Where We Are - AI and Writing - Truth-Telling: Critical Inquiries on LLMs and the Corpus Texts That Train Them

Oct 15, 2026

The text examines how LLMs and their training corpus shape authorship, research, labor, copyright, and writing, using the “literacy crisis” as a framework. It argues that composition studies must respond critically to linguistic punishment embedded in public datasets, where context and shared knowledge are missing and outputs can reinforce stereotypes and harmful sentiment. It highlights power and ideology across higher education, AI companies, workplaces, and marginalized communities, and proposes integrating critical inquiry to strengthen students’ rhetorical awareness of what is at stake when they publish with AI.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Where We Are: AI and Writing

Truth-Telling: Critical Inquiries on LLMs and the Corpus Texts That Train Them

Literacy crisis as a framework for AI and writing

How LLM training works and why context matters

Linguistic punishment, power, and ideology in corpus data

Implications for composition studies and student writing

Where We Are: AI and Writing Truth-Telling: Critical Inquiries on LLMs and the Corpus Texts That Train Them Antonio Byrd The commonplace concept ‘literacy crisis’ has framed ChatGPT’s popu larity, its rapid evolution (GPT-4’s release five months after ChatGPT 3.5), and its seemingly sophisticated language and knowledge performance. The concept helps scholars and teachers easily enter conversations about AI text generation technologies and how they transform our notions of authorship, research, labor, copyright, and writing. I’ve used the literacy crisis as a starting point for a 2023 CCCC Annual Convention panel called “ChatGPT, Magical Thinking, and the Discourse of Crisis”(Byrd et al.) and again at my home institution for a virtual workshop on LLMs that I co-facilitated with colleagues from computer science and the university writing center. On both occasions, I shared my thoughts on how resisting AI text generation technologies suggested that writing instructors want to protect ideologies and power that may restrict the flow of literacy practices. We have an opportunity to uncover conflicts or tensions over power and ideology among higher education, the companies that develop LLMs, workplaces, and marginalized communities as AI text generation technologies reshape our literacy practices. I again use the literacy crisis as a starting point in this essay but in the opposite direction. By responding to literacy crises with a back-to-basics pedagogy, composition studies has participated in histories of linguistic punishment that have shaped how writers produce, edit, and publish texts; these punishments create the presence of power and ideology in the corpus texts LLMs use for training. While we learn new literacy practices and teaching strategies with LLMs, we also bear the responsibility for participating in the creation of the next iteration of public data that contain our contemporary ideologies on language and culture. AI and writing involve critical inquiry on our relationships with these corpus texts. Integrating this critical inquiry may maximize students’ writing practices with AI and extend their rhetorical awareness of what’s at stake when they go public with their writing to participate in cultural and political conversations. ChatGPT, along with other LLMs, train on public content extracted from the Internet (called datasets or corpus texts) to learn the nuances and patterns of human languages (some can also train on datasets that contain millions of images to generate multimodal content and analyze multimodal inputs to generate text as output in response). As chatbots, LLMs seem to be an effective Composition Studies 51.1 (2023): 135–142 knowledge system and natural language processor to human users; however, they have really created mathematical formulas to predict the next token in a string of words from analyzing patterns of human language. They learn a form of language, but do not understand the implicit meaning behind it. Meaning comes from a “shared knowledge that the interactants bring to the scene . . . That right there is the complexity of rhetoric: It is an art that attempts to deal with the inherent messiness of the richness and variability of communication context” (Professional Communication and Network 147) . Without accessing context, they are “stochastic parrot[s]” making stereotypical word associationsand negative sentiments about marginalized people without understanding why those words are harmful (Bender et al. 616) . AI safety experts can intervene in the training to limit these outputs to the user and mitigate hallucinations.(We can still work around those guardrails, however. Visit [jailbreakchat.com](jailbreakchat.com)[ ](jailbreakchat.com)[and you](and you)’[ll find prompts that have generated](ll find prompts that have generated) harmful outputs. OpenAI uses this crowdsourced information to make patches in its LLMs.) But scholars and teachers in writing studies recognize

FAQs

How does the text connect “literacy crisis” to LLM-driven writing?

It uses the literacy crisis to frame how AI text generation technologies shift ideas about authorship, research, labor, copyright, and writing. It then reorients the framework to show how composition pedagogy can respond to the power dynamics embedded in training corpora.

What does the text say about how LLMs produce text?

It describes LLMs as systems that predict the next token based on patterns in human language datasets. It argues they learn linguistic forms but do not access implicit meaning that arises from context and shared knowledge.

What is “linguistic punishment” in the context of LLM training?

The text defines linguistic punishment as violence and oppression tied to bodies and land, including colonialism, imperialism, genocide, and slavery, which helps create English dominance. It links these histories to downstream effects in corpus creation, marginalization, and harmful model outputs.

Oct 15, 2026

Upload by WPS_1790064749