Contemporary writing practices rarely look like a straight line. A draft opens in Google Docs. A prompt is tested in ChatGPT. A source is skimmed through a database. Feedback arrives through comments. A sentence is revised, then rephrased by an AI tool, then revised again. It is tempting to describe this as a loop—a recursive, iterative process in which writers move back and forth among tools and tasks.
This metaphor works, to a point. It captures circulation, return, and adjustment. But it also risks smoothing over something more complicated; loops suggest coherence. They imply that writing moves along a track, even if that track bends and repeats. What they struggle to account for are the uneven, layered, and sometimes contradictory forces that shape writing as it unfolds.
Put differently, a loop describes how writing moves, while an assemblage helps explain why writing unfolds differently across changing relations, tools, constraints, and contexts. This post offers a slight but important reframing: what we often call “in-the-loop writing” is better understood as a rhetorical assemblage.
From Loops to Assemblages
Recent work on Rhetorical Load Sharing (RLS) (Knowles, 2022; 2024) as a framework for how writing tasks are distributed across human and nonhuman participants—writers, platforms, algorithms, and collaborative infrastructures—in ethical and productive ways is a valuable contribution to Writing Studies that moves us away from seeing writing as the product of a single author and toward a more distributed model of composing. However, RLS often relies on the idea of a loop: a system in which rhetorical work circulates among human and non-human participants, but these “in-the-loop” frameworks are inaccurate due to the inherent multifariousness present in many writing situations.
Assemblage theory offers one way to extend this thinking in metaphorically accurate ways that acknowledge the complexity of contemporary rhetorical situations. Stemming from the philosophical concept of the rhizome, assemblage theory “was meant to apply to a wide variety of wholes constructed from heterogenous parts” (DeLanda, 2006, p. 3). So, rather than imagining writing as a loop, we might understand it as a dynamic configuration of heterogeneous elements—people, technologies, institutional norms, genres, and material conditions—that come together, temporarily, to produce a text. These elements are not bound together in a stable system. Instead, they remain partially independent, capable of being recombined, reweighted, or disrupted. From this perspective, writing does not move smoothly through a loop. It emerges from shifting relations.
Consider a familiar scene. A student—call her Rebecca—is drafting a literature review. She begins by outlining her argument in a word processor. Unsure how to synthesize several sources, she turns to a generative AI tool and asks for help summarizing key themes. The AI produces a paragraph that sounds polished, even authoritative, so Rebecca pastes it into her draft. But she hesitates; the tone feels too generic. She then revises, checks sources, and reintroduces specificity. Later, she asks the AI to rephrase a sentence for clarity. She keeps some phrasing, discards other suggestions, and reshapes the sentence.
At no point is the rhetorical work simply handed off. Instead, it is reconfigured across interactions. This matters because AI systems do not understand rhetorical situations in the way human writers do. They generate language by predicting likely word sequences based on patterns in training data, not by engaging purpose, audience, or exigence in a situated way (Knowles, 2022). As a result, their outputs often require human intervention to align with disciplinary expectations, local contexts, and ethical considerations. If we describe Rebecca’s process as a loop, we capture repetition but miss instability. The process is not circular so much as layered and contingent.
Rhetorical Load as Emergent
This is where assemblage theory helps to reframe RLS. Rather than imagining rhetorical work as something parceled out across participants, we might see it as something that emerges from their relations. RLS is often described in terms of how humans and machines “split the effort expended for a given task” (Knowles, 2022, p. 259), but this framing can imply that rhetorical labor exists prior to its distribution.
A rhetorical assemblage perspective suggests otherwise. The “load” itself is not pre-existing; it is produced through interaction. Rebecca does not simply decide what the AI will do. The AI’s outputs are shaped by training data, probabilistic modeling, and interface design. Prompts themselves function rhetorically and, as Ranade et al. (2024) pointed out, “set the context for the conversation”, shaping what counts as relevant output (p. 711). Institutional expectations and genre conventions further constrain what counts as acceptable writing. Rhetorical load, then, is not a fixed quantity. It is a relational effect.
