Advice Concerning the Increase in AI-Assisted Writing

Edward Schiappa, Professor of Rhetoric
Nick Montfort, Professor of Digital Media
10 January 2023

[In response to a request, this is an HTML version of a PDF memo for our colleagues at MIT]

There has been a noticeable increase in student use of AI assistance for writing recently. Instructors have expressed concerns and have been seeking guidance on how to deal with systems such as GPT-3, which is the basis for the very recent ChatGPT. The following thoughts on this topic are advisory from the two of us: They have no official standing within even our department, and certainly not within the Institute. Nonetheless, we hope you find them useful.

Newly available systems go well beyond grammar and style checking to produce nontrivial amounts of text. There are potentially positive uses of these systems; for instance, to stimulate thinking or to provide a variety of ideas about how to continue from which students may choose. In some cases, however, the generated text can be used without a great deal of critical thought to constitute almost all of a typical college essay. Our main four suggestions for those teaching a writing subject are as follows:

  1. Explore these technologies yourself and read what has been written about them in peer-reviewed and other publications,
  2. understand how these systems relate to your learning goals,
  3. construct your assignments to align with learning goals and the availability of these systems, and
  4. include an explicit policy regarding AI/LLM assistance in your syllabus.

Exploring AI and LLMs

LLMs (Large Language Models) such as those in the GPT series have many uses, for instance in machine translation and speech recognition, but their main implications for writing education have to do with natural language generation. A language model is a probability distribution over sequences of words; ones that are “large” have been trained on massive corpora of texts. This allows the model to complete many sorts of sentences in cohesive, highly plausible ways that are sometimes semantically correct. An LLM can determine, for instance, that the most probable completion of the word sequence “World War I was triggered by” is “the assassination of Archduke Franz Ferdinand” and can continue from there. While impressive in many ways, these models also have several limitations. We are not currently seeking to provide a detailed critique of LLMs, but advise that instructors read about the capabilities and limitations of AI and LLMs.

To understand more about such systems, it is worth spending some time with those that are freely available. The one attracting the most attention is ChatGPT. The TextSynth Playground also provides access to several free/open-source LLMs, including the formidable GPT-NeoX-20B. ChatGPT uses other AI technologies and is presented in the form of a chatbot, while GPT-NeoX-20B is a pure LLM that allows users to change parameters in addition to providing prompts.

Without providing a full bibliography, there is considerable peer-reviewed literature on LLMs and their implications. We suggest “GPT-3: What’s it good for?” by Robert Dale and “GPT-3: Its Nature, Scope, Limits, and Consequences” by Luciano Floridi & Massimo Chiriatti. These papers are from 2020 and refer to GPT-3; their insights about LLMs remain relevant. Because ChatGPT was released in late November 2022, the peer-reviewed research about it is scant. One recent article, “Collaborating With ChatGPT: Considering the Implications of Generative Artificial Intelligence for Journalism and Media Education,” offers a short human-authored introduction and conclusion, presenting sample text generated by ChatGPT between these.

Understanding the Relationship of AI and LLMs to Learning Goals

The advisability of any technology or writing practice depends on context, including the pedagogical goals of each class.

It may be that the use of a system like ChatGPT is not only acceptable to you but is integrated into the subject, and should be required. One of us taught a course dealing with digital media and writing last semester in which students were assigned to computer-generate a paper using such a freely-available LLM. Students were also assigned to reflect on their experience afterwards, briefly, in their own writing. The group discussed its process and insights in class, learning about the abilities and limitations of these models. The assignment also prompted students to think about human writing in new ways.

There are, however, reasons to question the practice of AI and LLM text generation in college writing courses.

First, if the use of such systems is not agreed upon and acknowledged, the practice is analogous to plagiarism. Students will be presenting writing as their own that they did not produce. To be sure, there are practices of ghost writing and of appropriation writing (including parody) which, despite their similarity to plagiarism, are considered acceptable in particular contexts. But in an educational context, when writing of this sort is not authorized or acknowledged, it does not advance learning goals and makes the evaluation of student achievement difficult or impossible.

Second, and relatedly, current AI and LLM technologies provide assistance that is opaque. Even a grammar checker will explain the grammatical principle that is being violated. A writing instructor should offer much better explanatory help to a student. But current AI systems just provide a continuation of a prompt.

Third, the Institute’s Communication Requirement was created (in part) in response to alumni reporting that writing and speaking skills were essential for their professional success, and that they did not feel their undergraduate education adequately prepared them to be effective communicators.[1] It may be, in the fullness of time, that learning how to use AI/LLM technologies to assist writing will be an important or even essential skill. But we are not at this juncture yet, and the core rhetorical skills involving in written and oral communication—invention, style, grammar, reasoning and argument construction, and research—are ones every MIT undergraduate still needs to learn.

