Pedagogy of Artificial Intelligence

The educational tool “Dans le regard de l’IA” (“Through the Eyes of AI”) by Média Animation sets out to question the biases of generative artificial intelligences. Its theoretical content is solid. But its pedagogy, very school-like, seems to me to go against what should be done: starting from people’s existing knowledge, not from their supposed ignorance.

A generous tool, a considerable effort

I must say it first: the tool “Dans le regard de l’IA”, developed by Média Animation with the support of the CSEM and the Wallonia-Brussels Federation, is the result of serious and generous work. The accompanying theoretical booklet is rich, featuring interviews with Gil Bartholeyns and Philippe Descola, among others, a global history of images, and a sharp analysis of dataset biases. All of it shared freely. That’s remarkable.

The tool itself, with its image sets and its “Word-Image” activity, is quite ingenious in its design. It seeks to make tangible, through experience, the data labelling mechanism that shapes image production by generative AIs. The principle is interesting: having participants live out an analogy with how AIs work, through drawing, visual recognition, and sorting images by keywords. The ambition of revealing the cultural biases of Western representations is one I fully share.

And yet, when I read the facilitation instructions, when I put myself in the participant’s shoes, something doesn’t work, for me, and in connection with my experience as a pedagogue.

The feeling of incompetence as a methodological symptom

I read the tool’s objectives. One must “understand how generative image AIs work”, “grasp the concepts of prompt and predictive generation”, “understand the role of datasets”, “identify biases”. I read the facilitation approach: the facilitator “presents the objectives”, “offers a definition”. Then the participants do the activity, and the facilitator “debriefs” by explaining what the activity was illustrating.

This is the classic structure of explanatory pedagogy. You tell people what they’re going to learn. You have them do an exercise. You explain what they should have understood.

And right there, even as an adult, even as someone who’s been working with AI for years, I feel the anxiety of the bad student rising. The one school taught me so well: the certainty that I don’t understand the instructions, that everyone else got it and I didn’t, that I’m failing to meet what’s expected of me. This is a methodological effect. It’s not an accident. The device itself produces it.

When pedagogy operates through successive instructions to be correctly carried out, when there are “right answers” to find (which images go under which keyword, which cards are “correctly” labelled), when knowledge sits on the facilitator’s side—the one who “debriefs” and explains what the exercise “illustrates”—the relationship being built is one of looking down. The facilitator knows. The participants discover what they should have known.

This can work for some. The tool is well made, and the quality of the human relationship with the facilitator can change everything, save everything. But the structure of the device pushes towards normativity—that is, towards excluding those who don’t share the same way of thinking, the same form of intelligence.

People who are already competent, but don’t know it

The root problem is that this type of pedagogy starts from the assumption that participants don’t know, and that the device will teach them. On questions of artificial intelligence in particular, but on many things, this is often wrong, and above all it closes doors to intellectual autonomy instead of opening them, because of the roles assigned by the device.

Young people especially have daily AI practices that constitute considerable experiential knowledge. They use generative filters on their photos, chat with bots, produce images with generators built into their social networks, test the limits of systems, subvert prompts. This knowledge is concrete, embodied, often very sharp. But it isn’t formulated in the language of media education. It doesn’t fit into the device’s boxes. It doesn’t match the “objectives” predefined by the tool.

And this is where something serious is at play. The participants’ areas of knowledge, which are not the same as those of the tool’s designers, find themselves delegitimised—not through ill will, but through structural omission. They simply don’t exist in what is being offered. A person may then come to believe that their knowledge has no value. And this isn’t just a belief: within the framework of the device, their knowledge is objectively discriminated against, since it doesn’t match the form of knowledge the framework recognises as legitimate. And also, very often, this means that children and young people experience the knowledge being taught to them superficially, because it is only partial while passing itself off as total—so it isn’t really knowledge at all, but a code detached from reality, which is ultimately just a space of good conscience for the adults who built it, and absolutely not a space of genuine openness.

This is a question of cultural rights. Every person carries a culture, forms of knowledge, experiences that constitute their dignity. An educational device that makes no room for this knowledge, that doesn’t allow it to be fully expressed, recognised, or brought into dialogue with other forms of understanding—that device effectively undermines the cultural existence of the very people it claims to educate. It is therefore entirely normal for such a device to be rejected, except by those who are very school-minded and already trained in obedience within their family environment.

The designers’ knowledge is also partial

Because the tool’s designers don’t hold ubiquitous knowledge either. Their understanding is partial. Their theoretical booklet, however rich, is one perspective among others. They don’t, for example, compare the different generative AIs available, their respective philosophies, their distinct technical and ethical choices. There is no mention of Claude, which has been available for a long time and is an AI singular enough in its approach to deserve being distinguished. Euria, the sovereign European AI, doesn’t appear either (perhaps it didn’t yet exist at the time of writing). This absence is understandable—it speaks to the difficulty of staying up to date in a fast-moving field—but it illustrates precisely what I’m saying: all knowledge is partial, including that of the people who design educational tools. This partiality of knowledge should be mentioned, highlighted, but it isn’t—the knowledge is presented as universal, which it absolutely is not. As a result, the tool is intrinsically, through its pedagogical structure, out of step with reality and with people’s experience (even though its stated aim is precisely the opposite).

