I am preparing to run a month-long residency with NEF Animation on the ecology of storytelling in animation cinema, where the use of artificial intelligence is a matter of dispute among filmmakers. These quarrels bear on generated images, that is, on what can be seen. Drawing on the sociologist Susan Leigh Star’s notion of infrastructure, I show that most of artificial intelligence operates where nobody is looking, and that those who claim never to use it are using it all the time.
The quarrel over generated images
In creative circles, a conflict has taken hold between authors, aimed at those who generate all or part of their images with artificial intelligence. Animation cinema offers a telling example. Films have been attacked at festivals because a machine had made some of their visual elements, sometimes only the textures of the images, one component among many. What triggers the anger is the generated pixel, what can be seen, or what people believe they can detect.
Yet nobody asks how a film’s research material was gathered, how the successive versions of the script were compared, how the schedule and the team’s cooperation were organised. A film can be deeply worked on by these tools without a single image carrying any trace of it, and that is in fact the most common situation. My own practice largely works this way. I use generative tools upstream, where I search, where I compare, where I put my thinking in order, far more than in the visible surfaces of what I produce. Judging the presence of artificial intelligence by the final image is looking for your keys under the streetlamp because that is where the light is. To understand why we look in the wrong place, we need the notion of infrastructure.
What Susan Leigh Star calls infrastructure
The sociologist Susan Leigh Star published in 1999 an article that has become a classic, “The Ethnography of Infrastructure” (American Behavioral Scientist, vol. 43, no. 3), which she presents as a call to study boring things, electrical plugs, standards, bureaucratic forms. In it she describes the properties of what becomes infrastructure, and several of them bear directly on our subject.
An infrastructure is embedded, caught inside other structures and other arrangements, so that it cannot be pointed at as a separate object. It is transparent, supporting tasks without having to be reinvented at each use. It is learned through practice, by becoming a member of a community, which is why what is an invisible given for some remains a visible obstacle for others. Star takes the example of the water network, a limpid infrastructure for the city dweller, a daily object of work for the plumber. It is built on an installed base, inheriting the strengths and weaknesses of what existed before it. And it only becomes visible again when it breaks down. Nobody declares that they use electricity or the road network, these things are the ground we walk on, until the day the power goes out.
This description has a consequence that concerns every quarrel about artificial intelligence. When a technique becomes infrastructure, the question “do you use it?” gradually loses its meaning, because use is no longer a decision, it is embodied in tools, gestures and habits. One can answer in complete good faith that one does not use it, exactly as one would answer that one does not use the land registry or electrical standards, while relying on them at every moment.
Artificial intelligence has slipped below the waterline
The situation can be pictured with the image of the iceberg. The emerged part, the one everyone talks about, is the conversational bots, ChatGPT, Claude and their kind, which we use knowingly, by choice, and about which we argue. The submerged part, immensely larger, is the artificial intelligence embodied in the tools we no longer even think of questioning.
Google’s search engine is the most massive example. In October 2015 Google confirmed its use of RankBrain, a machine learning system that enters into the ranking of results, presented at the time as one of the three most important signals in its algorithm. In October 2019, Pandu Nayak announced on the Google blog the deployment of BERT, a language model used at launch on roughly one query in ten in English in the United States. To search on Google is to put machine learning models to work, and it has been for more than ten years. The list goes on with the spellcheckers in our word processors and their suggested phrasings, translation tools, the filters that sort out junk mail, the routes computed by our map applications, the order of messages on social networks, the matches on dating platforms, and the cameras in our phones, which recompute every image at the moment it is taken. Someone who states that they never use artificial intelligence, and who does all of this every day, is not lying. They are describing an infrastructure in Star’s sense, learned through practice, embodied, transparent, and their statement gives the exact measure of the degree of invisibility these systems have reached.