For these reasons, we suggest that you begin by considering what the objectives are for your subject. If you are aiming to help students critique digital media systems or understand the implications of new technologies for education, you may find the use of AI and LLMs not only acceptable but important. If your subject is Communication Intensive, however, an important goal for your course is to develop and enhance your students’ independent writing and speaking ability. For most CI subjects, therefore, the use of AI-assisted writing should be at best carefully considered. It is conceivable at some point that it will become standard procedure to teach most or all students how to write with AI assistance, but in our opinion, we have not reached that point.The cognitive and communicative skills taught in CI subjects require that students do their own writing, at least at the beginning of 2023.

Constructing Assignments in Light of Learning Goals and AI/LLMs

Assigning students to use AI and LLMs is more straightforward, so we focus on the practical steps that can be taken to minimize use when the use of these system does not align with learning goals. In general, the more detailed and specific a writing assignment, the better, as the prose generated by ChatGPT (for example) tends to be fairly generic and plain.

Furthermore, instructors are encouraged to consult MIT’s writing and communication resources to seek specific advice as to how current assignments can be used while minimizing the opportunities for student use of AI assistance. These resources include Writing, Rhetoric, and Professional Communication program, the MIT Writing & Communication Center, and the English Language Studies program It is our understanding that MIT’s Teaching + Learning Lab will be offering advice and resources as well.

Other possible approaches include:

• In-class writing assignments
• Reaction papers that require a personalized response or discussion
• Research papers requiring quotations and evidence from appropriate sources
• Oral presentations based on notes rather than a script
• Assignments requiring a response to current events, e.g., from the past week

The last of these approaches is possible because LLMs are trained using a fixed corpus and there is a cutoff date for the inclusion of documents.

Providing an Explicit Policy

Announcing a policy clearly is important in every case. If your policy involves the prohibition of AI/LLM assistance, we suggest you have an honest and open conversation with students about it. It is appropriate to explain why AI assistance is counter to the pedagogical goals of the subject. Some instructors may want to go into more detail by exploring apps like ChatGPT in class and illustrating to students the relative strengths and weaknesses of AI-generated text.

In any case: What this policy should be depends on the kinds and topics of writing in your subject. If you do not wish your students to use AI assisted writing technologies, you should state so explicitly in your syllabus. If the use of this assistance is allowable within bounds, or even required because students are studying these technologies, that should be stated as well.

In the case of prohibition, you could simply state: “The use of AI software or apps to write or paraphrase text for your paper is not allowed.” Stronger wording could be: “The use of AI software or apps to write or paraphrase text for your paper constitutes plagiarism, as you did not author these words or ideas.”

There are automated methods (such as Turnitin and GPTZero) that can search student papers for AI-generated text and indicate, according to their own models of language, how probable it is that some text was generated rather than human-written. We do not, however, know of any precedent for disciplinary measures (such as failing an assignment) being instituted based on probabilistic evidence from such automated methods.

Conclusion

The use of AI/LLM text generation is here to stay. Those of us involved in writing instruction will need to be thoughtful about how it impacts our pedagogy. We are confident that the Institute’s Subcommittee on the Communication Requirement, the Writing, Rhetoric, and Professional Communication program, the MIT Writing & Communication Center, and the English Language Studies program will provide appropriate resources and counsel when appropriate. We also believe students are here at MIT to learn, and will be willing to follow thoughtfully advanced policies so they can learn to become better communicators. To that end, we hope that what we have offered here will help to open an ongoing conversation.[2]


[1] L. Perelman, “Data Driven Change Is Easy; Assessing and Maintaining It Is the Hard Part,” Across the Disciplines 6.2 (Fall 2009). Available: http://wac.colostate.edu/atd/assessment/perelman.cfm. L. Perelman, “Creating a Communication-Intensive Undergraduate Curriculum in Science and Engineering for the 21st Century: A Case Study in Design and Process.” In Liberal Education in 21st Century Engineering, Eds. H. Luegenbiehl, K. Neeley, and David Ollis. New York: Peter Lang, 2004, pp. 77-94.

[2] Our thanks to Eric Klopfer for conversing with us concerning an earlier draft of this memo.

A Poetry Class for 36,000

December 10, 5:30pm in MIT’s 6-120

Al Filreis

Teaching Modern & Contemporary American Poetry to 36k

Al Filreis has taught his “ModPo” course at Penn for years; in Fall 2012 he
offered a 10-week version of the course online, via Coursera, to more than
36,000 students. The course, as in its previous versions, does not include
lectures, being based instead on discussion – the collaborative close
readings of poems. The course grows out of Filreis’s work at the Kelly
Writers House; he has been Faculty Director of this literary freespace since
its founding in 1995. Filreis is also co-founder of PennSound, the Web’s
main free archive of poetry readings, publisher of Jacket2 magazine, and
producer and host of “PoemTalk,” a podcast/radio series of close readings of
poems. In conversation with Nick Montfort, Filreis will discuss ModPo and
his perspective on writing, teaching, and digital media.

