols. Simondon wanted humans to be “conductors of an orchestra” of machines. This is the background to Citton’s “exiting the machines requires a bit of technical knowledge”.
   • Critique of hylomorphism. Simondon’s brick example: clay is not passive matter to which the mould gives form. Clay and mould meet through forces. Ingold later builds on this.
3. Walter Benjamin, “The Work of Art in the Age of Its Technological Reproducibility” (1935–36). The camera reveals an “optical unconscious”: slow motion and close-up show what the eye cannot see. For Benjamin, film is a training ground where humans learn to live with the enormous apparatus of modern technology. A hopeful view of the apparatus, close to Huyghe’s.
4. Vilém Flusser, Towards a Philosophy of Photography (1983). The pessimistic view of the apparatus. The camera contains a program. Most photographers are functionaries who only exhaust the program’s possibilities. The camera takes the photograph through them. Freedom means playing against the apparatus: forcing it to produce what its program did not foresee. Citton’s “photograph your pockets in front of the Eiffel Tower” is a Flusserian exercise in exactly this.
5. Bernard Stiegler (1952–2020), Technics and Time (1994–2001) and later work. He combines Leroi-Gourhan and Derrida:
   • Tertiary retention. Memory stored outside the body in technology (writing, film, data). Human consciousness has always depended on it.
   • Pharmakon. From Plato via Derrida: every technology is both poison and cure. Writing weakens memory and makes philosophy possible. The question is never “for or against”, but how to care for a technology so its cure side wins.
   • Proletarianisation. Workers lost making-skill to machines, consumers lost living-skill to marketing, and now thinkers lose theoretical skill to algorithms. The afternoon sessions discussed Stiegler. Citton’s split between “the technology is a miracle” and “the economy is a calamity” is a pharmakon reading.

Where others push back.

• Defaults are a gift too. Good defaults let non-experts take good photos. Removing settings widened access. Most people do not want to set aperture. The democratic case for automation is strong.
• Settings can be a regime too. Pro-mode cameras have their own culture and norms (the “correct” exposure). Access to settings does not guarantee freedom from a regime.
• Huyghe’s distinction is not used widely in English-language theory, so it is hard to place in that debate. Simondon’s open/closed machine is the closest well-known equivalent.

What he leaves out. That language is now a setting. With LLMs, the prompt is the control surface, so writing becomes a way of adjusting the apparatus. Whether a prompt box is an apparatus (open-ended, adjustable through language) or a regime of use (a chat that steers you to quick answers) is a real open question that his frame invites but does not answer.

Check yourself

• Give one tool you use that is an apparatus and one that is a regime of use. Could the same tool be both?
• Explain why Huyghe’s “apparatus” and Agamben’s “apparatus” are near-opposites.
• Why does exiting the machines need technical knowledge? Which thinker gives the deeper reason?

---

5. Generative AI breaks his own thesis

In short. Answering an audience member, Citton says generative AI breaks his own thesis. Expert systems fit his definition of a machine, but machine learning corrects itself, and even its programmers cannot control it. So there may be gestures in generative AIs, and he finds that interesting. He adds guardrails about life and about behaviour versus being.

🎙 In the talk. This is the most honest moment of the evening. Answering an audience member, Citton says AI comes in two kinds. Expert systems match his definition of a machine: “for the machine to be able to react to every situation, we have to program every possible situation … a closed set”. Machine learning does not: “there is a process of self-weighting within the computational system … the machine reacts to vast amounts of input; it corrects itself”. So “generative AIs are irreducible to the way they were programmed. The big problem is that we don’t know what they are going to do. Even the people who programmed them cannot control them.” This “throws everything I told you out the window”. The uncontrolled, the adaptive and the unexpected are all there. The beginning? “Chatonsky would say yes.” His conclusion: “Are there gestures in generative AIs? … Yes. And in my opinion, that becomes interesting”.

(Caption slip: “works very well with machine learning, with expert systems”. From context, he means expert systems only.)

He then sets guardrails:

• Life: “where does the electricity come from?” Until machines “can build and maintain power plants … we are wasting our time thinking about a certain type of thing”.
• Behaviour vs being: self-adjusting weights “resemble quite closely what we behaviorally identify as animate beings” and it is important to ask in what way, and also to say why it is not life.
• Technology vs economy: “The political economy of AI is an absolute calamity … criminal”, yet the technology itself is a miracle.