Netflix has been making series from its data since 2013
There is one more level below the waterline, that of industrial use, and it changes the scale of the subject. Netflix offers the most documented example. Carlos Gomez-Uribe and Neil Hunt, then in charge of product and personalisation at Netflix, published in 2015 a reference article, “The Netflix Recommender System” (ACM Transactions on Management Information Systems), in which they write that their recommendation system influences roughly 80% of the hours streamed on the platform, the rest coming through search, and they estimate the combined effect of personalisation and recommendation at more than one billion dollars a year. And recommendation is only part of the use, the same data guides production decisions. Netflix’s first original series, House of Cards, launched in 2013, had its framework defined from the analysis of viewing behaviour. I told that story in the article The connected audience. The audience believes it is choosing entertainment, when it is in fact watching the product of a chain in which artificial intelligence intervenes from audience analysis all the way to the thumbnail displayed on the screen.
What Netflix does spectacularly, most large organisations do routinely. The State of AI survey that McKinsey publishes every year indicates, in its 2025 edition based on around two thousand organisations, that 88% of them use artificial intelligence in at least one function, logistics, fraud detection, maintenance, pricing, recruitment, customer relations. The sample leans towards large, already well-equipped structures, and the figure nonetheless says something simple. Artificial intelligence is an ordinary industrial working tool, far ahead of our conversations with a chatbot. Believing that the heart of the phenomenon lies in our personal use of ChatGPT is an error of scale. We are part of these systems as customers, users, viewers and patients, far more than as willing operators.
Keeping the gaze on the individual
This error of scale is not a mere misunderstanding, it suits certain interests. When public debate concentrates on individual uses, should one use ChatGPT, should one declare it, should one abstain, people commit themselves, judge one another and quarrel among themselves, while industrial deployment proceeds without meeting that critical fire. The mechanism has precedents. In the early 2000s, the oil industry, with British Petroleum in the front rank, popularised the notion of the individual carbon footprint and put online the first personal calculator, shifting responsibility for global warming onto everyone’s private behaviour. I developed this parallel between the critique of artificial intelligence and the history of the carbon footprint in the article The last straw. The common point is the framing of the gaze. As long as we discuss the virtue of individuals, we do not discuss the organisation of industries, and the quarrels between authors over generated images take part in that framing, without anyone intending it.
Becker’s invisible workers, from the credits to the royalties
This shift of attention, from visible surfaces to the making, Howard Becker prepared long ago. He established, in Art Worlds (1982, French translation Flammarion, 1988), that no work is the product of a single person. Every work is born of a chain of cooperation in which what he calls support personnel does its work, equipment manufacturers, technicians, assistants, editors, distributors, and this chain is governed by conventions, shared habits whereby some people are credited and others remain invisible. The history of art is full of these invisible figures, the workshops of the Flemish masters where anonymous hands painted the skies, the ghostwriters whose books bear other people’s names, the women film editors of early cinema, the arrangers of popular song.
There is a further degree, which Becker does not take up and which weighs heavily in the fields where I work. Being credited and receiving royalties are two distinct things. French law offers the starkest version of this. The intellectual property code, in its article L113-7, presumes as co-authors of a film a closed list of writing and directing functions, the script, the adaptation, the dialogue, the original music and the direction. Everyone else, cinematographers, film editors, production designers, costume designers, is not considered an author, whatever the importance of their contribution, and will never receive author’s royalties. They appear in the credits, and the credits open no rights. It is into this world that the question of artificial intelligence arrives, a world where recognition of work and payment for it have never coincided, and where the debate on plagiarism and intellectual property goes back a long way. I have set out my position on this point in the article Copyright and Artificial Intelligence. Seen from Becker’s standpoint, the present unease becomes legible. Artificial intelligence has entered the chain of cooperation as new support personnel, and the crediting conventions concerning it do not yet exist. Everyone improvises, between silence, contrite confession and denial, which is the normal state of an art world in the process of renegotiating its conventions.
Norman McLaren drew sound directly onto the film stock
Animation cinema, the very field where the quarrel over generated images is fiercest, carries in its history a lesson about the invisible. Norman McLaren, who joined the National Film Board of Canada in 1941, engraved the soundtrack of his films with a stylus onto the optical track of the film stock, obtaining percussion that nobody had played. He also made films without a camera, painting on clear stock, or scratching opaque black stock with a blade and a needle, as in Blinkity Blank in 1955. He practised pixilation, in which real people are animated frame by frame. He almost never came to the image by the direct route.
