How to cheat on your final paper: Assigning AI for student writing

Vol.:(0123456789)

1 3

AI & SOCIETY

https://doi.org/10.1007/s00146-022-01397-z

OPEN FORUM

How tocheat onyour ﬁnal paper: Assigning AI forstudent writing

PaulFyfe

Received: 13 September 2021 / Accepted: 4 February 2022

Abstract

This paper shares results from a pedagogical experiment that assigns undergraduates to “cheat” on a ﬁnal class essay by

requiring their use of text-generating AI software. For this assignment, students harvested content from an installation of

GPT-2, then wove that content into their ﬁnal essay. At the end, students oﬀered a “revealed” version of the essay as well

as their own reﬂections on the experiment. In this assignment, students were speciﬁcally asked to confront the oncoming

availability of AI as a writing tool. What are the ethics of using AI this way? What counts as plagiarism? What are the con-

ditions, if any, we should place on AIassistance for student writing? And how might working with AI change the way we

think about writing, authenticity, and creativity? While students (and sometimes GPT-2) oﬀered thoughtful reﬂections on

these initial questions, actually composing with GPT-2 opened their perspectives more broadly on the ethics and practice

of writing with AI. In this paper, I share how students experienced those issues, connect their insights to broader conversa-

tions in the humanities about writing and communication, and explain their relevance for the ethical use and evaluation of

language models.

Keywords Language models· Plagiarism· AI literacy· Writing· Pedagogy· Ethics

“The question then becomes, whose writing is this;

who can take ownership? The answer to this question

is not easy to decide and seems to be more complicated

than the question of whether it is true or false. I don't

exactly have the right to claim ownership of it, but I

will argue that all writers borrow ideas and style from

others.” (undergraduate student with GPT-2 italicized)

This paper shares results from a pedagogical experiment

that assigns undergraduates to “cheat” on a ﬁnal class essay

by requiring their use of text-generating AI software. For

this assignment, students harvested content from an instal-

lation of GPT-2, then wove that content into their ﬁnal essay.

At the end, students oﬀered a “revealed” version of the essay

as well as their own reﬂections on the experiment. The epi-

graph above comes from a student’s paper in which GPT-2

seems to join the conversation about the very ethical ques-

tions it has provoked. For their assignments, students were

speciﬁcally asked to confront the oncoming availability of

AI as a writing tool. What are the ethics of using AI this

way? What counts as plagiarism? What are the conditions,

if any, we should place on AI-assistance for student writ-

ing? And how might working with AI change the way we

think about writing, authenticity, and creativity? While stu-

dents (and sometimes GPT-2) oﬀered thoughtful reﬂections

on these initial questions, actually composing with GPT-2

opened their perspectives more broadly on the problems

and practice of writing with AI. In this paper, I share how

students experienced those issues, connect their insights to

broader conversations in the humanities about writing and

communication, and explain their relevance for the ethical

use and evaluation of language models.

1 Research questions andframeworks

This teaching experiment invites students into an urgent

conversation about the ethics of using AI in coursework.

Especially for courses that involve writing, students and

teachers must necessarily confront the problem of plagia-

rism with the wide availability of electronic sources and

online essay mills. As will be discussed below, the concept

of “plagiarism” itself needs more nuance, and it certainly

gets blurrier in context of using an AI. But in its familiar

form, especially as deﬁned institutional settings like my own

* Paul Fyfe

paul.fyf[email protected]

http://go.ncsu.edu/pfyfe

Department ofEnglish, NC State University, Raleigh, NC,

USA

AI & SOCIETY

1 3

university, plagiarism appears as the uncredited, knowing,

and sometimes wholesale adaptation of work that is not

one’s own.

For courses involving writing, that often mani-

fests as text copied from online sources or acquired from

vendors. Interestingly, artiﬁcial intelligence has upped the

ante, disabling the usual strategies of detecting plagiarism by

Googling phrases or checking papers against a known data-

base (e.g. Turnitin). Some online resources now advertise AI

to students to generate usable, unique text that is untrace-

able by current plagiarism detection software. For example,

claiming in a tagline to be “Empowered by Artiﬁcial Intel-

ligence,” EssayBot.com promises itself as “your personal

AI writing tool. With your essay title, EssayBot suggests

most relevant contents. It paraphrases for you to erase pla-

giarism concerns.” Marketing itself as a “bot” designed to

outmaneuver plagiarism, the site feeds into concerns about

the automation of writing and the erasure of human eﬀort.

In response, learning management systems and plagiarism

detection software are now adapting AI tools of their own,

locked in an arms race between crisply deﬁned antagonists:

systems to cheat artiﬁcially versus systems to insure origi-

nal work. Staking out this battle line, one recent plagiarism

detection product simply calls itself “Turnitin Originality.”

However, this conﬂict presents a false binary, an over-

drawn contrast for the sake of selling essays or student

integrity insurance. These days, computer- and AI-assisted

writing is already deeply embedded into practices that stu-

dents already use. The question is, where should the lines

be drawn, given the array of assistive digital writing tech-

nologies that many people now employ unquestioningly,

including spellcheck, autocorrect, autocomplete, grammar

suggestions, smart compose, and others? Asking students

to “write with AI” can usefully provoke conversations not

only about extreme examples of essay bots, but about eve-

ryday technologies, too. Within this spectrum of practices,

what are the ethical thresholds? At what point, in what con-

texts, or with what technologies do we cross into cheating?

Should that concept be redeﬁned? The scholar Margaret

Price has recommended that “the most constructive way to

approach teaching on plagiarism is to invite students into

a dialog about the subject, welcoming their perspectives

on its complexities” (2002, 105). Those complexities have

only increased with the ubiquity of AI. Oﬀering a similar

take on “How Teachers Can Prepare for AI-Based Writing,”

Jonathan Bailey (of Turnitin, note) suggests that teachers

talk to students about these issues now, before “the inevita-

ble day when students can push a button and the computer

writes their paper” (Bailey 2020). That forecast seems a bit

dramatic, but dramatic forecasts (as in science ﬁction about

humans and AI) often help us reconsider the nuances of our

own values, ethics, and relations to technology. In evolving

forms, AI is the present and the near future of what faces

students and educators.

Instead of a prescriptive approach to AI-assisted writing

and plagiarism, assignments like this try something more

proactive, allowing students to experience and then articu-

late for themselves the issues at stake. Furthermore, this

approach moves students beyond antagonistic discussions of

plagiarism to consider the potential uses of AI as a writer’s

tool. As Grant Otsuki argues, there’s no putting the AI genie

back in the bottle and little sense banning its use outright.