📚 Deeper theory

• His reversal follows AI history exactly (Roots). His definition of “machine” is the GOFAI/expert-system definition. Dreyfus showed why that kind of AI could not reach expertise. Connectionism and deep learning gave up hand-written rules. Sutton’s “Bitter Lesson” says that giving up rules is why they work. So Citton’s gesture/machine line was, without his saying so, a line between Dreyfus’s embodied skill and GOFAI. Deep learning moves across it.
• What “irreducible to its programming” means technically. The code that trains a model is short. Its behaviour comes from billions of weights learned from data, which no one wrote and no one can fully read. So “not programmed by anyone” is fair. But the trained model is deterministic given its weights and input, plus controlled randomness (temperature sampling). “Unpredictable to us” is not the same as “uncontrolled” in the sense of “passes through me”. The open question is whether opacity plus learning is enough for “gesture”, or whether gesture also needs a body with something at stake.
• The two camps (Roots). Stochastic parrots and mean images: outputs are averages, and nothing new begins. Emergence: new capacities appear that no one designed. Citton leans toward the second, as an artist like Chatonsky would. Steyerl, whom he cites approvingly for address, leans toward the first. This tension is unresolved in the talk.
• Simondon would not be surprised. A machine with a large margin of indeterminacy, open to information from its environment, is Simondon’s advanced machine. Deep learning is a radical case. What Simondon would ask: are we its “conductors”, who understand and coordinate it, or its users and worshippers? That is the regime-of-use question again.
• The electricity test, unpacked. It is a materialist move. Living bodies are part of metabolic cycles (water, food, sun) “that come from very far away”. Machines depend on a grid that humans build and maintain. Autonomy, in this view, is metabolic before it is cognitive. This links to research on the material footprint of AI (data centres, minerals, water), which Kate Crawford maps in Atlas of AI (2021).
• Agency vs life vs soul. In the poem discussion, Citton separates three questions. Alive? No. Soul? “Probably not, because the way humans have constructed the notion of a soul does not apply.” Agency? Yes, and agency “deserves to be recognized”. This is Latour’s position (idea 7): actants without life. His nominalism adds that “intelligence” and “alive” are words that cultures use to orient and reassure themselves. They are illusions, but necessary ones.

📚 Roots: A short history of AI — why “programmed” stopped meaning “machine”

The problem. Citton’s whole thesis rests on “machine = programmed”. AI history shows when that stopped being true, and so why he reverses himself in the Q&A.

1. Symbolic AI / GOFAI (“good old-fashioned AI”), 1956–late 1980s. From the Dartmouth workshop (1956) on, intelligence was treated as symbol manipulation by rules. Its peak was the expert system: MYCIN (Stanford, 1970s) diagnosed blood infections using about 600 hand-written rules. Edward Feigenbaum’s motto: “knowledge is power”. The approach broke on Polanyi’s paradox: experts could not state their rules, and the systems failed on cases nobody foresaw. This led to the “AI winter” of the late 1980s. This is exactly Citton’s “machine”: you must program every case.
2. Dreyfus’s critique (1965 onward). He predicted this failure from Heidegger and Merleau-Ponty: intelligence is embodied skill, not rules. The AI establishment mocked him. Later many accepted much of his diagnosis, at least for rule-based AI.
3. Connectionism — the other tradition. The perceptron (Rosenblatt, 1958) learned from examples. It was sidelined by Minsky and Papert (1969), then revived by backpropagation (Rumelhart, Hinton and Williams, 1986). These systems are not given rules. They adjust their own weights from data. This is the “self-weighting” in the talk.
4. The deep-learning turn. AlexNet (2012) for images, the transformer architecture (2017) for language, then GPT-style large language models. Richard Sutton’s essay “The Bitter Lesson” (2019) summarises 70 years: methods that build in human knowledge lose, in the long run, to general methods that learn from large amounts of data and compute. In other words, the less we program, the better it works. That is the opposite of the expert-system world, and the reason Citton says it “throws everything I told you out the window”.
5. The live debate: interpolation or emergence?
   • “Stochastic parrots” (Bender, Gebru et al., 2021): LLMs stitch together statistically likely text with no understanding. Steyerl’s “mean images” belongs to this camp: the output is the average.
   • “Emergent abilities” (Wei et al., 2022): new abilities appear suddenly at scale. Critics (Schaeffer et al., 2023, “a mirage?”) showed that some of this suddenness comes from how abilities are measured.
   • Interpretability research tries to open the black box: what do the internal features actually represent?
   • Where Citton stands: closer to the emergence side than you might expect from a critic of technology. Generative AI is irreducible to the way it was programmed … even the people who programmed them cannot control them. He keeps his critique for the economy, not the technology.

