A 200-gram steak emits as much greenhouse gas as 4,000 queries to an artificial intelligence. A person who writes fifty prompts a day for a year produces a carbon impact equivalent to eating three or four steaks. I checked this calculation against figures published in 2025 by ADEME, Mistral AI and Google, then asked myself why we feel guilty in front of our screens and not at the restaurant. The history of the individual carbon footprint, invented by the oil company BP, and the robustness of living things described by Olivier Hamant help me answer that question.
Five kilos on one side, one gram on the other
In The skill as a work of art, I had compared a steak and queries through the water they consume. For carbon, I went back to the sources.
For beef, I used the Agribalyse database of the ADEME, which is the reference in France: between 26 and 28 kilos of CO2 equivalent per kilo of French beef, counting livestock farming, animal feed, slaughter and transport, which comes to about 5.5 kilos for a 200-gram steak. Joseph Poore and Thomas Nemecek, who published a study of food at global scale in Science in 2018, also count the forest cleared for pasture and fodder crops, and they reach nearly 100 kilos per kilo for beef from beef herds, around twenty kilos for the same steak.
For queries, the figures move a lot. In 2023, people spoke of about 4 grams per query. In July 2025, Mistral AI published the first full life-cycle assessment of a large language model, carried out with the consultancy Carbone 4 and supported by ADEME: 1.14 grams of CO2 equivalent for a 400-token answer from its assistant Le Chat (a few hundred words), including a share of model training and server manufacturing. A month later, Google announced 0.03 grams for a median Gemini prompt. Shaolei Ren, at the University of California, Riverside, and Alex de Vries-Gao, at the Vrije Universiteit Amsterdam, pointed out as soon as it was published that Google counts the renewable electricity it buys by contract, and not the electricity its data centres actually consume. One should add that a generated image or video costs much more than a short answer, and that the same query emits several times more where electricity comes from coal than where it comes from nuclear power or dams.
Dividing 5.5 kilos by 1.14 grams gives a little under 5,000 queries per steak. I had read elsewhere the figure of 3,000, which you get with a 2-gram query, as in older estimates, and a 6-kilo steak. With Google’s figure, more than a hundred thousand. I kept 4,000 in the title because it is the figure that circulates, and because it is rather unfavourable to artificial intelligence.
A person who writes fifty prompts a day, which means working with artificial intelligence from morning to night, writes a little over 18,000 in a year, about 21 kilos of CO2 equivalent at Mistral’s figure: a little under four French steaks.
Of course, the figures I put forward are only an estimate. They could vary by a factor of ten depending on the data centre, the length of the queries, and the way the cost of training is spread over the number of users. It is obviously much more complex than that. But the orders of magnitude are robust enough to draw a consequence from them, and the main consequence is this: the choice we make at a restaurant, which seems trivial when we are among friends or with family, has an ecological impact several thousand times greater than the choice we make in front of a screen.
The individual carbon footprint, a BP invention
I told in The last straw how the oil company BP popularised the notion of the individual carbon footprint in the early 2000s, and put the first personal carbon footprint calculator online in 2004. The “Beyond Petroleum” campaign had been designed by the advertising agency Ogilvy & Mather, and the calculator was used by 278,000 people in its first year. The expression carbon footprint was then almost unknown to the public. Geoffrey Supran, who studies oil companies’ discourse on climate at the University of Miami, estimates that BP spent more than a hundred million dollars a year on this campaign between 2004 and 2006. In his view, the aim was to place responsibility for climate change on consumers, and thereby to hide the fact that the emissions of the industries themselves are out of all proportion with what changes in citizens’ habits could achieve.
Karl Smerecnik and Valerie Renegar, two communication researchers, analysed BP’s next campaign, Helios Power, in 2010, and called this mechanism “capitalistic agency”: through its advertising, BP offers consumers a capacity to act for the environment, but only in the forms that produce capital, and rules out those that would harm profit. In Political awareness and political conscience, I called political good conscience this feeling of being on the right side that takes the place of political awareness, that is, of the capacity to think a problem through at the level where it is played out.