Instead of teachers “pretending AI doesn't exist” or treating

it merely as an antagonist, Otsuki encourages us to bring it

into the classroom: “it might be time to train people to write

with AI” (2020). What would that look like? While many

writers are familiar with various assistive technologies, there

are practically no examples for training writers with text-

generating AI and language models like GPT-2.

Thus, the

research questions from my experiment also explore stu-

dents’ strategies for AI-assisted composition. How might

we write with these tools? What skills might writing with

AI newly require? How does this challenge our assumptions

about textual communication? And what potential risks and

harms must we navigate? “Cheating” oﬀers a starting point,

but it opens onto much more complex and consequential

topics, from human and nonhuman agency to the threats of

disinformation and algorithmic bias. Thus, the assignment

moves students from a familiar ethical context (is this plagia-

rism?) to new ethical questions (whose writing is this? what

eﬀects does it have?) and to new forms of writing practice

(can I collaborate with AI?). The resulting insights can join

a much broader conversation about AI literacy, including

whether or how we might ethically and productively work

with AI.

Additionally, in this paper, I want to suggest how discus-

sions of AI literacy can be enriched by ongoing scholarly

conversations about plagiarism, authorship, and writing

pedagogy. For example, composition and rhetoric scholars

have been working thoughtfully on these issues for dec-

ades, having challenged the concept of plagiarism long

before the ubiquity of the internet or machine learning. As

Plagiarism appears as its own section of “Academic Misconduct”

(8.4) in NC State’s student conduct policy (“POL 11.35.01 – Code of

Student Conduct” 2020).

See Elkins and Chun for a report on an assignment oriented more

toward literary writing (2020). See also the reﬂections by Lang etal.

(2021).

As deﬁned by Long and Magerko, AI literacy includes “a set of

competencies that enables individuals to critically evaluate AI tech-

nologies; communicate and collaborate eﬀectively with AI; and use

AI as a tool online, at home, and in the workplace” (2020, 2). See

also Ng etal. for a review of recent approaches to deﬁning AI literacy

(2021).

AI & SOCIETY

1 3

Price argues, “plagiarism is not stable” but situated diﬀer-

ently “across historical time periods, across cultures, across

workplaces, even across academic disciplines” (2002, 90).

In our own time, educational institutions continue to deﬁne

plagiarism in ways that idealize originality, a romanticized

concept of a masculinized, independent, unified author

which ignores all the rich dialogic messiness of language

and the intertextuality of writing (Miller 2004; Flannery

1991). And emphasizing “originality” tends to punish the

legitimate ways in which students come to learn a new dis-

course, often by imitation. For instance, in “patchwriting,”

students might swap or stitch their own words into a source

text which, according to Rebecca Moore Howard, is part

of a “healthy eﬀort to gain membership in a new culture”

(1992, 236). Howard reframes patchwriting as a “pedagogi-

cal opportunity, not a juridical problem” under the regime of

plagiarism (1995, 789). Compositionists have also updated

this pedagogical ethos for the digital age, suggesting how

plagiarism anxieties tend to restrict the creative possibili-

ties of networked multimedia and the “nature of writing in a

remix culture” (Johnson-Eilola and Selber 2007, 381). When

instructors strategically decriminalize plagiarism, they also

open opportunities for creative and critical exploration.

Along these lines, Johnson-Eilola and Selber propose that

writing-as-remix activities “oﬀer important new ways for

thinking critically and productively about what it means to

write, about what it means to read, and about what we value

as texts” (2007, 377).

Throwing AI into the mix updates these questions in

interesting ways—and helps to move discussions about

language models beyond simplistic contrasts of human

originality vs. machine imitation. That binary currently

shapes so many responses to text-generating AI, whether in

amazement about the coherent humanness of its outputs, or

in sharp criticism of its failures to make sense or construct

logical arguments (Branwen 2020; Marcus and Davis 2020;

Seabrook 2019; Lea 2020). Yet such responses are entirely

consistent with how AI—especially AGI, or artiﬁcial general

intelligence—has historically been framed. In the long view,

AI frequently appears as an imitation game or even as a prac-

ticed illusion (Standage and Stevenson 2018). Commenta-

tors still tend to fall into the Turing trap, as it were, evaluat-

ing any AI-generated language within a reductive framework

of “bot or not” (Hua and Raley 2020). That reaction may

show an anthropocentric bias, judging artiﬁcial intelligence

only to the degree it resembles a human’s. According to

Miller, it also shows our deep “commitment to agency” that

automation seems to challenge: in other words, it’s not just

whether synthetic writing sounds like a human; it’s whether

we can relinquish the humanistic concept of agency that has

informed philosophy, politics, and pedagogy for a very long

time (2007, 141).

Discussions of posthumanism are more open to the alter-

natives, and help signal how to approach writing with AI

beyond the human/machine binary. Well before AI started

to blend into our normal practices, Haraway declared that

“writing is pre‐eminently the technology of cyborgs” (2016,

57). Put diﬀerently, writing simply does not exist outside of

the technologies used to produce it. And when technologies

for writing change, so do its conceptual horizons [Perhaps

writers change too, in a symbiotic process of “technogen-

esis” (Hayles 2012, 10)]. Writing with AI may force us into

a heightened awareness of these posthuman dependencies,

not only asking us to reexamine our deﬁnitions of writer,

text, and reader, but to reevaluate our very identities within

technological systems. What does that look like in prac-

tice? Certainly discussions of these issues are important,

but active experimentation has a major role to play. As Casey

Boyle argues, discussion alone preserves the split between

humanistic minds and the technological objects they sepa-

rately consider. Instead, Boyle recommends helping students

to realize their embedded relationships to technological sys-

tems (2016, 540). A “posthuman practice” of writing, as

Boyle calls it, would immerse students in mediated envi-

ronments and stretch the ways they conceive of agents, text,

networks, and communication. Perhaps most importantly

for my argument, Boyle claims that such immersion oﬀers

a better place to create an ethics appropriate to our own

complex technological systems. As he suggests, “ethics in

a posthuman practice are not ideals imposed upon actions

we ought do but are instead ongoing exercises whose aim is

to compose new capacities for conducting ourselves within

expanded media ecologies” (Boyle 2016, 549).

In such a spirit, this assignment places students within the

media ecology of text-generating AI and asks them to articu-

late the ethics practiced therein. Plagiarism oﬀers a famil-

iar starting point for that exploration, and the varied ideas

and critiques students propose—soon to be discussed—are

interesting enough for continuing discussions of AI and writ-

ing. In a larger sense, students’ experiments with GPT-2

also represent a diﬀerent approach to AI literacy, including

evaluating language models, their risks and harms, and the

responsible practices of using (or not using) them.

