On compar:IA, the comparison tool opened by the French Ministry of Culture in October 2024, you can now question thirty-eight artificial intelligence models, and yet we say “AI”, in the singular, as if there were only one. Drawing on the Egyptologist Jan Assmann, on Max Weber, Bruno Latour and Jacques Rancière, I propose what I call methodical polytheism: asking several models the same question and comparing their answers, the way we compare the opinions in a group, mine included.
“I asked the AI”
“I asked the AI”, people say, the way they say “I looked it up in the dictionary”, with a definite article that assumes there is only one. In The biases of trust in machines, I called “algorithmic authority bias” the trust we then place in the machine. I explained there that I follow the GPS even when it sends me down an absurd route, in the neighbourhood where I have lived for twenty years. With “AI”, it seems to me that the effect is even stronger, since it answers us in sentences.
Behind the word, though, there are ChatGPT, Gemini, Claude, Mistral, DeepSeek, Grok, etc., and each one comes out of a company that has chosen its training data and had its employees write the model’s rules of behaviour. Some of these rules are published. OpenAI published a “Model Spec” for ChatGPT on 8 May 2024, which it has revised regularly since. Anthropic published a new “constitution” for Claude on 22 January 2026, freely available. In the 2024 version of OpenAI’s text, we read for example that the assistant should not “try to change anyone’s mind”, and that it should ask clarifying questions when a request is unclear. This is an editorial line, like a newspaper’s, and Anthropic’s is not the same.
I saw this while preparing the article « The flavor of data ». All I wanted was for Gemini to turn the photo of a page into text, a page in which the author denounced the grip of the GAFAM on the functions of the State. Gemini gave me the opposite, a passage praising the understanding between these companies and States. I insisted several times: it apologised, then did it again. I only noticed because I had the page in front of me. When you ask a question whose answer you do not know, you do not have that page, and only a second answer, from another model, can take its place.
The gods called to witness
Historians of religion thought about the difference between one god and many long before us. In 2003, the German Egyptologist Jan Assmann published a book translated into French in 2007 by Laure Bernardi (Aubier) as Le prix du monothéisme, and into English as The Price of Monotheism. In it he calls “the Mosaic distinction” the divide between true religion and false ones, and he places a first form of it under Akhenaten, in the fourteenth century BCE. Before this divide, the gods of one people were matched with those of another. The sun god was called Utu in Sumerian and Shamash in Akkadian, and lists of gods set the two names side by side. When Ramesses II and the Hittite king Hattusili III concluded a peace treaty, around 1259 BCE, they called to witness “a thousand gods” of the Hittite land and “a thousand gods” of Egypt. Assmann speaks of the “translatability” of gods. The god of Israel, for his part, has no equivalent among the neighbouring peoples: with monotheism, there is nothing left to translate.
When we say “AI”, we are doing a little what Akhenaten did: there is only one answer, the one we received. At TUMO Paris, Poe made it possible to send the same request to several models, so the teenagers discovered that there were several manufacturers, each with its own choices; I described this in Media literacy begins with our imaginaries. They were then doing what the polytheistic peoples of Antiquity did: looking for what, from one answer to the next, corresponded, what diverged, and why.
The word “polytheism” also has a modern meaning, given to it by Max Weber. On 7 November 1917, in Munich, he gave a lecture to students, published in 1919 and known in English as “Science as a Vocation” (in French, “Le métier et la vocation de savant”, in Le savant et le politique, Plon, 1959, translated by Julien Freund). For him, the modern world is a polytheism of values. He gives an example: science cannot say whether French culture is worth more than German culture, because here too different gods are fighting one another. He draws a rule from this for his own profession: the professor does not have to tell his students which god to serve, for, in the words of the French translation, “the prophet and the demagogue have no place in a university chair”. Weber is speaking of ultimate values, not commercial products, and I do not want to force the comparison. His rule still applies to the person leading a workshop on AI; I will come back to this.