Where others push back.

• From the critical side (Bender, Steyerl): calling statistical output “gesture” romanticises it and hides the labour, data extraction and averaging behind it.
• From the embodiment side (Dreyfus’s heirs, the audience member): gesture needs a living body with emotion and risk. Without stakes, there is no gesture, only variation.
• From the AI side: the models are less mysterious than “no one can control them” suggests. Alignment training does shape behaviour heavily, and interpretability is making progress. “Uncontrolled” is an overstatement, though “not fully understood” is fair.

What he leaves out. The fact that the regime of use shapes the gesture of the model itself. Alignment training, system prompts and product design narrow what a model does. In his own terms, a generative model is an apparatus with enormous settings, wrapped in a strong regime of use. That would connect deep dives 4 and 5, but he does not make the link.

Check yourself

• Why do expert systems fit his definition of “machine”, and why does machine learning break it?
• What is the difference between “unpredictable to us” and “uncontrolled” in the gestural sense?
• State the electricity test. What does it take as its basis for autonomy?

---

6. Address: singular vs machinic (Q&A)

In short. Platforms address us through our profiles, which hook us. He calls this machinic address. Gestural address goes between singular bodies and minds. Steyerl’s “Mean Images” says statistical images are mean in both senses because they do not address our singularities.

🎙 In the talk. Discussant A asks about toucher à distance and “the power of address”, and how to teach gestures that are wasting away. Citton answers in three moves.

1. Expressive gesture. His 2012 book treated gesture as expressive. This talk set that aside. Writing is a gesture: “Rousseau … made writing gestures in the 18th century, and a century and a half later, it moves me”.
2. Momentum. A gesture is “caught up in a momentum that goes beyond us”, and that momentum is contagious. Arab Spring phone videos carried “an emancipatory momentum … from a distance through the image of the gesture”.
3. Address. Platforms address us too, “because it’s my profile that suggests to the machine that if it recommends such image after such image … I will get hooked”. That is machinic address. Gestural address goes between singularities: “it is my body … I know I have before me minds, bodies also completely singularized”. He can address ten people on Zoom; with “blank screens and 200 people … I’ll create a profile of the audience”. Hito Steyerl’s “Mean Images” says that statistical images are mean in both senses, because they don’t address our singularities.

📚 Deeper theory

• Deleuze’s “dividual” is the exact concept here (Roots). The profile that platforms address is not you. It is you divided into data points and recombined with millions of similar data points. Machinic address speaks to the dividual. Singular address speaks to the individual.
• Why averaging is cruel, historically. Statistics of the “average man” (Adolphe Quetelet, 1835) and Francis Galton’s composite photographs (1870s–80s), which overlaid faces to find “types”, including criminal types, fed directly into eugenics. Steyerl’s “mean” draws on this history: the average was never neutral. It was used to rank and exclude. The Cartema summary of her essay names the “hygienist and even eugenicist” potential of averaging.
• Benjamin’s aura vs reproduction. Rousseau’s writing still “touches” from a distance because it was a singular address, even when printed millions of times. The question for AI writing: can a text addressed to no one in particular ever touch from a distance?
• Contagion of gestures. This connects to Gabriel Tarde’s Laws of Imitation (1890). Society is built from imitation spreading between people, an idea that has returned with network science. Citton has written about Tarde elsewhere. The Arab Spring example also connects to debates about whether social media helped or only looked like it helped (Zeynep Tufekci, Twitter and Tear Gas, 2017: networked movements grow fast but lack the organising capacity that slower movements built).
• Address as design. Rouvroy’s algorithmic governmentality (Roots) explains why machinic address is politically different: a subject who is addressed can answer back. A profile that is predicted cannot.

