Introduction
Table of Contents
General Introduction
Personal Reflection
Why this project, and why now?
Even decades ago, as a student of linguistics and cognitive science, I was aware of the concept of a general-purpose type of “linguistic engine” or “language model” that could take human text as input and provide coherent, semantically plausible output. This type of work had been foreshadowed at least as early as the mid-1960s ELIZA chatbot; I had played with the chatbot SmarterChild myself in my teenage years; and John Searle’s 1980 Chinese room thought experiment was already a classic touchstone in the fields of linguistics and cognitive science in arguing that consciousness was not necessary for a machine to produce plausible general-purpose language output (in other words, an argument against inferring that an advanced chatbot had any sort of mind in the sense that human beings understand the term).
To the best of my recollection, I became aware of the current wave of language model development with the publication of “A robot wrote this entire article. Are you scared yet, human?” in The Guardian in September 2020. The article was a showpiece for the output of the GPT-3 model, arguably the first large language model capable of producing coherent paragraph-length texts. (Somewhat misleadingly, the piece in The Guardian was not produced solely by GPT-3, but by human editors who generated several small texts using GPT-3 and then stitched them together into a more coherent-sounding composition.)
It occurred to me that this sort of text generator might be of concern in the areas of fraud and disinformation, but I gave it little thought as being relevant to authorship per se.
In June 2022, I obtained access to the OpenAI Playground, a website which provided limited access to variants of the GPT-3 model. This was a “pre-chatbot” version of the tool that would essentially auto-complete a text the user entered. (For example, instead of giving direct commands such as “Write me a sonnet about cats,” the user would write something like, “The following poem is a sonnet about cats:” and let GPT-3 complete the composition.)
I explored the tool in two main ways.
First, I played with GPT-3 with my son (then five years old) to generate silly outputs, such as a recipe for bug soup or pretend dialogs between fictional characters he liked. I kept screenshots of our favorites in an album called “Appa talks to a robot 🤖”. The limitations of the tool were clear- it lacked imagination and was often only marginally coherent- but I felt it facilitated a sort of playful reading practice for my child, as the outputs were novel and silly.
Second, as an instructor of ESL writing and as an educational technology specialist, I made my first tentative explorations of how the tool might impact ESL, authorship, and academic integrity. I wrote memos to my colleagues explaining the new tools and what they appeared to be capable of. In one October 2022 email, I wrote:
In this short AI activity [in one of my writing courses], I opened OpenAI Playground and entered a prompt like the following:
The following is a well-structured five paragraph argumentative essay. The essay argues that the English Language Institute should give its students free ice cream every Friday.
That prompt created a skeletal- but recognizable- five-paragraph essay. It was actually a really useful activity for reviewing essay structure because the AI-written essay had a thesis statement; topic sentences for the body paragraphs; and a conclusion that restated the thesis. It was very short on supporting detail, but it was a pretty good outline for a workable essay. The AI chose believable key points for the body paragraphs, for example, such as talking about how free ice cream would improve student morale and attract more students to the ELI.
ChatGPT launched the next month, November 2022, built on those GPT-3 era models I had already been experimenting with. Through its intuitive conversational interface and free-tier access, ChatGPT was the commercial product that made Generative AI accessible to a mass audience in the United States. As GenAI tools became a cultural phenomenon, I continued intensively exploring these tools and their implications for education as they emerged, but it is safe to say the topic was no longer “niche” from November 2022 onward.
Eventually, I realized my GenAI experiments were becoming too scattered to keep easy track of. In October 2023, I created a GitHub repository called llm-prompts-for-education to keep a record of the “prompts” I was developing, i.e. carefully structured and worded tasks and parameters for a large language model chatbot to follow in a given context window. (This type of “prompt engineering” has gone somewhat out of fashion, but lives on in various contexts such as .md context files.)
My impression as a writing instructor was that fall 2024 was the first semester in which students’ essays almost all bore noticeable signs of AI assistance, often in ways that damaged the coherence of the texts or violated genre norms. For example, students who submitted outlines for 7-paragraph essays might turn around and submit drafts with 15 to 25 unfocused paragraphs, sometimes interspersed with ordered or bulleted lists or other extraneous materials, all strongly in the “AI chatbot voice.”