Gods we make
In 1996, Bruno Latour published a short book with Les Empêcheurs de penser en rond, Petite réflexion sur le culte moderne des dieux faitiches (in English, On the Modern Cult of the Factish Gods). In it he recalls that the Moderns reproach “fetishists” for worshipping objects they have made with their own hands, and believe themselves safe from this naivety. To get out of this, he coins the word faitiche (“factish” in English), from “fait” (fact) and “fétiche” (fetish), which both go back to the Latin facere, “to make”: a factish is something we make and which nevertheless makes us act. A language model is one. I know that engineers built Gemini, and yet it confidently handed me the opposite of the page I had entrusted to it.
The teenagers at TUMO, with Poe, sent the same request to several models and looked at what carried over from one answer to the next. This is what I propose to call, for want of a better term, methodical polytheism. The word “methodical” comes from Descartes’s doubt (Discourse on the Method, 1637), which he practised as a working method. After comparing, you may very well prefer one model, or decide to use none at all!
A hospital in Lyon, and rain
I have had very different audiences make this comparison. In 2024, at TUMO Paris, we had the teenagers install the Poe app as soon as they arrived; it brought together models from several manufacturers in a single interface, for text as well as images (Tumo x IA - Creative journeys to the heart of AI). There were eight sessions of around seventy people, in a cinema at the Forum des images. Each pair had fifteen minutes to write a script in dialogue with language models, fifteen to draw three or four images from it, then twenty to edit the film and add sound. The films then appeared one by one on the big screen, and the discussion that followed was often heated.
On a training course I led for librarians, one participant was surprised to see a model add to his image a hospital in Lyon and some rain, which nothing in his request called for. Another participant found that the model she was using went well beyond the instruction she had given it.
By comparing, we also see the similarities, and I find as much to learn there as in the differences. Almost all models reproduce the same gender stereotypes and, by default, set their images in a Western, well-off world; many also push stories towards what pleases. These similarities come from the data they were trained on, largely what we ourselves have put online over the past thirty years, and from a commercial logic shared by their manufacturers. In spring 2026, I asked six models (Gemini, Perplexity, ChatGPT, Claude, DeepSeek and Grok) a question about the origin of an open letter, deliberately slanted. Most of them went along with the question, whereas a few weeks earlier, asked the other way round, models from the same manufacturers had given me equally solid arguments for the opposite view. I described this in The Tilting Mirror.
The bilingual Telemachus
In Le Maître ignorant (Fayard, 1987; in English, The Ignorant Schoolmaster), Jacques Rancière tells the story of Joseph Jacotot, a French teacher exiled under the Restoration who became, in 1818, a lecturer in French literature at the University of Louvain. Many of his students did not speak French, and he did not speak their language. Through an interpreter, he had them study a bilingual edition of Fénelon’s Telemachus, and they learned French by comparing the two texts. From this Rancière draws the idea of the equality of intelligences, and a critique of the “explicator”, who persuades the pupil that they cannot understand without him. Jacotot’s students learned by putting two versions side by side, without their master explaining anything to them. I took up this figure of the ignorant schoolmaster in Succeeding without convincing, to evaluate a collective session differently.
And what about the person leading the group in all this? Polytheism applies to them too. If they explain to the group what “AI” is and what to think of it, they take the place of the one god, and the group receives their word the way one receives an oracle’s. In Pedagogy of Artificial Intelligence, I argued for having people create rather than explaining to them. When a group compares models, it also compares points of view, mine included.
The word “polythéisme” came to me during the librarians’ course I mentioned above. As I scrolled through the list of available models in front of the group, I said that we were not dealing with a magic god, but with lots of different manufacturers, and the word followed. I only grasped its significance afterwards. The group, for its part, taught me something else. One trainee pointed out that, in another training course she had attended, most people had never used generative AI, whereas I was starting from the opposite assumption. I have been using these tools every day for several years. In that course, I was looking at them from the angle of cinema and cultural mediation. It would be easy to conclude that I know more than the people I train. About the way these tools are used, or not used, in circles other than mine, they often know more, and it is by exchanging in this way, then continuing the exchange without me, that each person builds their critical thinking.