As Minh

Hua and Rita Raley have argued of the story-generating

game AI Dungeon, its players become “necessary partners”

with its underlying AI in a creative, critical, and functional

relationship to it. Hua and Raley describe that relationship as

a kind of “citizen NLP” and an example of “how humanists

Long and Magerko recommend that “Researchers seeking to foster

AI literacy may want to avoid misleading tactics like Turing decep-

tions and black-box algorithms” (2020, 8)—and I could not disagree

more. Inviting students into this experience instead resembles how

Meredith Broussard works with AI “to commit acts of investigative

journalism” into its consequences (2018, 6).

AI & SOCIETY

1 3

might more meaningfully and synergistically contribute”

to the assessment of language models (2020). While that’s

a reach goal for an undergraduate writing assignment, it

underscores the value of their participation, even if—and

especially if—those students do not hail from computational

disciplines. In this spirit, I oﬀer my students’ experiments

as test case of the kind of immersive, applied ethics that the

humanities can bring.

2 Materials andmethods

This assignment was scaﬀolded into the ﬁnal module of the

honors course “Data and the Human” at NC State University

(HON 202–006, Fall 2020).

The class enrolled 20 students

from ﬁrst-years to seniors with a variety of majors and con-

centrations. The course was designed as an interdisciplinary

seminar, welcoming any motivated student, and requiring

only the most basic technical abilities. Its goal was to cul-

tivate what has been called “critical data literacy,” combin-

ing a practical understanding of data and its transformations

with a critical awareness of how data organizes knowledge

and orchestrates power (Boyd and Crawford 2012; D’Ignazio

and Klein 2020; Battista etal. 2020; Crawford 2021). The

course carried students through three modules: data, pri-

vacy, and surveillance; deﬁnitions and critical dimensions

of data; and ﬁnally machine learning and AI. Each module

concluded with a hands-on experiment and accompanying

report, in which students practiced with some of the tech-

nologies in question, and then reﬂected on what they expe-

rienced in context of the assigned readings and class discus-

sions from that module. The NC State University Libraries

generously provided students workshops for students to gain

basic skills in all the tools the assignments required.

For the course’s ﬁnal module on AI, students had to write

an essay that integrated the output from a text-generating

language model into their own writing—but without reveal-

ing which was which. Styled as the “Professor Fyfe Turing

Test,” students were encouraged to try and make indistin-

guishable what essay text came from the AI or from them-

selves. I imposed no quotas for the amount of AI-generated

text nor strictures about the ways it appeared. In the essay’s

ﬁnal section, students composed exclusively in their own

voices (i.e. without AI), reﬂecting on the following questions

given in the assignment prompt:

How easy or not was it to write this way? What worked

or what didn’t? How did the AI-generated content

relate to your own? How did it aﬀect what you might

have thought about or written? Do you feel like you

“cheated”? To what degree is this paper “your” writ-

ing? Do you expect a reader would notice GPT-2’s text

versus your own? Would you use this tool again, and

in what circumstances? And, ultimately, what ideas

about writing, AI, or humanness did the experiment

test or change?

Finally, all papers included an Appendix in which stu-

dents provided a “revealed” version of their essay with the

AI-generated text highlighted. The assignment yielded 20

student essays of 1500 + words each. All students gave their

permission to have their work considered and quoted in my

own reports about the experiment.

Many of the assignment’s questions were open ended,

leading students into certain considerations while also let-

ting them articulate their own insights. While evaluating

these papers for our class, I captured responses to those

guiding questions to include in the metrics and discussion

to follow. I also compared those responses to students’ aca-

demic disciplines and majors—which, as it turned out, did

not factor at all into the patterns of responses. This explora-

tory project oﬀers its ﬁndings as preliminary, encouraging

more research perhaps with focused survey instruments or

keying assignment questions and responses to emerging

frameworks for AI literacy.

The language model students used to generate text was

a medium-sized out-of-the-box installation of GPT-2, com-

passing 774 million language parameters. GPT-2 or “general

purpose transformer” is an unsupervised language model

trained on a large corpus from the web, released by OpenAI

in 2019 (Radford etal. 2019).

It takes an initial “prompt”

and attempts to predict, based on its learned parameters,

what text is likely to follow. Scott Bailey, then a Digital

Research and Scholarship Librarian at NC State, collabo-

rated on the assignment design and created the interface for

students.

We chose GPT-2 after reviewing what text-gener-

ation software, available at that time, might be most acces-

sible to students.

We then created a user-friendly Python

For an extensive introduction to this class of machine learning soft-

ware as well as its emerging uses and risks, see Bommasani et al.

(2021).

Scott has since moved to a position at the software company Sourc-

egraph.

A quick demo of GPT-2 can be done online with the site “Write

With Transformer” created by the company Hugging Face: https://

trans former. huggi ngface. co/ doc/ distil- gpt2 In a subsequent version

of this assignment, students used the open-source language model

GPT-J which also makes a web-based demo available in a browser:

https:// 6b. eleut her. ai/.

Such cross-disciplinary approaches might especially be warranted

given the landscape of ethics training in CS programs (Raji et al.

2021).

The course syllabus and module assignments are available open

access in the MLA CORE repository: http:// dx. doi. org/ 10. 17613/

0h18- 5p41.

AI & SOCIETY

1 3

notebook in Google Colab which students copied into their

own university-aﬃliated Google Drives to customize. With

just a few clicks, students could easily activate the model and

adjust a few settings, including the initial text prompt, the

number of outputs, and the length of each output. Basically,

they could enter their own text and have the model generate

chunks of what it predicted came next. From that point on,

“writing” the paper was up to them.

3 Results anddiscussion

At ﬁrst glance, many students expected that this assignment

would be easy, as if the time had arrived when they could

“press a button and the computer writes their paper” (Bailey

2020). But students discovered that this was not easy at all.

Most reported (87%) that it became far more complicated

than just writing the paper themselves. When AI gets associ-

ated with automation or claims to save time and labor, apply-

ing AI to writing can appear like a shortcut. This may hold

for generating text in shorter, formulaic genres like news

spots or sports items, as we had discussed earlier in the mod-

ule. But, as gets frequently observed, GPT-2 at least cannot

sustain logically subordinated writing in longer genres like

essays (Marcus and Davis 2020; Seabrook 2019; Elkins and

Chun 2020).

Students quickly realized this for themselves,

noting that “as it wasn’t easy as I thought it would be,” and

“integrating artiﬁcially generated text into my writing was

more of a curse than the blessing I thought it might be when

this project was ﬁrst explained,” among similar reactions.