One of the most useful aspects of this reframing is that it helps us attend to moments of friction.
In Rebecca’s case, friction appears when AI-generated text does not align with disciplinary expectations. The language is too general, the claims too broad. This mismatch forces revision. Loops tend to hide this friction by emphasizing flow, but assemblages foreground it. This is especially important pedagogically. Scholars have noted that in-the-loop approaches can sometimes reduce human involvement to a final quality check, “like a human inspector placing a sticker on a machine-made product” (Mason & Dvorak, 2026, p. 264). When that happens, students may engage less with the writing process itself, but moments of friction resist this reduction. They demand engagement. They are not breakdowns—they are sites of rhetorical activity.
Rethinking Agency
This shift also reframes how we understand agency. If we rely on loop models, we tend to ask: Who is doing the work? How much belongs to the human, and how much to the machine? These questions often lead to debates about authorship and control. Assemblage theory lends itself to a different approach. Agency is not something that can be cleanly divided because it is produced through interaction.
This aligns, in part, with RLS frameworks that position AI collaborative writing along a spectrum between human-authored and synthetic text, rather than as a binary (Knowles, 2024). But assemblage theory pushes further by emphasizing that agency is not just distributed; it is emergent. Rebecca’s capacity to act depends on her relation to tools, knowledge, and constraints. The AI does not possess human-like agency, but it participates in shaping what becomes possible within the writing situation.
Finally, thinking in terms of assemblages allows us to scale up. Rebecca’s writing process is not just an individual workflow. It is nested within larger systems: courses, writing programs, institutional policies, and platform infrastructures. As Mason & Dvorak (2026) noted, workflows are useful, but they can become too “rigid” if they fail to account for how writers adapt, deviate, and reconfigure processes in practice (p. 263). Assemblage thinking accounts for this flexibility.
Not a Closed Loop
What appears as an individual loop is always part of a broader network of relations. None of this is to say that the loop metaphor should be discarded. It remains useful for describing iteration and feedback. But if we stop there, we risk overlooking the complexity of contemporary writing. A rhetorical assemblage framework offers a complementary perspective. It shifts our attention from circulation to configuration, from smooth repetition to dynamic interaction, from shared tasks to emergent effects.
“In-the-loop writing,” then, is not best understood as a closed system of exchange. It is an open, evolving assemblage continually shaped by human and nonhuman actors, by friction and alignment, and by the broader conditions in which writing takes place. And once we see writing this way, the question is no longer just how to stay “in the loop,” but how to understand—and intervene in—the assemblages that make writing possible.
References
DeLanda, M. (2006). A new philosophy of society: Assemblage theory and social complexity. Bloomsbury.
Knowles, A. M. (2022). Human-AI collaborative writing: Sharing the rhetorical task load. Proceedings of the IEEE International Professional Communication Conference (ProComm), Limerick, Ireland, 257-261. https://doi.org/10.1109/ProComm53155.2022.00053
Knowles, A. M. (2024).Machine-in-the-loop writing: Optimizing the rhetorical load. Computers and Composition, 71, Article 102826. https://doi.org/10.1016/j.compcom.2024.102826
Mason, E., & Dvorak, K. (2026). Looping generative AI into writing center consultations. In E. H. Buck & J. Botvin (Eds.), Writing centers and AI: Generating early conversations (pp. 261–270). The WAC Clearinghouse/University Press of Colorado. https://doi.org/10.37514/PER-B.2026.2791
Ranade, N., Saravia, M., & Johri, A. (2025). Using rhetorical strategies to design prompts: A human-in-the-loop approach to make AI useful. AI & Society, 40, 711–732. https://doi.org/10.1007/s00146-024-01905-3
Generative AI Disclosure
Generative AI (GenAI) was used to consider (re)organizational approaches and polish the prose. For polishing, it was explicitly prompted to “Revise this writing for clarity and coherence while maintaining my own voice and style. Check with me for any changes you want to make, identifying exactly where the change(s) may need to be made in the text”.
