AI serves as a new kind of participantin the learning process: a participant that can offerinstant feedback, generate creative prompts, andeven model revision strategies at any moment in thewriting process

Thesemodels don’t understand the text the same wayhumans do

Our discus-sions can shift from asking what changes are needed

We are both con-cerned that the struggle with writing, the effort ittakes to make a decision about the next best step inour piece, remains part of the process of writing

We used ChatGPTto save a bit of our mental energy trying to think upour students’ next steps and spent that energy insteaddiscussing the best options among the next steps thisprogram was suggesting

the setting, her main character was male, the twistwas more involved, and she removed the supernatu-ral elements entirely. She only needed ChatGPT toflood her with ideas.

In her final draft, much of what the studentincorporated significantly deviated from ChatGPT’ssuggestions

to what changes are suggested and whether they rep-resent the writer’s true intentions.

The primary function of anLLM is to predict the next word in a sequence basedon the words that came before it (think about theword suggestions when you text).

Historically, access to quality writing supporthas been unevenly distributed, often influenced bythe resources available in the school district. Studentsfrom under-resourced backgrounds have faced sig-nificant disadvantages in this regard. This AI tool,accessible to anyone with an internet connection,offers an opportunity to help level the playing field

Interacting with ChatGPT is an iterative pro-cess. You can and should re-prompt it to meet yourgoals and needs. This creates what is known as a chainprompt.

How peo-ple choose to use ChatGPT speaks to their intent,”observes Yoder-Wise. “That intent could be to enrichsomeone’s understanding of how better to articulatemessages. . . . The intent could also be to deceive oth-ers in the presentation of written materials”

the responsibilityof fully reading and evaluating the submitted work must always rest solelywith the human reviewer.

I’m particularly concerned with a tendency to “ask AI” aboutwhat it “thinks” about how it should be used or what it “knows” about ourfields of inquiry, as these prompts implicitly treat the AI as a subject withagency

Ideally, editorial policies should help authors use these tools in specific,targeted ways that don’t drive out linguistic variation and the richness ofglobal Englishes.

Eachprompt is, however, still significantly higher than the cost of a search enginequery (at least before the search engines added AI overviews), and the longerthe ChatGPT output, the more energy it uses (You).

I would prefer to see AI-produced text to be quotedand cited in the same way that we require recognition and citation of anyother text that an author draws on in a manuscript, regardless of the source.

although this use is, of course, counterto the goals of linguistic justice and ultimatelypushes academic language toward a moreconsistently and generically white male voice.

Perhapsthe most effective use I’ve seen has focused on assessing and improvingorganizational structure.

They will also make style suggestions to try to makethe text conform to the most common expression, based on its statisticalunderstanding of its training data (which is why much of the output, if notguided by extensive prompts, sounds so very generic)

In practical terms, this means that authors have an obligation toreview and evaluate any textual output from LLM-based AI systems. Evenwhen being used to help edit work originally produced by an author, thesesystems will correct most grammatical errors—but will also introduce newones in the process. They will also make style suggestions to try to makethe text conform to the most common expression, based on its statisticalunderstanding of its training data (which is why much of the output, if notguided by extensive prompts, sounds so very generic). In short, the outputcan never be trusted and thus must always be reviewed.

the system doesn’t “think” and doesn’t “know,” and if it outputsnew text that a human sees as false or incorrect, it’s not a “hallucination”—it’s just doing precisely what it was designed to do.

the system turns all those words and symbols into numbers and then findsthe relationships between and among those data points. When providedinstructions to produce new text, the system uses its statistical model toprovide human-sounding language in response. But, under the hood, it’sconverting everything to numerical values, performing analyses on thosevalues, and then converting the results back into words and parts of words.