All reported that GPT-2’s outputs were diﬃcult to control

and often strayed oﬀ topic or into nonsense. They had to

keep giving GPT-2 “a shove in the right direction,” and then

wrestle its outputs back into their contexts. At least without

more extensive ﬁne-tuning of the model, GPT-2 didn’t often

sound like them, and it certainly didn’t follow where they

wanted to lead.

These initial frustrations made it relatively

easy to counter any student’s received impressions of AI as

an automatized writer.

Immediately confronted with its limitations, students

also discovered GPT-2’s capacity to fabricate information,

including plausible-sounding false statements and even

quotes from non-existent experts. For instance, one student

noted that “sometimes the passage said the exact opposite

of what was true, but the way it was worded seemed so pro-

fessional and authentic I was almost convinced.” The risks

of such AI-generated misinformation have certainly been

noted, including by OpenAI itself (Hern 2019). Even more

striking than authoritative sounding text were quotes from

non-existent authorities. Another student included this ironic

quote from GPT-2 apparently about itself:

“It's deﬁnitely a breakthrough, and I think it will have

a signiﬁcant impact on the ﬁeld," said Professor Hervé

Gagnon from the department of Design at the OpenAI

and one of the researchers behind the project. "But I

think it's not yet ready to be used by the average per-

son,” he added.

Of course, there is no such professor. Fabricated quotes

like this shifted students’ attention from plagiarism to the

threats of “synthetic disinformation” and fake news (Kreps

and McCain 2020). As one student reﬂected, “I think that

the false information could be classiﬁed as cheating a lot

more than the actual process itself.” We had earlier read

articles about the potential threats of GPT-2 in generating

fake personae, expert testimony, and false reports (Metz and

Blumenthal 2019). But it’s a diﬀerent matter when students

become responsible for that material and whether or how

to deploy it within their own essays, requiring a “shift in

literacy practices” that accommodates algorithmic as well

as human writing (Laquintano and Vee 2017).

Beyond its obviously fabricated personae, GPT-2 also

raised questions for students about where or from whom it

derived its statistical imitation of English. When GPT-2 did

make sense, whose viewpoints did it express, and how could

we even identify or cite them? Even if a student wanted to

quote or credit GPT-2’s output for a statement or an idea—in

other words, even if they wanted to play fair by standard

rules of plagiarism—how would they go about it? There are

no citation styles for text generated by a language model,

much less any disciplinary frameworks that would accept

those contributions as valid. Conversely, whose viewpoints

did GPT-2 exclude, and thus what paradigms could it never

represent? Was GPT-2 a compromised source by default,

untrustworthy because of its blind spots, and at worst a

potential source of harm? Along these lines, several stu-

dents reﬂected on the model’s biases: as one put it, “[the

assignment] showed me that programs like these do contain

unjust prejudice from the data that they were trained on.” As

scholars have recently suggested of such language models,

“large datasets based on texts from the Internet overrepresent

hegemonic viewpoints and encode biases potentially damag-

ing to marginalized populations” (Bender etal. 2021, 1). We

had prepared for these problems earlier in the module with

assigned readings and discussion about algorithmic bias,

which four students linked directly to their own experiments.

Similar discussions can be found in the use of AI or automated

systems to grade student writing (Anson 2006; Anson and Perelman

2017).

We considered training the model on student’s own writing for

better results, but realized that would require a prohibitive and

potentially invasive amount of training data from students: 50 + MB

their writing in plain text. We agreed a simpler approach would still

achieve the assignment’s goals.

AI & SOCIETY

1 3

Students rightly presented a lot of critique, perceiving

their diﬀerence from GPT-2 in terms of viewpoints, goals,

and writing style. Yet many students were also surprised

by the degree to which GPT-2 seemed to resemble or even

anticipate their own thoughts. For some, the writing partner-

ship even worked nicely: “I was genuinely surprised with

how well some of the content ﬂowed with my personal writ-

ing and how it continued to sound like me.” Several others

likewise reported on the uncanniness of GPT-2’s language

imitation. Noting the “similarity between my writing and the

predicted text,” a student shared that “I felt like I was read-

ing potential sentences that I would have written in another

timeline, like I was glancing into multiple futures.” In this

case, AI unsettled the writer’s orientation to narrative: an

essay no longer unspooled along one single thread controlled

by the author, but seemed almost fractal, as if multiplying

the author and the possible pathways of their prose. It gave

this student pause, at least, by unsettling the familiar dimen-

sions of their writing and how they related to it.

A few students were startled that GPT-2 gave them sub-

stantive help with their essays. Occasionally, it did more

than pump out plausible text: its outputs could contain useful

ideas. A student revealed that “some of the things that the

program came up with were not things I would have thought

of on my own.” In a related way, other students found that

GPT-2 sometimes helped them articulate ideas still forming

or that they struggled to express. As one explained: “those

were the words that were on the ‘tip of my tongue’ and

GPT-2 was able to articulate them for me in a better way.”

Echoing this statement, another student acknowledged that

“GPT-2 aided me in formulating my thoughts, something

which I notoriously struggle with.” At a glance, these reports

show students gratefully working through writer’s block.

Interestingly, these writers continue to claim the writing

represents “my thoughts,” just with mechanical help from

GPT-2 in ordering them into grammatical sequence. But

how can a student recognize the untrained outputs from a

language model as their own ideas? In a sense, these stu-

dents are “patchwriting,” or borrowing and manipulating the

structures of expert writing to join a discourse more credibly

(Howard 1992, 233). According to Howard, patchwriting has

real pedagogical value in helping students overcome self-

consciousness about their writing or being outsiders. These

students adapted AI outputs in a similar spirit: “the way the

sentences were generated helped catalyze my thinking and

writing process and helped me continue the paper. For me

personally, I have all my ideas in my head but sometimes

my brain just blanks out at how to write coherent and easy

to read sentences and I end up getting frustrated.” Fears

about plagiarism can further cement this frustration, locking

students into feeling disqualiﬁed because of their perceived

deﬁciencies as writers.

As students accommodated AI outputs into their essays,

whether gratefully or begrudgingly, they all noticed diﬀerent

ways it impacted their writing. Some began to invent func-

tions for GPT-2 as a writer’s tool—ways that it might be

used credibly and in earnest. However, even they acknowl-

edged that results were somehow not their own. They faced a

tradeoﬀ: they could ﬁnd more outputs matching their tone or

style if they cared less about the content. For instance, “I was

usually able to ﬁnd sentences that ﬂowed in a similar way to

my writing even if it wasn’t what I had in mind in terms of

content.” Other students took the opposite approach: chang-

ing their own writing style to better match GPT-2’s outputs

and “to create a smoother ﬂow.” One student realized that, to

make this assignment work, “instead of desperately trying to

make GPT-2 generate text to convey my ideas, I should just

write my paper around whatever ideas GPT-2 conjured up.”