For France, the consultancy Carbone 4 calculated in 2019 what individual actions can achieve, in a study by César Dugast and Alexia Soyeux entitled “Faire sa part ?” (“Doing one’s share?”). A person who did absolutely everything, what the two authors call heroic behaviour, would reduce their footprint by about a quarter, and by 5 to 10% realistically. Yet the Paris Agreement requires an 80% reduction. At the global level, the FAO estimated in 2023 that livestock systems account for about 12% of human-caused greenhouse gas emissions, and the International Energy Agency in 2025 that data centres, all activities combined, emit about 0.5% of the world’s CO2.
I did the calculation with public figures. According to the French Ministry of Ecological Transition, the average carbon footprint of a person in France was 9.4 tonnes of CO2 equivalent in 2023, so Carbone 4’s realistic actions save between half a tonne and a tonne a year. BP, for its part, publishes its own emissions for 2023: 32 million tonnes of CO2 for its own facilities, and 315 million for the oil and gas it sold, once burned. To offset those 315 million tonnes, roughly three hundred to six hundred million people would have to adopt these actions, five to nine times the population of France. Individual actions are not completely useless, but they are not on the same scale.
On 4 August 2020, BP pledged to cut its oil and gas production by 40% by 2030 compared with 2019, and to raise its low-carbon investment to about 5 billion dollars a year. In 2023, the reduction target was lowered to 25%. On 26 February 2025, BP dropped it: its oil and gas investment rose to 10 billion dollars a year, transition investment fell to 1.5 or 2 billion, and its production is set to increase by 2030.
A carbon assessment for a filmmakers’ residency
From an ethnological point of view, there is an imaginary of ecology. I use the word in Cornelius Castoriadis’s sense in The Imaginary Institution of Society (1975): we can only represent things to ourselves through an imaginary, inner representations that allow us to represent them, and this applies to everything we think, not only to our mistakes. Saying that there is an imaginary of ecology does not mean that ecological problems are imaginary: the impact of data centres is real, and it is legitimate to be concerned about it. The imaginary of ecology is the representation of what a good action would be, the one that gives us the impression of being committed. Yet what we believe to be good at one moment can turn out to be more complicated, or not good at all. For a long time, nothing was said about the dangers of cigarettes, and milk or red meat were associated with good health; it was an imaginary of good health. There have always been people saying that these things did not bring health, and then the imaginary changed. In Media literacy begins with our imaginaries, where I also draw on Castoriadis, I traced our guilt in front of a prompt to the images that accompanied the arrival of these tools, photos of data centres in particular, whereas no image of that kind surrounds a rib steak.
I had an illustration of this during a working meeting with the design team of the residency I will lead in February 2027 for NEF Animation, a month in Nantes with ten animation filmmakers, on generative artificial intelligence and the ecology of narratives. The ecological question came up straight away, almost dogmatically: carbon footprint, energy responsibility, digital frugality. Since 1 January 2024, the CNC has made its production funding conditional on a double carbon assessment, provisional then final, drawn up with an approved tool such as Carbon’Clap from the association Ecoprod. Producers find themselves facing forms whose boxes they do not know how to fill in, so complex are the technical parameters and so uncertain the answers.
The incoherence is complete. The same people who earnestly debate digital sobriety during the day go to a restaurant in the evening and order meat. What they save by sending a few fewer queries is multiplied by thousands in the other direction by a month of meat-based meals. I am a vegetarian, and I presented this calculation to the design team of the residency as a biased calculation. It is not even the story of the hummingbird. It is a mechanism by which we make ourselves believe we are committed when we are not, and I described in The paradox of digital criticism how this kind of contradiction poses less a moral problem than a problem of understanding the mechanisms at work.