Why this matters for you. It is a design test for any AI product that speaks to people. Does it answer this person, or the average person like them? The test fits personalisation in government services, where averages fail the people at the edges.

📚 Roots: Address, profile and control — why “mean” means cruel

The problem. The Q&A’s most original thread: platforms speak to profiles, not persons. This has deep roots.

1. Rousseau. Citton and Discussant A credit him with “touching from a distance” (toucher à distance). Writing lets one singular person move another across centuries.
2. Walter Benjamin’s “aura” (1935). The “unique appearance of a distance, however near it may be”: the here-and-now presence of an original. Mass reproduction strips the aura. A singular address keeps something like it.
3. Gilles Deleuze, “Postscript on the Societies of Control” (1990). Foucault had described disciplinary societies (schools, factories, prisons) that mould individuals in closed spaces. Deleuze says we are moving into societies of control: continuous, open, modulated by codes and data. His key line: “individuals have become ‘dividuals’, and masses, samples, data, markets, or ‘banks’.” A platform profile is a dividual. It is you broken into data points and recombined with millions of others.
   • Note: the radar-operator model of attention (idea 2) is disciplinary in Foucault’s sense. The platform feed is control in Deleuze’s sense. The Friends of Attention attack the first. Steyerl and Citton’s “machinic address” attack the second.
4. Antoinette Rouvroy and Thomas Berns, “algorithmic governmentality” (2013). Algorithms govern by predicting behaviour from data and shaping the environment in advance, without ever addressing a subject who could argue back.
5. Hito Steyerl. Earlier: “In Defense of the Poor Image” (2009) and The Wretched of the Screen (2012). Later: “Mean Images” (c. 2023). AI images are mean (statistical averages of training data) and mean (cruel), because averaging has a history in statistics and eugenics (Galton’s composite photographs, 1870s–80s) and because it rests on precarious, low-paid data workers (Cartema abstract).

Check yourself

• What is the difference between address to a profile and address to a person? Use Deleuze’s term.
• Why does the history of the “average man” make “mean” mean cruel?
• Can momentum spread through machinic address? What did Tufekci find?

---

7. Losses and recompositions: photography as the model for AI (Q&A)

In short. Each medium loses something, and handwriting carries what typing loses. Citton takes photography as his model. It was a scandal at first, yet painters were not all ruined, and they started painting differently. He expects new forms of writing and teaching that accept what is done automatically, and he sees large fields left for writing as art.

🎙 In the talk. An audience member asks why gestures are not recorded alongside writing. Citton turns this into a question about what each medium loses. Handwriting carries what typing loses: graphology, Guérin’s drawing. If students don’t build sentences themselves, “even if it’s with a keyboard … they will undoubtedly lose something. Is that the end of the world? Maybe, I don’t know”. His model is photography. It was “a scandal at first … much like AI is a scandal for translators today”, because portraitists lost their trade. He himself can’t draw, and “I would live very differently if I knew how to sketch like Tim Ingold”. Yet “the painters were not all ruined by photography. We started doing painting differently.” So he expects new forms of writing and teaching “that will acknowledge what is being done automatically”, which could be “fantastic … for my little farm of literary studies”. Machines can imitate Beckett or Mallarmé “based on averages”, but écriture, writing as art and not as communication, has “enormous fields” left.

📚 Deeper theory

• The photography story in more detail. The daguerreotype arrives in 1839. Painter Paul Delaroche is said to have declared “from today, painting is dead” (probably apocryphal). Baudelaire’s Salon of 1859 attacked photography as “the refuge of every would-be painter”, fit only to be “the servant of the sciences and arts”. What happened next: miniature portrait painters did lose their market. Painting then moved away from what photography did well (likeness) toward what it couldn’t do: Impressionism (light and perception), then abstraction. Photography itself became an art. Pattern: a medium automates one function, and the older art redefines itself around what was not automated.
• Benjamin’s twist (1935): the debate over “is photography art?” missed the point. Photography changed what art is: reproducibility, the loss of aura, new forms of seeing.
• Ong and the history of writing (Roots). Each shift (speech to script, script to print, print to screen) restructured thought, not just output. If LLMs automate composition, writing may shift from producing sentences to choosing, judging and arranging them, closer to editing or curating.
• Handwriting research. Studies of note-taking (e.g. Mueller and Oppenheimer, 2014, “The pen is mightier than the keyboard”) found better conceptual understanding with handwritten notes. Later replications gave mixed results. The effect is debated but not dismissed. This is a real-world example of a gesture (handwriting) doing cognitive work.
• Écriture vs communication. In French theory since Barthes (Writing Degree Zero, 1953), écriture means writing that is aware of itself as a material practice, not just a channel for messages. Citton’s hope is that if machines take over communicative writing, the literary or artistic gesture of writing becomes more visible and valuable, as painting did after photography.