In these cases, GPT-2 was no longer articulating ideas for

students, but dictating what they said and how they should

write: “my own writing style adapted to more closely match

the style of the generated responses rather than the other way

around.” That had negative consequences for some, as when

a student judged that the “program decreased the quality of

my writing. Since I was trying to closely match the writing

style of the AI, I felt like I was not able to truly write like

myself.” Yet even that frustration may yield a net positive.

Ironically, when encouraged to cheat, these students instead

ended up reinforcing their own capacities as writers, the dis-

tinctiveness of their voices, and the tradeoﬀs if not outright

sacriﬁces of using AI.

The eight students who concluded that they had indeed

“cheated” on the assignment were committed to their own

voice and pedagogical value of writing. Using GPT-2

seemed to short circuit the very intellectual goals of an

essay: “I see writing papers as a far more expressive and

individualistic manner of displaying comprehension …

Writing is a skill that is able to be cultivated, but only

through practice and understanding of your own identity.”

That student’s reﬂection accords with some of the ways in-

ﬁeld experts also explain why neither writing nor evaluation

can be mechanized. As ChrisAnson summarizes, “The point

of writing in a course is for students to explore and reﬂect

on ideas through language, convey their own interpretations

and informational discoveries to others, and in the process

intersubjectively create purpose and meaning” (2006, 54).

Asking students to write with AI may paradoxically help

solidify these values, deepening the commitment of some

This may especially apply in context of ESL and language diver-

sity, as students and/or instructors disqualify their expression against

standard written English.

AI & SOCIETY

1 3

writers to do without AI altogether. As a student concluded:

“You’d think after years of writing papers I wouldn’t think

twice at letting a machine write one for me, but instead

the opposite happened.” Using GPT-2 only deepened their

desire for control, as another student echoed: “Ironically, I

feel like ‘cheating’ actually sabotaged my ability to write

this lab report how I would have liked.”

But other students were less sure they had really cheated,

accepting GPT-2’s contributions less as “plagiarizing a

paper from a peer,” and more as a kind of collaboration

we have yet to deﬁne. In general, these students were more

open to diﬀerent conﬁgurations of authorship and writ-

ing. One student insightfully explained how this requires

conceiving AI less antagonistically: “AI assistants were

not meant to replace or impersonate humans but provide a

bridge that connects our ideas with theirs—a hybrid.” She

moved beyond the binary of bot or not to understand AI as

a collaboration between partners who are good at diﬀer-

ent things. This insight resembles how Ted Underwood has

tried to reframe popular understandings of machine learn-

ing: “models will matter not because they’re ‘intelligent,’

but because they allow us—their creators—to collabora-

tively explore a latent space of possibility” (2021).

Thus,

this student did not mistake GPT-2 for a “peer” at all, but

approached it instead as a tool for her own exploration. Such

exploration became possible not only because of GPT-2, but

because “the threat of punishment wasn’t hovering over me,”

she said. This comment underscores two things. First, stu-

dents can perceive creative and critical opportunities when

allowed to experiment beyond the juridical framework of

plagiarism. Inviting students to openly vs. clandestinely use

AI may have shaped their responses to the assignment, but

removing the threat of plagiarism opens participation in

new ways and to writers of diverse backgrounds. Second,

conventional deﬁnitions of plagiarism already construct AI

as an intelligent peer, which is not only a mistake, but may

foreclose on AI’s more interesting possibilities, including

what “hybrid” or collaborative conﬁgurations writers might

discover.

LeahHenrickson has noted that “we do not know where

computer-generated texts ﬁt within our current conceptions

of authorship and reading” (2021). In many ways, they don’t

ﬁt at all, scrambling the ways we recognize and deﬁne how

texts communicate. These ambiguities tended to interest

students in the “not cheating” camp. As one wrote of their

hybridized essay, “it doesn’t feel like something I’d write

but it also doesn’t not feel like something I’d write.” The

vagueness of this statement is appropriate, as the demarca-

tions of agency (what “I’d write”) are no longer clear. The

double negative—it’s not not my writing—suggests our lack

of vocabulary for recognizing ourselves in such an entangled

relationship. Similarly, another student noted the strange-

ness of not recognizing the sources of their own essay after

the fact: “when reading this essay, I often have to glance at

the appendix to remind myself which phrases are not my

original thoughts.” They have not merely fooled themselves

with a Turing test. Rather, this student implies how they’re

caught between a familiar framework for writing (“my origi-

nal thoughts”) and a hybrid one we have yet to build.

Several scholars have tried to articulate what this frame-

work might be. Instead of using a “simple author-not author

dichotomy” to judge computer-generated texts, Henrick-

son proposes that we place them along “a continuum from

authorship to generatorship” (2019). This results in a more

ﬂexible concept that Henrickson calls “algorithmic author-

ship” (Henrickson 2019; 2021). David Rieder sees this less

in terms of authorship than as the “domain of the digital

rhetor,” in which new conﬁgurations of digital media con-

tribute to the rhetor’s goals of persuasion.

Other scholars

have proposed “co-creativity” and “synthetic literature” to

describe writing with AI and the hybridity of computer-gen-

erated texts (Manjavacas etal. 2017). Notably, these argu-

ments derive primarily from experiments in creative writing,

new media art, and electronic literature rather than exposi-

tory writing. That makes sense, given the relative openness

of these studies to ambiguity and the sociotechnical interests

of contemporary critics. Far less studied are how such hybrid

frameworks might apply to argumentative prose, or how AI-

generated materials might factor in scholarship on posthu-

man rhetoric and assemblage. Perhaps research and practice

in new media arts point the way toward how language mod-

els might be used in good faith, used as “an instrument that

extends, and fundamentally transforms, the human writing

process” (Henrickson 2021).

Ten students overall answered that they had not cheated.

Some of them hedged that while they didn’t technically pla-

giarize, it still felt wrong. But about a quarter of the class

had no need of such nuances. Intriguingly, they remained

conﬁdent in their status as authors of their papers, and ﬁrmly

defended their work as “not cheating” even within the con-

ventional framework of plagiarism. They claimed to have

done original, intellectual work—just not necessarily by

writing. Instead, they explained their eﬀort more in terms

of assembly and editing. The AI could generate grammatical

text, but until that text was sifted, organized, and stitched

into an essay, the AI was not writing, per se. One student

See also Reid on the “possibility space” of nonhuman rhetoric

(2020).

