The Teacher as Text Designer, the Student as LLM: How AI Disrupts Traditional Writing Pedagogy
Words never truly belong to us. As Bakhtin taught, they're borrowed from our collective word bank, coming alive in our mouths without charge, joining chains of utterances stretching back through history and forward into possibility. Consider this phrase: "Mitochondria are the powerhouses of cells." Have you encountered it before? Who does it belong to? Once meaningful, now somewhat hollowed out of the specifics I once knew. Should I now footnote it? Who owns it? If AI produced these words, would they be stolen? This question of ownership now haunts writing classrooms where Large Language Models (LLMs) have disrupted the landscape of text production, forcing us to reconsider longstanding theories of how writing works and how we should teach it.
The Fingerprint vs. The Dataprint
Luciano Floridi's concept of 'dataprint' (2025) represents more than a minor semantic shift from 'stylistic signature' when discussing LLM output. This distinction marks a crucial boundary between natural and computational text—one we must acknowledge to understand the relationship between human and artificial writing processes.
Human authors develop "voice" through intentional choices, personal experiences, and both conscious stylistic decisions as well as unconscious habits. Jane Austen, William Faulkner, John Updike—each crystallized a signature style visible throughout their writing lives. Voice and style are deeply personal, varying by genre and purpose, but unmistakably human.
LLM output operates differently. Their dataprints emerge from training data and architectural design rather than lived experience or intentionality. Researchers can identify these dataprints by analyzing patterns in language use—the frequency of parts of speech, common word combinations, and character patterns. When plotted on radar charts, each model reveals a distinctive shape. This statistical approach achieves remarkable accuracy (94-98%) in identifying which LLM produced a given text.
Recent studies by Bitton, Bitton, and Nisan (2025) reveal that different LLM families display distinct linguistic patterns—GPT-4 uses 18% more determiners than Claude and Llama 2, while exhibiting 22% higher adverb frequency compared to human authors. These patterns remain consistent across domains; a GPT model "sounds like" itself whether discussing physics or cooking a recipe. When classifiers trained on one model's text are tested on another model from the same family, performance drops only slightly (0.01), but when tested on different model families, performance plummets (0.62 drop).
Knowledge-Telling: Students as Human LLMs
This revolution in text production forces us to revisit Bereiter and Scardamalia's (1982) influential model of writing development, which distinguished between novice "knowledge-telling" and expert "knowledge-transforming" processes. In the knowledge-telling approach, novice writers operate under the assumption that knowledge exists pre-assembled in their minds. Their task is merely to retrieve and transcribe this information in acceptable form according to the teacher's specifications. The writer becomes a passive conduit, transferring existing knowledge onto the page according to externally imposed requirements.
Here's the revelation that has gone unnoticed until now: in traditional educational settings, teachers have functioned as "text designers"—establishing parameters, setting goals, and defining evaluation criteria. Students, meanwhile, have been reduced to text producers who generate content to satisfy these external demands without personal investment in the communicative purpose.
Sound familiar? In traditional education's knowledge-telling paradigm, students have essentially functioned as human LLMs—producing text on demand according to specifications set by an authority, lacking agency over the text's purpose or design. The teacher acts as a "distant author," providing prompts to novices who then produce texts that satisfy the teacher-as-writing-manager.
The following visual also from Bereiter and Scardamalia reveals previously non-remarkable assumptions. What are students thinking about when they begin a writing task? We would hope they are thinking about what they might write, what they might say about the topic or idea. what they need to learn or find out, etc. Instead, the opening phase requires them to create a “mental representation of the assignment.” How many of us writing on Substack begin our writing process by constructing a mental model of the assignment? To move into the details of planning, students analyze the “problem,” which is the problem posed by the assignment. Then they create sub-problem spaces: content is a problem (What can/should I say?) and c constructing discourse (text) is also a problem (How do I have to say it?). Their own thinking as a writing agent is completely submerged in the waters of the assignment. Indeed, any teacher giving a highly detailed writing assignment with scoring criteria and the rest of the traditional paraphernalia might try feeding the material to an LLM to experience what students experience.
Distant Writing: The Great Inversion
Floridi's concept of "distant writing" flips this dynamic entirely. In this new paradigm, humans (students) become text designers, establishing requirements, constraints, and goals, while the LLM takes on the execution role. The human reclaims agency over the conceptual and creative dimensions of writing—deciding what to communicate and why—while delegating the mechanical production of sentences to the machine.
This inversion challenges us to reconsider what constitutes authentic writing. If the problem with knowledge-telling was its reduction of students to text-producing mechanisms serving teacher-designed purposes, then distant writing potentially liberates human writers to focus exclusively on intentionality, purpose, and meaning—the very elements missing from traditional knowledge-telling approaches.
The most profound implication is that distant writing might actually promote what Scardamalia called "knowledge-transforming"—the more sophisticated approach where writers actively reconstruct and transform, not reproduce, knowledge through writing. By assuming the role of designer rather than say-write scribe, the human writer must engage more deeply with goals, audience needs, and the transformative potential of the text. The mechanical aspects that often preoccupy novice writers (sentence construction, word choice), which in the mind of a mature writer are of the utmost significance, are delegated to the LLM, potentially allowing for greater focus on the conceptual work of shaping knowledge. The writer-in-charge then has open season on the AI text, reworking the simulated language painstakingly to accomplish the original vision, taking full advantage of Bakhtin’s insights about multivocality.
Paradoxically, the advent of LLMs might push human writers toward more sophisticated knowledge-transforming processes by removing the cognitive burden of knowledge-telling mechanics. In the same way, LLMs seen as tools put to work by a meta-author suggests that breaking the task into manageable subtasks will become a routine part of getting ready to write. Knowledge-tellers typically start writing with word one and continue until they have finished. Rarely do they do deep revision of the sort mature distant writers perform on hybrid texts.
New Questions for Writing Classrooms
This shift introduces urgent questions for writing instructors:
How do we help students understand the distinction between human voice and LLM dataprint?
What constitutes authentic writing when the mechanical production of text is delegated to machines?
Should we embrace distant writing as a way to accelerate student development toward knowledge transformation?
How do we assess writing when the boundary between human and artificial contribution blurs?
A New Writing TextScape
The concepts of 'dataprint' and 'distant writing' don't invalidate Bereiter and Scardamalia's developmental framework but expand it to address new realities of AI-assisted writing. As writing increasingly involves collaboration between human designers and LLM systems, understanding the dataprints of different AI models becomes crucial for effective composition. Writers as we speak are choosing among models of LLMs to complete particular tasks, even using two different models to parse and express ideas in text of different dataprints for options. Writers must now consider not only their own cognitive processes but also the characteristics of their AI collaborators. This shift introduces new questions about authorship, creativity, and writing development that weren't contemplated in traditional models.
As educators, we face the challenge of helping students navigate this complex terrain, guiding them to become thoughtful text designers rather than passive think-say text producers. Perhaps in this new infosphere, the most essential writing skill isn't crafting perfect sentences; understanding when, why, and how to shape knowledge through an intentional collaboration with artificial systems may be the coming standard.
The words may never truly belong to us, but the purpose and meaning we create with them still can—and must.