A teacher named Solveig is walking around a Norwegian classroom while her carpentry students build a presentation. She stops at a table where two boys are using ChatGPT, watches for a moment, and lands on a verdict that ends the conversation: “It becomes like a source, just like reading the textbook.”
That single line is the hinge of a new paper by Leander et al. They sat with video of real classrooms and asked: when students and teachers use ChatGPT, what do they actually decide it is? Their answer points straight at what critical AI literacy needs to become.
Why “Just a Source” Is a Comforting Lie
The “source” label is reassuring because it makes the tool familiar. A source is something you already know how to handle. You read it, you check it, you cite it, you write the assignment in your own words. Drop ChatGPT into that slot and the whole anxious question of AI in school seems to resolve itself.
The authors show why that move is doing more work than it looks. Media researchers call it domestication, the way a strange new technology gets tamed into an existing routine. And in these classrooms, the taming is really about permission. Underneath every exchange about what ChatGPT is, Leander and colleagues find a quieter question: is this allowed? One student says it out loud, insisting his group isn’t allowed to use ChatGPT any way other than as a source they rewrite. The “it’s just a textbook” framing isn’t a neutral description. It’s the version of AI that school already knows how to permit.
I’ve felt the pull of this myself. It’s tempting to tell teachers that AI is just another reference tool and move on. This paper is a good reminder that the tidy answer smooths over the thing we actually need students to understand.
The Students Already Sense It’s Something Else
The most striking part is what the students do next. They keep breaking the frame their teachers hand them. In the talk, ChatGPT shows up as a tutor they learn from, an editor, an idea generator, a machine that summarizes and simplifies, even a kind of external memory. One boy pushes his group directly: we’re not copying from ChatGPT, we LEARN from it.
The second classroom stages the tension even more sharply. One student argues the group could never have found this information without the tool, treating it as a knowledgeable partner. His classmates hold the official line, and one of them makes an interesting observation: ChatGPT can’t be trusted as a source because it can’t name where its information came from, and it hands back a different answer every time you ask. The text has no stable identity anyone else can check.
Notice what’s happening. The kids are constructing AI as an agent, something that acts, while the school keeps pulling it back to being a resource, something you consult. Both are true at once, and the friction between them is exactly where the learning lives.
What a Textbook Never Did
The authors give the genuinely new part a name : ChatGPT is an abstraction machine. A textbook was written by people, for a context, with a fixed and checkable identity. A language model pulls language away from any specific context and hands back a smooth, finished-looking whole assembled from patterns across millions of texts. In the carpentry lesson, that showed up in a small, telling way. The teacher noticed the AI’s answer was “a bit vague,” because a model, by design, strips out the specific, local detail her students needed about installing a real door.
That difference changes how we should teach writing. The multiliteracies tradition, the New London Group model many of us were trained on, puts the human at the center of design, choosing from available materials to compose and remix texts. Leander and colleagues argue GenAI scrambles that picture.
When a student writes a prompt, they aren’t selecting from a shelf of resources. They’re writing a text whose only job is to make another text, addressed not to a reader but to a model. The finished output doesn’t arrive as raw material. It arrives as what the authors call AI saturated design, with most of the decisions already baked in.
Leander and colleagues offer one concrete response they call Undesign. You don’t let students paste an AI answer in whole. You have them break it apart, pull out the useful pieces, and recombine them with their own examples and other sources. It’s a small move with a big effect: it turns a finished-looking block of text back into raw material a student has to think with, which is exactly what the carpentry teacher was reaching for when she pushed the boys to add examples from their own door assignment.
Critical AI Literacy Means Seeing the Machine and the Platform
This is where the paper hands teachers something usable. If the old critical literacy taught students a “metalanguage” for reading texts, the authors argue AI demands two new capacities on top of it.
The first they call modeling language: a working understanding of how these systems actually operate, where the training data comes from, what biases ride along with it, why the output is confident and generic at once, and why words like “thinking” and “intelligence” mislead us when we attach them to a machine. This is the same competence I’ve been building toward in my own critical AI literacy work, and it’s the same skill instructional designers named when they called prompt writing a craft you have to learn.
The second capacity is meta-platform consciousness, and it’s the part schools almost never reach. To read AI critically, the authors argue, means seeing the platform behind the chat box: whose company owns it, what data it harvests, what it does with student information, and the physical cost most classroom talk ignores, the data centers and the energy and water these systems consume. That’s a demanding curriculum. It’s also the accurate one. A single good question can start it: ask students who profits when they type into the box, and where their words go afterward.
There’s a policy thread here too. Because so much of the classroom talk was really about permission, the paper is an argument for schools to decide, out loud, what AI is authorized to be, which is the same case I’ve made for clear course-level AI use rules.
Language Equal to the Thing
The deepest point in this paper is quiet but large. When we call ChatGPT “just a source,” we hand students a small vocabulary for a strange new object, and a small vocabulary keeps them from seeing what’s really in front of them. The New London Group wanted schools to raise “designers of social futures.” You can’t design a future you don’t have words for.
So the work isn’t to make AI feel familiar. It’s to give students language big enough to match the thing, words for the agent, the abstraction machine, and the platform all at once. That’s harder than calling it a textbook. It’s also the only version of AI literacy that respects how strange and consequential these tools actually are.
References
Leander, K. M., Arnseth, H. C., Silseth, K., & Erstad, O. (2026). “It becomes like a source, just like reading the textbook”: Towards an expanded GenAI-informed multiliteracies model. Pedagogies: An International Journal, 21(2), 234-252. https://doi.org/10.1080/1554480X.2025.2545203