Personal conversation with the author. Rieder’s own experiments

in physical computing and new media composition are similarly

interested in hybridity. He describes this as “everting” new media, or

“infusing the real with aspects of the virtual.”

AI & SOCIETY

1 3

reasoned that they hadn’t cheated because “I ended up

doing most of the work to help make the statements useful.”

Another agreed, making a fascinating comparison to explain

their point: “I would say that I still wrote this paper. The

diﬀerence is that I did not feel like a traditional writer. I felt

more like Dr. Frankenstein, stitching together half sentences

and incoherent AI words into something more cohesive.”

While the student did not pursue the analogy to Franken-

stein, we might consider the provocative relation of suturing

texts from GPT-2 to the surreptitious assembly of dead body

parts, each resulting in an eloquent monster, each seeming

to threaten the boundaries of the human, each asking hard

questions about the relation of creature and creator. The

student had played mad scientist in a text laboratory where

ethics rules were temporarily suspended. And they realized

that bringing such an essay to life required their substantive

work. That realization—rather than their hideous progeny—

made the experiment a success.

The intellectual labor of assembly includes the work of

editing, which scholars have described as “not merely a

mechanical procedure, but a complex and creative process”

(Dragga and Gong 1989, 9). In their study of editing, Dragga

and Gong detail the extensive “rhetorical repertoire” that

editors must comprehend and employ in diﬀerent situations

(1989, 11). Within this repertoire, “arrangement” probably

best describes what these students defended as their honest

labor, having done all the work on the “organization of a

text, the ordering of information according to appropriate

cognitive patterns” (Dragga and Gong 1989, 12). Scholar-

ship on writing-as-remix sees these thought processes acti-

vated in related ways. Johnson-Eilola and Selber propose the

term “assemblage” and emphasize the critical, rhetorically-

situated thinking it requires: “assemblages are texts built

primarily and explicitly from existing texts to solve a writing

or communication problem in a new context” (2007, 381).

In this light, while one student confessed they were “barely

writing,” they similarly defended their work as conceptual

assembly: “it was more constructing an idea from the many

voices.” Ultimately, as another student concluded, “that

involvement in my thought process turns the use of the gen-

erated text from cheating to not cheating … Doing all that

work … makes me feel like I did more writing rather than

less.” Here, the word “writing” represents a far more com-

plex, engaged activity than the act of putting down words,

and using AI ironically required them to do more of it.

When writing with AI becomes more like the assem-

bly and editing of texts, we change the role of author to

something more like an editor, curator, or mediator. In these

roles, students found themselves right in the middle of “a

relationship between two entities who will attribute agency

to each other”—authors and audiences, whose relationship

seems profoundly unsettled by AI (Miller 2007, 149). Writ-

ing makes its meaning through the terms of that relation-

ship, or what Henrickson calls the “hermeneutic contract”

between authors and readers (2019). When new text tech-

nologies appear, those terms have to be renegotiated, eventu-

ally becoming normalized into the expectations for a given

medium, to the point where we take them for granted. But no

norms yet exist for writing with AI, no shared expectations

for how it credibly makes meaning. This assignment asks

students to renegotiate that hermeneutic contract, shifting

writers and readers into a posthuman context where agency

matters less. Put diﬀerently, dealing with the problems of

expectations and norms makes student reﬂect upon the social

protocols of an emergent text technology, rather than just

seeing it as a tool. Thus, not only does this assignment aim

at debunking AI as an automatized writer, it reframes writ-

ing with AI as a social negotiation in which students have

a stake.

As Rieder puts it, this is the “age of digital persuasion.”

But in expository genres, writing with AI has a persua-

sion problem. How can readers be persuaded to accept its

outputs? Where might AI make credible contributions to

the writing process? What aspects of writing with AI can

we agree to value, or in what contexts? We have already

accepted many aspects of computational writing assistance

into our norms: “the modern textual landscape,” says Hen-

rickson, is “permeated with varied modes of human–com-

puter collaboration” (2019, 7). But text generation contin-

ues to be a sticky problem, and consensus is not coming

soon. My students realized that we were nowhere close to

push-button solutions to writing essays automatically, but

that writing with AI has emerged as an important issue for

us all to grapple with now, to reject with reason, and/or to

reimagine with speciﬁc collaborative practices.

4 Closing

Collectively, my class did not conclude one way or another

that writing with AI was tantamount to plagiarism. They

split down the middle on the question, with all sorts of quali-

ﬁcations about their votes. But this assignment was never

meant to settle a debate. Instead, it immersed students within

the debate to broaden and deepen their perspectives, and

even to raise other questions we had not yet collectively dis-

cussed. As one student perceptively judged: “Overall, the

experiment was able to test multiple ideas … It was interest-

ing to experience these ideas after having read about them.”

The “hands-on” experience let students explore the argu-

ments for themselves, as well as test out new conﬁgurations

For an extensive introduction to assemblage in writing studies,

as well as current examples in pedagogical practice, see Yancey and

McElroy (2017).

AI & SOCIETY

1 3

of writing beyond the potentially limiting concept of humans

vs. AI. Furthermore, the assignment proved useful not only

in reckoning with text-generating AI, but in having students

reﬂect on the values of their own expression. The assign-

ment’s framework of AI-powered plagiarism led students

to assume that they would somehow avoid the intellectual

work of writing. But the opposite happened, reinforcing or

expanding students’ deﬁnition of what that work encom-

passed. As the author Robin Sloan reports of his own experi-

ments writing with machines, “the goal is not to make writ-

ing ‘easier’; it’s to make it harder” (2016). In other words,

such experiments refuse to let us take our ideas about writing

for granted, pushing us to rearticulate or re-envision them.

And, happily, even though this required more work and some

occasional aggravation, students were engaged in ways sur-

passing my expectations.

Whether or not writing with AI becomes cheating is a

much more complicated question than software purveyors

and institutional plagiarism policies would have us believe.

And humanities scholars can play an important role in help-

ing explain and navigate those complexities. For starters,

we should reframe that originating question in a few diﬀer-

ent ways. How do institutionalized deﬁnitions of plagiarism

block the possibilities of writing with computational assis-

tance or even partnership? What kinds of creative or critical

frameworks do plagiarism policies impose or preclude, espe-

cially compared to what we really value for students? How

are these values supported (or not) by the learning activities

that AI seems likely to impact (like writing essays)? How

might we modify these activities or emphasize diﬀerent out-

comes? How do plagiarism regimes or AI platforms diﬀer-

ently aﬀect populations of students? Are there new modes

of creativity, critical reasoning, rhetoric, assemblage, and

expression that computational assistance helps us iden-

tify? How might AI-assisted writing exercises align with

frameworks for teaching AI literacy?