Georges Sifianos reports, in Esthétique du cinéma d’animation (Le Cerf, 2012), the definition McLaren gave of his art. Animation is not the art of drawings that move but the art of movements that are drawn, the difference between two successive frames matters more than the image each one carries, and animation is therefore the art of manipulating the invisible intervals that lie between the frames. Here is an art one of whose great figures locates the essential in the interval nobody sees, whose sound is drawn as much as its images, and which is judged today on its pixels. There was a political position in McLaren’s work, held in these technical detours and in his refusal of heavy means. He ran audiovisual training courses for Unesco, in China in 1949 and in India in 1952, and he said that if all his films had to disappear but one, he would keep Neighbours, his 1952 film against war.
The debate about film stock has ended, the video tapes have remained
Quarrels over the legitimacy of a technique have a lifespan, and the history of cinema provides two instructive cases. I remember the debates that accompanied the arrival of digital cinema, and the question of whether digital was still cinema. They were fierce, they filled the journals and the festivals, and they have ended. The same had happened with video, held to be a medium without nobility. Yet the nobility of cinema rested in large part on its cost, and video, far cheaper, was taken up by the women who had no access to the means of cinema.
Carole Roussopoulos, a pioneer of lightweight video in France, taught video at the University of Vincennes from 1973 to 1976 and trained many women in the practice. In 1982, with Delphine Seyrig and Ioana Wieder, she founded the Centre audiovisuel Simone de Beauvoir, precisely because the tapes shot over the previous ten years were fragile and had to be preserved. The work continues. Between 2009 and 2013, the Médiathèque Valais digitised most of the Carole Roussopoulos collection with the support of Memoriav.
Here we find, almost point by point, what Star describes. When the legitimacy debate dies down, the medium disappears into infrastructure, and what remains is an infrastructure problem, materials that decay, abandoned formats, playback machines that no longer exist, on which what reaches us depends. The quarrel over generated images will follow the same path as the one over digital and the one over video, and what will remain after it are the infrastructure questions it covers up today, those of the tools, the data, and the people who keep them running.
The invisible humans inside the machine
Among these covered-up questions is that of the humans hidden inside artificial intelligence. Mary L. Gray and Siddharth Suri documented in Ghost Work (Houghton Mifflin Harcourt, 2019) the vast invisible workforce that keeps supposedly automatic systems running, the people who annotate training data, who filter violent content, who correct the outputs of models, often from low-wage countries, without status or recognition. Every generated image, every fluent answer rests on this work, as the Flemish master’s workshop rested on its anonymous hands.
To be consistent, we would have to state both, what share the machines took in the work, and what share of human labour was taken into the machines. Asking the second question shifts our attention away from aesthetic suspicion about the authenticity of a work and towards the conditions in which all of this is made, and to my mind that is the question that counts.
The production note I write for my own work
Legislators have taken up the subject. The European regulation on artificial intelligence, adopted in 2024, provides in its article 50, applicable from 2 August 2026, that content generated or manipulated by artificial intelligence systems must be marked in a machine-readable format, and disclosed to the public in several situations. This is useful against deceptive uses, where the question really is whether an image or a voice is synthetic. For works of art, marking answers yes or no where the making is a chain, with degrees and with locations. The drafters partly acknowledged this, since the text provides exceptions for systems that perform a standard editing assistance function, in other words for artificial intelligence that has already passed into the state of infrastructure.
I argue for something else, a production note, voluntary and precise, in the spirit of the credits. In it I would state, step by step, writing, research, image, sound, editing, what was done by hand, what was assisted, what was generated and then reworked, what was generated as is, and with which tools. This is what I make a point of doing for my own work, including for the writing of texts such as this one, and I find that this precision makes what I say more credible rather than weakening it.
I have no illusions about how complete such a note can be. It will always remain partial, for the very reason I have just described, some of the tools used are not perceived as artificial intelligence tools by the people using them. And I think the quarrel over generated images is becoming obsolete, as the one about digital against film stock has become. In a few years, nobody will ask a filmmaker whether their film was made with artificial intelligence, for the same reason that nobody is asked any more whether their film was edited on a computer. What will remain are the infrastructure questions, those of the tools embodied in our gestures, of the data that steers productions, and of the people who work inside the machine, which are matters of pay and status before they are matters of authenticity. It is in order to describe that work that the production note is worth writing.
