The popular imagination of AI-assisted writing tends toward fluency. Writers are promised a seamless co-author, one that generates, suggests, and refines without complaint. My own writing loop resists this story. Friction, in my experience, is not an exception to be smoothed over; it is a constitutive feature of working with generative AI. This piece thinks through those moments of breakdown; the hesitations, refusals, and quiet negotiations that structure my encounters with AI suggestion, and argues that attending to them carefully reveals something important about who these tools are built for, what kinds of knowledge they recognize, and where rhetorical power resides.
Writing scholars have long understood that technologies of writing are never neutral. In their foundational analysis of computer interfaces as political sites, Selfe and Selfe (1994) demonstrated that the electronic writing classroom reproduced dominant cultural values—Standard American English, corporate literacy norms, and rationalistic ways of knowing—under the guise of democratizing access. What they identified in the graphical interface of the 1990s applies with renewed urgency to the AI writing assistant of the 2020s. When a language model flags my syntax as problematic, suggests I compress an argument, or offers a sentence that is cleaner but emptier, it is not malfunctioning. It is doing what it was trained to do, that is to reproduce the dominant way of meaning making. The question Selfe and Selfe asked of computer interfaces—whose interests do these designs serve? —is precisely the question we must ask of AI writing tools today.
Knowles (2022) developed the framework of rhetorical load sharing to describe how AI writing assistants distribute cognitive and rhetorical work across the five canons of rhetoric. Tracing how tools affect a writer’s burden at the stages of invention, arrangement, style, memory, and delivery, Knowles shows that collaboration with AI is never merely additive, rather, it is redistributive. But redistribution carries costs. When algorithmic suggestion dominates stylistic decisions, the human writer does not simply gain time; she surrenders a site of agency. When AI shapes invention, the range of what feels thinkable narrows. Rhetorical load sharing is not a simple gain; it is a negotiation in which voice and judgment are perpetually at stake.
My loop resists at particular, recurring points. When I draw on epistemological traditions that are not Anglo-American in origin, those traditions are smoothed into a generic academic idiom. For instance, I wanted to integrate Nyāya Sūtra (an ancient Indian Justice rhetoric) to my Technical Communication class to situate, broaden, and diversify the needs of globalization in the curriculum. When I tried to have AI generate some ideas regarding an user experience assignment for my class, it kept giving me responses that are based on the traditional US centric such as. It flattened Nyaya by retrofitting it into Anglo-American frameworks, reducing holistic, soteriological systems to abstract, secular logic. It described the Nyāya Sūtra as a manual on “formal deductive syllogisms”, stripping away its core spiritual purpose. To Illustrate-
- The “Epistemology” Framework: “Gautama’s text presents a foundational empiricist epistemology, evaluating reliable truth conditions (pramāṇa) through perception, deductive inference, analogy, and testimony.”
- The “Syllogism” Translation: “Nyāya offers a formal, five-step logical argument (the avayava model) used to deduce empirical truths, mirroring Western categorical logic.”
When a delayed subject or an inverted clause is doing rhetorical work, it is flagged as awkward and corrected away. They are patterns, and patterns have politics. Vee (2023) has described large language models as systems trained to produce the kinds of prose that count as authoritative within the dominant Anglophone academic tradition. Writing that departs from that tradition risks being registered not as difference but as error. The empirical record on this point is sobering. Hofmann et al. (2024) demonstrated that language models embody covert racism in the form of dialect prejudice, producing raciolinguistic stereotypes about speakers of African American English more severe than any human stereotypes ever experimentally recorded. Critically, safety training designed to suppress overt racist outputs did not resolve the underlying bias—and in some cases intensified it, teaching models to conceal discriminatory judgments beneath apparently neutral language. The implications extend well beyond any single dialect. They point to a structural truth: the training process does not merely learn language; it learns whose language counts.