What are the risks

and harms of these technologies that we need to factor in?

And how can we bring diﬀerent disciplinary perspectives

on these questions to bear upon the development and imple-

mentation of AI?

This essay has recontextualized AI-powered plagiarism

with students’ reﬂections alongside examples from schol-

arship in writing studies, rhetoric, book history, media

theory, HCI, and science and technology studies. All of

these ﬁelds—in their work past and present—continue to

oﬀer vital perspectives on whether or how writing with AI

might be embedded in society. Assignments like this may

help amplify for students, too, a call for their participation

in the discourse. My course included students from various

disciplines and, as an English professor, I especially wanted

to embolden students specializing in ﬁelds other than com-

puter science.

But CS students, too, often lack program-

matic opportunities for creative and critical experimenta-

tion. The most glowing review of this technically simplistic

assignment came from a student in computer engineering.

Another CS student chose to pursue internships in AI and

ethics as a result. As Hua and Raley argue, we all have a

stake “in training, evaluating, and collaborating with the

autonomous systems that will continue to speak and write

on our behalf” (2020). Embedding students within these

debates helps show them not only the important issues at

stake, but invites them into the evolving ethical project of

dealing with AI in our world.

Acknowledgements My thanks to the students in the HON 202 semi-

nar “Data and the Human” for their enthusiastic participation in this

experiment. Many thanks also to NC State University colleagues

including Scott Bailey, Zachary Beare, Chris Anson, Helen Burgess,

and David Rieder for sharing their perspectives and recommenda-

tions. Thanks also to peer reviewers for constructive suggestions that

improved this article.

Author contributions Not applicable.

Funding Not applicable.

Availability of data and material The course syllabus and module

assignments are available open access in the MLA CORE repository:

https:// doi. org/ 10. 17613/ 0h18- 5p41.

Code availability Not applicable.

Declarations

Conflict of interest The authors declare that they have no competing

interests.

Additional declarations for articles in life science journals that report

the results of studies involving humans and/or animals Not applicable.

Ethics approval Federal regulations allow speciﬁc categories of human

subjects research to be exempt from continuing IRB review [45 CFR

46.104(d)]. Exemption Category 1 applies to research conducted in

established or commonly accepted educational settings involving nor-

mal educational practices, including research on classroom instruc-

tional strategies.

Consent to participate All enrolled students completed the major

assignments as course requirements.

Consent for publication All students gave their emailed consent to have

their essays included and quoted in my report. Quotes from student

papers have been anonymized in this article.

For speciﬁc examples from those frameworks, see Long and

Magerko (2020), Ng etal. (2021).

As Lauren Goodlad complains, “while there is increasing talk of

making AI ‘ethical,’ ‘democratic,’ and ‘human-centered,’ scholars

in the humanities seldom shape these discussions” (Goodlad and

Dimock 2021, 317).

AI & SOCIETY

1 3

References

Anson CM (2006) Can’t touch this: reﬂections on the servitude of

computers as readers. In: Patricia FE, Richard H (eds) Machine

scoring of student essays: truth and consequences. Utah State

University Press, Logan, pp 38–56

Anson CM, Perelman L (2017) Machines can evaluate writing well.

In: Ball CE, Loewe DM (eds) Bad ideas about writing. Digital

Publishing Institute/West Virginia University Libraries, Mor-

gantown, pp 278–286

Bailey J (2020) How teachers can prepare for AI-based writing. Tur-

nitin. https:// www. turni tin. com/ blog/ how- teac h ers- can- prepa r e-

for- ai- based- writi ng. May 21, 2020.

Battista A, Katherine B, Marybeth M (2020) Data literacy in media

studies: strategies for collaborative teaching of critical data

analysis and visualization. J Interact Technol Pedag. https://

jitp. commo ns. gc. cuny. edu/ data- liter acy- in- media- studi es- strat

egies- for- colla borat ive- teach ing- of- criti cal- data- analy sis- and-

visua lizat ion/.

Bender EM, Timnit G, Amy M-M, Schmargaret S (2021) On the dan-

gers of stochastic parrots: can language models be too big?” In:

FAccT ’21: Proceedings of the 2021 ACM Conference on Fair-

ness, Accountability, and Transparency, 1–14. Virtual Event,

Canada: ACM. https:// doi. org/ 10. 1145/ 34421 88. 34459 22.

Bommasani R, Drew AH, Ehsan A, Russ A, Simran A, Sydney VA,

Michael SB, etal (2021) On the opportunities and risks of foun-

dation models. http:// arxiv. org/ abs/ 2108. 07258.

Boyd D, Crawford K (2012) Critical questions for big data: provoca-

tions for a cultural, technological, and scholarly phenomenon.

Inform Commun Soc 15(5):662–679. https:// doi. org/ 10. 1080/

13691 18X. 2012. 678878

Boyle C (2016) Writing and rhetoric and/as posthuman practice. Coll

Engl 78(6):532–554

Branwen G (2020) GPT-3 creative ﬁction. Gwern.Net. 2020. https://

www. gwern. net/ GPT-3.

Broussard M (2018) Artiﬁcial unintelligence: how computers mis-

understand the world. MIT Press, Cambridge

Crawford K (2021) Atlas of AI: power, politics, and the planetary costs

of artiﬁcial intelligence. Yale University Press, New Haven

D’Ignazio C, Klein L (2020) Data feminism. MIT Press, Cambridge

Dragga S, Gong G (1989) Editing: the design of rhetoric. Routledge,

London

Elkins K, Jon C (2020) Can GPT-3 pass a Writer’s turing test? J Cult

Anal. https:// doi. org/ 10. 22148/ 001c. 17212

Flannery KT (1991) Composing and the question of agency. Coll

Engl 53(6):701–713. https:// doi. org/ 10. 2307/ 377895

Goodlad LME, Dimock WC (2021) AI and the human. PMLA

136(2):317–319. https:// doi. org/ 10. 1632/ S0030 81292 10000 79

Haraway DJ (2016) A cyborg manifesto: science, technology, and

socialist-feminism in the late twentieth century. In: Manifestly

Haraway. University of Minnesota Presshttps:// doi. org/ 10. 5749/

minne sota/ 97808 16650 477. 003. 0001

Hayles NK (2012) How we think: digital media and contemporary

technogenesis. The University of Chicago Press, Chicago London

Henrickson L (2019) Towards a new sociology of the text: the her-

meneutics of algorithmic authorship. PhD Dissertation, Lough-

borough University.