By fall 2025, it remained clear to me that AI usage in student papers remained widespread, but the quality of the GenAI tools had improved enough that the fall 2025 papers were not as obviously deficient as the fall 2024 papers. By this I mean that essays that appeared to show strong AI influence tended to have structures that more closely matched the assignment; tended to be more coherent and disciplined in content; and exhibited far fewer genre-breaking features such as bulleted or ordered lists.
What has been invisible to me, especially as an instructor focused on ESL support, is to what degree the AI tools unproductively replaced student effort and thought in these cases, versus to what degree students were making good-faith attempts to have an AI tool massage their own arguments, drafts, and notes into prescriptively correct English prose.
Either way, this phenomenon represented a crisis for traditional assessment methodology. The instruction and assessment of writing has long been predicated on performance assessment: the idea that by analyzing a student’s submitted work, you can make reasonable (if imperfect) inferences about their underlying writing competence. This holds doubly true for ESL writing assessment, as the performance assessment endeavors to measure both writing competence and English language competence. If AI tools were now responsible to a large degree for the final submitted form of student essays, the theory and mechanism behind performance assessment breaks, as though one is assessing a golfer’s physical fitness by how fast they can drive a golf cart.
These issues resist any easy resolution.
One school of thought championed by a surprisingly large segment of educators is to return to fully-offline, paper-and-pencil writing; but for most purposes, this strategy strikes me as deeply and harmfully artificial. The simple truth is that academic and professional writing is a digital-first, heavily tool-mediated process. We use templates, word processors, citation managers, and numerous other digital tools and resources to compose serious texts in our fields. In the messy apprenticeship of writing pedagogy, we need to treat our students as junior colleagues and enable them to attempt ever-more-ambitious approximations of both workflow and product sophistication. It would be difficult to persuade me that a reactionary paper-and-pencil pedagogical program could ever prepare students to produce the sorts of products expected in real-world academic and professional settings. Perhaps fully-offline and analog authorship may suffice for short written answers on an exam, but it is not compatible with serious modern authorship.
Although I reject the full-analog movement, I am open-minded about what a digital, tool-mediated authorship process ought to look like. I consider myself neither to be an AI “booster” nor a prohibitionist. I can see plausible workflows and circumstances in which Generative AI tools can take on roles similar to an administrative assistant or research assistant and better enable an author to tackle large, significant projects successfully; but it is equally obvious that overreliance on (or poorly considered usage of) the same tools can enable an author to cognitively disengage from the drafting process, harming both learning and the fundamental notion of authorship.
To some degree, the problem is one of tooling and interface. It is plausible that better-designed and better-guardrailed tools, deployed within the context of well-designed assignments, can scaffold learning and accelerate authorship in ways that enhance both learning and output. Yet even if such a place for AI in the author’s and student’s toolkit reaches proper maturity, I am unable to easily imagine how any author or student could be kept within those boundaries of productive practice; there will always be less-helpful, less-productive tools and workflows available as well, whether prohibited by academic and industry standards or not.
But perhaps this shadow-side of authorship is not a wholly new issue: plagiarism, ghostwriting, and even substance abuse have long been available to authors. Perhaps the misuse of Generative AI tools- whatever consensus, if any, emerges on what “misuse” looks like- is just one more vice to add to a much older list.
I offer this work, then, not in a utopian (or even particularly optimistic) spirit, but simply in a spirit of an open-minded tinkerer. That there is presently a crisis in both the field of writing pedagogy as well as in the broader conception of authorship is obvious. I do not presume to be able to resolve this crisis, but I believe that looking at the issue through the lens of extended cognition- to acknowledge that authorship has always been a social, cognitive, and tool-mediated phenomenon larger than our individual brains, and to try to identify some of the ways in which Generative AI tools might play productive roles in that process- at least avoids the Scylla of unrealistic prohibitionism and the Charybdis of harmful, unreflective, uncritical overuse of Generative AI tools.
Bill Price, 2026
Acknowledgements
Instructor Notes
Implementation guidance:
- Note that this guide is written for students at the University of Pittsburgh. It may contain useful information for those in other contexts, but instructors may need to edit these materials to provide institution-specific guidance.
Prerequisite knowledge:
- tips here
License: This chapter is licensed under CC BY 4.0.
Cite this chapter as: Price, B. (2026). Introduction. In Authorship & AI: Modular OER for Responsible Academic Writing with Generative Tools. University of Pittsburgh. https://billcprice3.github.io/authorship-and-ai/005-introduction.html