Filreis is Kelly Professor of English and Director of the Center for
Programs in Contemporary Writing at the University of Pennsylvania. He is
the author of Wallace Stevens and the Actual World, Modernism from Right to
Left, Counter-Revolution of the Word: The Conservative Attack on Modernism,
1945-60, and other works. He was chosen as Pennsylvania Professor of the
Year by the Carnegie Foundation in 2000.

Co-sponsored by the SHASS Dean’s Office and the Literature Section.

All Purple Blurb events are free and open to the public. The Purple Blurb
series is supported by the Angus N. MacDonald fund and Writing and
Humanistic Studies.

On Reading

I was asked to discuss reading (and reading education) from my perspective recently. Here’s the reply I gave…

The students I teach now, like other university students I have taught, have the ability to read. They are perfectly able to move their eyes over a page, or a screen, and recognize the typographical symbols as letters that make up words that make up sentences or lines.

The problems they face usually relate to a narrow concept of reading, which includes an unwillingness to read a wider variety of texts. These are not problems that are restricted to well-qualified, well-educated university students who are expert readers. As the networked computer provides tremendous access to writing and transforms our experience of language, all of are asked to rethink and enlarge our reading ability.

One problem with reading too narrowly is the view that reading is only instrumental. People who spend a significant portion of their lives speaking to their friends and family members about nothing – simply because they enjoy company and conversation – will sometimes refuse to believe that reading can a pleasure in and of itself. Reading is too often seen only as a tool, to allow one to follow instructions, determine the ingredients in packaged food, or to learn about some news event or underlying argument.

To try to enlarge a student’s idea of reading, I might present a text that, whether said aloud or imagined to one’s self, communicates almost nothing and is simply beautiful. To take a very conservative example:

Glooms of the live-oaks, beautiful-braided and woven
With intricate shades of the vines that myriad-cloven

  Clamber the forks of the multiform boughs,—
      Emerald twilights,—
      Virginal shy lights,
Wrought of the leaves to allure to the whisper of vows,
When lovers pace timidly down through the green colonnades
  Of the dim sweet woods, of the dear dark woods,
    Of the heavenly woods and glades,
That run to the radiant marginal sand-beach within
      The wide sea-marshes of Glynn;—

This is from an American poem, Sidney Lanier’s The Marshes of Glynn, published in 1878. It shows that it is not necessary to turn to some almost unrecognizable avant-garde poem to see how sound can overflow meaning, providing a pleasure that has almost nothing to do with communication. Lanier had a very musical view of poetry, but this poem does not need to be sung or played by an expert musician. Anyone who can read English can bring it alive.

Another problem with a narrow concept of reading is not understanding the full range of what can be read. Students are comfortable reading pages and feeds, and reading emails and IMs, but it is not always as clear how they might read around in the library to explore as researchers, how they might read an unusual Web site or other complex digital object, or how they might read their daily urban environment. Those of us familiar with art and literature will habitually “view” something in the former category, perhaps not even noticing that it may be legible.

To address this issue, I do turn to an almost unrecognizable avant-garde poem, Steve McCaffery’s Carnival, the first panel. This Canadian concrete poem, typed in a red and black in an amazing configuration of characters, seems to many to be an artwork but was created by a poet and published as poetry. Unlike the most famous Brazilian concrete poems, this is an example of “dirty concrete” that many do not even see, initially, as legible. Students who are able to assume the attitude of readers find that it can be read, however, and that it has an amazing ability to disclose things about reading: Our assumptions in how to trace words and letters in space, our ability to fill in partly-missing and entirely missing letters, the question of how to sound fragments of words and patterns of punctuation marks.


The first page of Carnival, panel one, is from the
Coach House online edition.

When people learn how to speak a language – whether as an infant or later in life – they sometimes simply babble or chat. Everyone would agree that a learner should be allowed to enjoy speaking and listening, to enjoy making and hearing the sounds of a language, in addition to caring about the purposeful uses of that language. I believe it’s the same for reading. Reading is more than just a process of decipherment that provides an information payload. It gives us special access to the pleasures of language and to its complexities. Any sighted person, with or without any English, can look at the first panel of Carnival. But only a reader can both view it and read it, comprehending it as a visual design and as language. If a reader is unwilling to hear The Marshes of Glynn, it could be understood simply as a botanical catalog with some lovers traipsing about here and there. To hear, instead, the play of sound and sense, is to encounter something new, to understand another aspect of words and how they work.