Henrickson L (2021) Reading computer-generated texts. Camb Univ

Press. https:// doi. org/ 10. 1017/ 97811 08906 463

Hern A (2019) New AI fake text generator may be too dangerous to

release, say creators. The Guardian. https:// www. thegu ardian.

com/ techn ology/ 2019/ feb/ 14/ elon- musk- backed- ai- writes- convi

ncing- news- ﬁcti on. February 14, 2019

Howard RM (1992) A plagiarism pentimento. J Teach Writ

11(2):233–245

Howard RM (1995) Plagiarisms, authorships, and the academic

death penalty. Coll Engl 57(7):788–806. https:// doi. org/ 10.

2307/ 378403

Hua M, Rita R (2020) Playing with unicorns: AI dungeon and citizen

NLP. Digit Hum Quart 14(4). http:// www. digit alhum aniti es. org/

dhq/ vol/ 14/4/ 000533/ 000533. html.

Johnson-Eilola J, Selber SA (2007) Plagiarism, originality, assemblage.

Comput Compos 24(4):375–403. https:// doi. org/ 10. 1016/j. compc

om. 2007. 08. 003

Kreps S, Miles M (2020) Not your father’s bots. https:// www. forei

gnaﬀ airs. com/ artic les/ 2019- 08- 02/ not- your- fathe rs- bots. April

16, 2020.

Lang A, Quinn D, Annie KL (2021) The ghost in Anouk’s Laptop.

The Data-Sitters Club (blog). https:// datas itter sclub. github. io/ site/

dsc9. February 17, 2021

Laquintano T, Annette V (2017) How automated writing systems aﬀect

the circulation of political information online. Lit Compos Stud

5(2):43–62. https:// doi. org/ 10. 21623/1. 5.2.4

Lea R (2020) If a novel was good, would you care if it was created by

artiﬁcial intelligence?. The Guardian. http:// www. t hegu ardian.

com/ comme ntisf ree/ 2020/ jan/ 27/ artiﬁ cial- intel ligen ce- compu

ter- novels- ﬁcti on- write- books. January 27, 2020

Long D, Brian M (2020) What is AI literacy? Competencies and design

considerations. In: Proceedings of the 2020 CHI Conference on

Human Factors in Computing Systems, 1–16. CHI ’20. Associa-

tion for Computing Machinery, New York https:// doi. org/ 10. 1145/

33138 31. 33767 27.

Manjavacas E, Folgert K, Ben B, Mike K (2017) Synthetic literature:

writing science ﬁction in a co-creative process. In: Proceedings of

the Workshop on Computational Creativity in natural Language

Generation (CC-NLG 2017), 29–37. Association for Computa-

tional Linguistics, Santiago de Compostela. https:// doi. org/ 10.

18653/ v1/ W17- 3904.

Marcus G, Ernest D (2020) GPT-3, bloviator: OpenAI’s language

generator has no idea what it’s talking about—MIT technology

review. In: MIT Technology Review (blog). https:// www. techn

ology review. com/ 2020/ 08/ 22/ 10075 39/ gpt3- openai- langu age-

gener ator- artiﬁ cial- intel ligen ce- ai- opini on/. August 22, 2020

Metz C, Scott B (2019) How AI could be weaponized to spread disin-

formation. The New York Times, June 7, 2019.

Miller S (2004) Rescuing the subject : a critical introduction to rhetoric

and the writer. In: Paperback (ed). Southern Illinois University

Press, Carbondale

Miller CR (2007) What can automation tell us about agency? Rhetor

Soc Q 37(2):137–157. https:// doi. org/ 10. 1080/ 02773 94060 10211

Ng DT, Kit JK, Leung L, Chu KWS, Qiao MS (2021) AI literacy: deﬁ-

nition, teaching, evaluation and ethical issues. Proc Assoc Inform

Sci Technol 58(1):504–509. https:// doi. org/ 10. 1002/ pra2. 487

Otsuki GJ (2020) OK computer: to prevent students cheating with AI

text-generators, we should bring them into the classroom. In: The

Conversation (blog). https:// theco nvers ation. com/ ok- compu ter-

to- preve nt- stude nts- cheat ing- with- ai- text- gener ators- we- should-

bring- them- into- the- class room- 129905. January 23, 2020.

POL 11.35.01—code of student conduct (2020) NC State University.

2020. https:// polic ies. ncsu. edu/ policy/ pol- 11- 35- 01/.

Price M (2002) Beyond ‘gotcha!’: situating plagiarism in policy and

pedagogy. Coll Compos Commun 54(1):88–115. https:// doi. org/

10. 2307/ 15121 03

Radford A, Jeﬀrey W, Dario A, Daniela A, Jack C, Miles B, Ilya S

(2019) Better language models and their implications. OpenAI.

https:// openai. com/ blog/ better- langu age- models/. February 14,

2019.

Raji ID, Morgan KS, Razvan A (2021) You can’t sit with us: exclusion-

ary pedagogy in AI ethics education. In: Proceedings of the 2021

ACM Conference on Fairness, Accountability, and Transparency,

AI & SOCIETY

1 3

515–25. FAccT ’21. Association for Computing Machinery, New

York. https:// doi. org/ 10. 1145/ 34421 88. 34459 14.

Reid A (2020) Synthetic speech and negotiation: AI’s nonhuman rheto-

ric. Enculturation. http:// encul turat ion. net/ synth etic_ speech_ and_

negot iation.

Seabrook J (2019) Can a machine learn to write for the new yorker?

https:// www. newyo rker. com/ magaz ine/ 2019/ 10/ 14/ can-a- machi

ne- learn- to- write- for- the- new- yorker. October 7, 2019.

Sloan R (2016) Writing with the Machine. Robin Sloan (blog). 2016.

https:// www. robin sloan. com/ notes/ writi ng- with- the- machi ne/.

Standage T, Seth S (2018) The box that AI lives. In: The Secret History

of the Future. http:// www. slate. com/ artic les/ podca sts/ secret_ histo

ry_ of_ the_ future/ 2018/ 09/a_ 200_ year_ old_ chess_ playi ng_ robot_

expla ins_ the_ inter net. html. Accessed September 28, 2018.

Underwood T (2021) Science ﬁction hasn’t prepared us to imagine

machine learning. In: The Stone and the Shell (blog). https:// tedun

derwo od. com/ 2021/ 02/ 02/ why- sf- hasnt- prepa red- us- to- imagi ne-

machi ne- learn ing/. February 2, 2021

Yancey KB, Stephen M (eds) (2017) Assembling composition. In:

CCCC studies in writing and rhetoric. In: Conference on Col-

lege Composition and Communication of the National Council

of Teachers of English, Urbana

Publisher's Note Springer Nature remains neutral with regard to

jurisdictional claims in published maps and institutional aﬃliations.