📚 Roots: Two shorter threads

Sensation and literacy

• Walter Ong, Orality and Literacy (1982): writing restructures thought. Oral cultures think in formulas, stories and the present situation; literate cultures in lists, abstractions and analysis.
• David Abram, The Spell of the Sensuous (1996): the alphabet turned our senses away from the animate landscape and toward the page. Citton: “I believe in that a little bit, too”.
• Baptiste Morizot: philosophy as tracking. Learning to read wolf tracks retrains attention toward other living beings. This links back to Ginzburg’s hunter (idea 3).

Agency without life–end

• Bruno Latour, actor-network theory (1980s on): non-humans (microbes, door-closers, speed bumps) are actants. They make a difference to what happens, so they belong in the account.
• Donna Haraway, “A Cyborg Manifesto” (1985): the human/machine boundary was never clean.
• Rights of nature: New Zealand’s Whanganui River received legal personhood in 2017. Citton’s extension: if rivers, why not, one day, machines?
• The surrealists: André Breton’s Manifesto (1924) defined surrealism as “psychic automatism in its pure state”: automatic writing to free the unconscious from rational control. This is the audience’s point. In art, automatism has long been a route to the uncontrolled, not its opposite.

---

Where others push back.

• Scope. Photography automated one skill (likeness). LLMs touch many cognitive skills at once: summarising, drafting, translating, coding, explaining. The recomposition may be broader, faster and harder.
• Economics. Painters were not “all ruined”, but many portrait painters were. A recomposition at the level of culture can still mean real loss for real people. Citton’s “is that the end of the world? Maybe” recognises this without resolving it.
• Learning vs production. A painter after 1839 could choose not to paint likenesses. A student who never builds sentences may never get the skill that écriture depends on. The recomposition only works if teaching protects the base gesture.

Check yourself

• What did painting do after photography? What is the general pattern?
• What is the strongest objection to the photography analogy for AI?
• Why does Citton think literary studies might gain from LLMs?

Q&A

(Add questions and answers here only if worth keeping.)

---

About the source and how this page was made

• Speaker: Yves Citton
• Event: Rencontres de Sophie (3rd edition), architecture school, session on machines
• Date: ~Feb 2026 (inferred — he says Attensity! came out in English “a month ago”)
• Length: 1:31:12

Speakers

Label	Who	Role
Host	Maël Tessier (name from captions; spelling may differ)	Professor at the host architecture school; introduces
Citton	Yves Citton	Professor of literature and media, Paris 8. Main speaker
Discussant A	Anne Philippe (named by host)	Educator and artist; asks about “touching from a distance” and address
Discussant M	Mélanie Richer (named by host)	Asks about “animated gestures” and sensation
Audience	4–5 people, not named	Ask about gesture vs writing, life vs machines, a poem on machine agency, surrealist automatism

Caption warnings. This is an auto-translation of French speech. Names are often wrong in the captions. Corrections used on this page: “attité / intensity” = attensity; “Teaming Gold” = Tim Ingold; “Châonski / Chatonski” = Grégory Chatonsky; “Pierre Damia / Hug” = Pierre-Damien Huyghe; “Guéri” = Michel Guérin. The French appareil is translated as both “apparatus” and “device”. Here it is apparatus. Factual slip in the talk: Citton guesses that Frank Gehry died “around 80”. Gehry died on 5 Dec 2025, aged 96 (Al Jazeera).

---

🎙 marks what the speaker says. 📚 marks added background and the wider debate, not the speaker’s claim. † marks background from general knowledge that was not checked against a source.