So why build a device that presupposes knowledge on one side (the tool’s, the facilitator’s) and ignorance on the other (the participants’)? This asymmetry is problematic, not because the designers are wrong in what they propose, but because this structure prevents the collective expansion of knowledge. We limit ourselves to transmitting what we already know, instead of creating the conditions for new knowledge to emerge from the encounter.

What can be drawn from it

I don’t want to throw the baby out with the bathwater. There are very interesting elements in this tool that can be appropriated differently.

The idea of working with image sets, for example, is fertile. But rather than distributing pre-made image sets and asking participants to sort them into predefined categories, we could ask people to make their own. In fact, the tool does suggest this, at the end of the process, in a section called “Appropriation”. But it’s relegated to the end, as an optional bonus. I think that’s where we should start.

The theoretical booklet on AI biases, the history of images, the question of dataset labelling—all of this is a valuable resource for the facilitator. Not as content to transmit, but as a personal grounding that enriches their own understanding and allows them to accompany the group’s discoveries with relevance. Knowing that it lacks a great deal of information and concepts.

Proposals for a creative pedagogy of AI

What I propose is to reverse the logic. Start from creativity, not from explanation. Start from people’s knowledge, not from their supposed ignorance. Here are some concrete avenues, some tested, some envisaged.

Create first, reflect afterwards. Rather than explaining how AIs work and then doing an exercise that “illustrates” the explanation, we invite participants to create something with AI. A portrait, a landscape, a scene, a story, a character. During the creation, participants discover for themselves the resistances, biases, and surprises of the tools. It’s an embodied discovery, not a lesson received. Reflection comes afterwards, from lived experience.

Compare AIs with each other. Submit the same prompt to three or four different AIs (ChatGPT, Midjourney, Claude, Euria, generators built into social networks) and compare the results. The differences are immediately obvious. Participants see for themselves that AIs are not a homogeneous block, that each tool carries its own biases, its own aesthetics, its own refusals. This concrete comparison produces more critical thinking than an hour of explanation about datasets.

Start from real uses. Ask participants what they already do with AI. Which tools they use, in what contexts, with what results, what surprises, what disappointments. Start there. From what they already know, even if they don’t yet know they know it. The facilitation then consists of putting words to experiences, making connections between individual practices and collective stakes, making visible competences that were invisible.

Create knowledge-exchange situations. In any group, there are always people who know things others don’t, and vice versa. Organising the conditions for this exchange is pedagogy. You might, for example, suggest that each person shows the others something they know how to do with AI—a trick, a use, a subversion. The facilitator is no longer the sole source of knowledge, but the one who enables knowledge to circulate within the group.

Use AI advertisements as educational material. The ads for Claude aired during the 2026 Super Bowl, for example, are very rich objects to analyse. They show how a technology company represents itself, what promises it makes, what imaginaries it mobilises. Working on these ads means doing media education and critical thinking about AI at the same time, with concrete, current material that speaks to people, while also putting into perspective the fact that they are advertisements.

Subvert AIs. Propose creative challenges: get an AI to produce an image it isn’t supposed to be able to make. Test its limits, its refusals, its possible workarounds. This kind of playful activity develops a fine understanding of how the tools work, through practice and experimentation, not through the transmission of notions.

Pedagogy as recognition

What is ultimately at stake is the question of recognition. Recognising that the people we work with already carry knowledge and competences. Recognising that our own knowledge, however elaborate, is only partial. Recognising that intelligence is collective, that the most fertile pedagogical process is one where everyone learns, including the facilitator.

Jacques Rancière, in The Ignorant Schoolmaster, laid the theoretical framework for this approach: truly emancipatory teaching starts from the postulate of the equality of intelligences, not from ignorance to be filled. Paulo Freire, in Pedagogy of the Oppressed, showed that educational devices that fail to recognise people’s knowledge reproduce the very relations of domination they claim to fight.

When it comes to artificial intelligence, these stakes are all the more acute because knowledge is being constructed in real time, practices are constantly evolving, and young people in particular are developing competences that many adults simply don’t have—and that this type of device tends to render invisible and to delegitimise. It seems to me that the role of pedagogy is not to explain AI to people, but to create the conditions for people to think AI together, starting from what they live, with what they already know.

This is more demanding for the facilitator. It requires improvisation, listening, openness to the unexpected. It’s what I call structured improvisation: arriving with a framework, tools, resources, but staying open to what emerges from the group. The result is always richer, more surprising, more formative—for everyone—than the application of a pre-established device, however well designed it may be.

“Dans le regard de l’IA” is a tool by Média Animation, available at media-animation.be. The theoretical booklet and the facilitation guide can be downloaded for free.

Thanks to Noémie Rubat du Mérac for the tip about this project.