This is not about denying the ecological impact of artificial intelligence. In 2025 I wrote a detailed article on the energy consumption of data centres, Ecology and Artificial Intelligence, and another in 2026, Data Centers : We Fill Them Ourselves, on the sovereignty issues they raise. According to the International Energy Agency, the electricity consumption of data centres accounted for 1.5% of the world’s electricity in 2024; it is expected to more than double by 2030 and reach 945 terawatt-hours, slightly more than Japan consumes today, mainly because of artificial intelligence.
The disappearance of critical thinking in the service of capitalism
I described the same mechanism for the Covid period in What fear did to us and in Placing ourselves after the fact, or during: in the name of “protecting others”, no one exercised critical thinking any longer about the measures taken or about who really benefited from them. Oxfam calculated in January 2022 that the wealth of the world’s ten richest men had grown from 700 to 1,500 billion dollars since the start of the Covid-19 crisis, while the incomes of 99% of humanity fell.
On the ecological question, the mechanism is comparable. Bureaucratic rules are built, carbon calculation forms, regulatory constraints, decrees, which are quite restrictive for small cultural players but have strictly no measurable effect on the overall energy economy of artificial intelligence. Artificial intelligence is built in everywhere, whether we like it or not. Google’s search engine has used it to rank its results since 2015, and according to the survey the consultancy McKinsey published in 2025, 88% of the roughly two thousand organisations surveyed use it in at least one of their functions; I gave other examples in Invisible artificial intelligence. These uses, over which we have no individual control, consume incomparably more energy than the conversational queries of a few people with an ecological conscience, and I showed in Data Centers : We Fill Them Ourselves that public authorities themselves are customers, with our emails, our hospitals and our town halls.
Who benefits from this state of affairs? I ask the question without claiming to have a definitive answer. What is certain, on the other hand, is that the artificial intelligence industry does not suffer in the slightest from this ecological good conscience and pursues its development unhindered, that the environment gains nothing from it, since its effect is negligible, and that those who practise it perhaps gain above all the feeling of being on the right side.
Olivier Hamant’s robustness against the ecology of performance
The biologist Olivier Hamant distinguishes performance, where one optimises, anticipates and maximises results, from robustness, where one maintains the functions of a system despite fluctuations, at the cost of apparent inefficiencies (La troisième voie du vivant, Odile Jacob, 2022; Antidote au culte de la performance, Gallimard, “Tracts” collection, 2023). I presented this framework in Robustness, antifragility, resilience.
When I apply it to the ecological question, I see what does not work in current discourse. The dominant ecological imaginary is an imaginary of performance, where one would like to perform well in one’s eco-gestures, where one optimises one’s consumption and measures one’s footprint. But that is not how we get better ecosystems; what we get is good conscience, bureaucracy and, paradoxically, blindness to the real places where things are decided.
Drawing on this ethnological view of these discourses, I suggest entering into complexity rather than staying with simplism. As soon as there is simplism, “prompts pollute, so let us be frugal”, it is a dictatorship of thought. The reality is that the ecological impact of artificial intelligence depends on highly technical parameters (GPU or TPU, electricity mix, location of data centres, models used, generative or agentic AI), that our individual uses weigh almost nothing next to the uses built into every digital system, and that if one cares about ecology, it is not at all in front of one’s computer that it is played out first.
It is played out in everyday decisions in other parts of our lives: what we eat, how we travel, what we buy, the political choices we support. And for digital technology itself, it is played out collectively, in the tools an organisation chooses and in the place where it hosts its data, far more than in the number of queries each of us sends. Entering into complexity is essential to critical thinking. It means accepting that we do not control everything, that there are redundancies and uncertainties, that we can be wrong, and that it is in this play with the unforeseen, as Hamant says, that we build real robustness rather than displayed performance.
The point, then, is not to make people who use artificial intelligence feel guilty, nor indeed to relieve them of guilt. The point is to have a more complete, more accurate awareness of ecological reality. That means looking at things from a little higher up, seeking out information other than what we are given, and above all not letting ourselves be blinded by guilt over our small individual actions, to the point of no longer seeing the structural and political responsibilities, where something could change.