Nowhere is this more consequential than for international and multilingual scholars writing in English-dominant academic contexts. My positionality as an international scholar complicates every interaction I have with AI suggestions. Authority and credibility are negotiated differently across cultural, linguistic, and disciplinary contexts, and AI tools are largely incapable of recognizing these negotiations. What they offer instead is a normalized standard that presents itself as quality while encoding, invisibly, assumptions about who the default academic writer is and how that writer should sound. Sano-Franchini, McIntyre, and Fernandes (2024) have argued for the importance of refusal as a critical practice in writing studies, drawing attention to how generative AI tools amplify existing linguistic hierarchies and marginalize diverse literacies. Their call for refusal resonates with the earlier work of Selfe and Selfe (1994), who reminded us that the rhetoric of technology—the promise of democratization and improved communication—consistently obscures the ways that technological systems enact and sustain existing structures of power.
The moments when I refuse a suggestion, hesitate before accepting a revision, or restore what the model has deleted, are not interruptions to the collaborative flow. They are where authorship lives. Reza et al. (2025) found that writers consistently express a need not for smoother AI assistance but for more meaningful control, especially for the capacity to direct the collaboration rather than merely react to it. Their research reveals a persistent gap between what current AI writing systems offer (suggestions optimized for surface fluency) and what writers need (recognition of their rhetorical intentions and authorial agency). Friction is the symptom of that gap. It signals that the model has misrecognized something; registered as an error, what is in fact intention, flattened what needed to remain complex, optimized where the writer needed to stay rough.
There is a subtler dimension I want to name: the way that repeated AI-assisted writing begins to shape what feels writable. When certain constructions are consistently flagged, when certain epistemological moves are repeatedly smoothed away, the writer’s own sense of possibility narrows quietly. This is not dramatic censorship; it is a gentle, cumulative constraint. Robbins (2025) has theorized this through the concept of rhetorical debt, arguing that the responsible use of AI outputs requires human expertise and rhetorical judgment that the model itself cannot supply. A judgment that is, paradoxically, eroded by over-reliance on the very tools that demand it. Dependency on AI writing infrastructure can shape not only how we write but what we imagine we could write.
To position breakdown as a critical resource is not to romanticize difficulty or reject the genuine affordances of AI writing assistance. It is to insist that moments of friction in the writing loop are epistemologically rich—a map of what the system does not know, cannot recognize, and will not accommodate. Who gets to decide what counts as good writing? Which epistemologies are encoded within suggestion engines, and which are quietly erased? These are not technical questions. They are the constitutive questions of rhetoric and writing studies, and they return us, in every moment of breakdown, to the irreducible work of the human writer: to decide what the writing is for, what it must do, and who it needs to be.
References
Hofmann, V., Kalluri, P. R., Jurafsky, D., & King, S. (2024). AI generates covertly racist decisions about people based on their dialect. Nature, 633, 147–154. https://doi.org/10.1038/s41586-024-07856-5
Knowles, A.M. (2022). Human-AI Collaborative Writing: Sharing the Rhetorical Task Load. 2022 IEEE International Professional Communication Conference (ProComm), 257-261. https://doi.org/10.1109/ProComm53155.2022.00053
Reza, M., Thomas-Mitchell, J., Dushniku, P., Laundry, N., Williams, J. J., & Kuzminykh, A. (2025). Co-writing with AI, on human terms: Aligning research with user demands across the writing process. [Preprint].arXiv. https://arxiv.org/abs/2504.12488
Robbins, S. (2025). Large language models and the problem of rhetorical debt. AI & Society. https://doi.org/10.1007/s00146-025-02403-w
Sano-Franchini, J., McIntyre, M., & Fernandes, M. (2024). Refusing GenAI in writing studies: A quickstart guide. Refusing Generative AI in Writing Studies. https://refusinggenai.wordpress.com/
Selfe, C. L., & Selfe, R. J. (1994). The politics of the interface: Power and its exercise in electronic contact zones. College Composition and Communication, 45(4), 480–504. https://doi.org/10.2307/358761
Vee, A. (2023). Large language models write answers. Composition Studies, 51(1), 176–181.
AI Disclosure: I’ve used Claude to see if and how it smooths and erases the non Anglo-American rhetorical traditions.