The leisure to compare
Methodical polytheism is easier to describe than to practise. In daily life, we use artificial intelligence tools to get a result: an email to send, a summary to hand in, an image for a poster. When the answer is good enough, we move on, and we do not bother asking a second model the same question, then a third. And even when the question is worth it, comparing costs credits, sometimes several subscriptions, etc. In the context of ordinary work, my proposal therefore remains somewhat theoretical.
I propose giving it moments of its own: times for training, meeting or collective intelligence, without constraint or pressure to be productive, where people can play together with these tools and step back from them. The Greeks called skholè the time freed from business and urgency, and the word gave us “school”. Pierre Bourdieu, in Méditations pascaliennes (Seuil, 1997; in English, Pascalian Meditations), made it the condition of the kind of thinking he calls “scholastic”: we can only step back from the world if we have this time, and not everyone has it. If you organise training courses, meetings or collective work sessions, it is therefore up to you, in my view, to offer this time, rather than counting on each person to take it alone, in the evening, using their own credits. In Proper time, I proposed calling proper time the time whose use we decide for ourselves; during a comparison workshop, everyone gets a little of it back.
compar:IA and Arena
Poe displayed, before each request, its cost in points, and this cost varied a great deal from one model to another. Participants thus discovered that an image is not free, which gave them a concrete sense of what these tools cost to run, an order of magnitude I gave in Ecology and Artificial Intelligence. At the end of March 2026, Poe cut its free allowance from 3,000 to around 300 points a day, without announcing it, according to the sites that track its pricing. With 300 points, you can barely make a few images, and for a group the tool becomes useless.
I now turn to two free services. The first, compar:IA, is a public service launched in October 2024 by the Atelier numérique of the French Ministry of Culture. Two models drawn from the thirty-eight on the list answer the same question, without giving their names. You vote, and the names then appear, along with the energy consumed by each answer, in watt-hours. Each model also has its own page, with an “energy score” from A to F and an openness level out of 6: Apertus 70B, a Swiss open-source model, scores 6, Claude Sonnet 5.5 scores 1. The votes are then used to build a ranking of the models, published on the site. Only texts are compared there, not images.
The second, Arena, was born in 2023 at the University of California, Berkeley, under the name Chatbot Arena, and was called LMArena until 28 January 2026. There you can choose the two models yourself, and you can also compare images. It is free in October 2026, as Poe was before March.
Points not to miss
- Setting aside time. At TUMO, each pair had fifty minutes to make their film, and nothing else to hand in. If you organise a training course or a meeting, what moment in it do you leave free of any expected result?
- Sending the same request to several models. At TUMO, we had the teenagers install Poe as soon as they arrived, so that they could switch from one model to another. And you, how many models do you ask a question that matters?
- Voting before knowing the names. compar:IA hides the names until the vote, and I suggest letting the discussion start from the moment they appear.
- Looking at the differences, then the similarities. With the librarians, it was a participant who noticed the hospital in Lyon that nothing in his request called for. I then look with the group at what the answers have in common, starting with gender stereotypes.
- Turning the question around. In spring 2026, I questioned six models about an open letter from the SACD (the French society of authors and composers for stage and screen) with a slanted question, and most of them followed its slant (The Tilting Mirror). What do they answer if you ask the question the other way round?
- Looking at the cost. At TUMO, the teenagers saw the price of each request in Poe points; on compar:IA, it is in watt-hours.
- Letting the group name what it sees. “Polytheism” was my word, said to the librarians, but it was a trainee who taught me that, in another course she had attended, most people had never used generative AI. The group often knows more than I do about its own professions!