IF in College Education?

Mary Dooms, a middle school teacher in Illinois who has used interactive fiction in her teaching, recently asked me if I knew about any uses of IF in teaching in higher education. That’s a good question.

She had found Utah State’s Voices of Spoon River and Myth Mechanic. I know right off that Jeff Howard has taught The Crying of Lot 49 using IF, and that students read IF and create it as a digital literary practice in two of my classes, Interactive Narrative and The Word Made Digital.

I’ve certainly heard of many other uses of IF in education, and when I can, I’ll begin to collect and list those. Rather than wait, I thought I’d call for links from Post Position/Grand Text Auto readers, since I know there are many classroom deployments of IF that I’m not aware of. If they have well-packaged downloads like the Utah State projects, that’s particularly nice, but I imagine that other information about the classroom use of IF would also be helpful. Any ideas?

Purple Blurb – Digital Writing, Fall 2009

Once again, Purple Blurb offers readings and presentations on digital writing by practitioners of digital writing. All events are at MIT in room 14E-310, Mondays at 6pm. All events are free and open to the public. The Purple Blurb series is supported by the Angus N. MacDonald fund and Writing and Humanistic Studies.

Noah Wardrip-Fruin.

September 14 — Noah Wardrip-Fruin is author of Expressive Processing: Digital Fictions, Computer Games, and Software Studies (MIT Press, 2009), co-creator of Screen (among other works of digital writing), and assistant professor in the Department of Computer Science at the University of California, Santa Cruz.

Mary Flanagan.

November 2 — Mary Flanagan is author of Critical Play: Radical Game Design (MIT Press, 2009), creator of [giantJoystick], and author of [theHouse] (among other digital writing works). She is Sherman Fairchild Distinguished Professor in Digital Humanities at Dartmouth.

D. Fox Harrell.

November 16 — D. Fox Harrell is the creator of the GRIOT system for computational narrative and author of several works in this system, including Loss, Undersea and The Girl
with Skin of Haints and Seraphs.
He is assistant professor of digital media in the School of Literature, Communication, and Culture at the Georgia Institute of Technology.

Marina Bers.

November 30 — Marina Bers is author of Blocks to Robots: Learning with Technology in the Early Childhood Classroom (Teachers College Press, 2007) and creator of the system Zora. She is associate professor in the Department of Child Development and adjunct professor in the Department of Computer Sciences at Tufts University.

Digital Writing and Readings

[As I wrote on netpoetic.com:] Adam Parrish recently taught a class at NYU in the ITP program: Digital Writing with Python. I was very interested to learn about it and to see documentation of the final reading/performance, with some links to students’ blog entries about their projects. Here at MIT, I teach a class called The Word Made Digital in which students do poetry, fiction, and less classifiable writing projects using Python and other systems and languages. And, I know that Daniel Howe has taught the RISD and Brown class Advanced Programming for Digital Art and Literature.

I suspect, though, that these classes that are mainly focused on writing and programming are rather rare – much more rare, I’d bet, than design and art classes that are heavy on programming. It may have something to do with the number of galleries and curated Web sites exhibiting programmed visual art, which seems to me to be much greater than the number of similar edited venues for digital writing that’s driven by code. I’m not sure which way the causality flows. But several of the art-loving among us have some idea that, say, Processing programs can be aesthetic, even though they’re made of code. It’s not as common for literary folks to think of Python, Perl, or other programming languages (whether or not they start with P) as ways of creating literary art.

My sense is that having readings, of the sort that Parrish hosted at the end of his class and of the sort that the Electronic Literature Organization has sponsored and organized over the years, is a useful way to address this gap between literature and the visual arts. (Full-blown festivals, of course, don’t hurt either.) A reading allows writers to show off a program, which may be intricate, and explain how it works. It’s fun for those who are already into digital literature, and an accessible way for other literati to see what computational writing is about and how it bring certain literary qualities into the digital realm – even if it does radically subvert others. And since there aren’t as many official, edited, and well-promoted publication options for computational writers, going to do a reading can be a good way to appear in a context of other writers and reach a public.

I’m trying to do my part here by running a reading series for digital writing, but that’s grist for the next post [of mine on netpoetic.com].

This was posted here on Post Position for the convenience of those of you who subscribe to the feed or visit the site. If you want to leave a comment, please head over to this post on netpoetic.com.

Pythonic Textuality at NYU

I was very interested to learn that Adam Parrish, whose own Interactive Telecommunication Program (ITP) masters project was “New Interfaces for Textual Expression,” is now teaching Digital Writing with Python at NYU’s ITP. The course is concluding; Parrish and his students will mount a final performance on August 5 at 7pm. Parrish eschewed powerful, cryptic Perl for clarity of Python in this course on creating text machines, as I did in putting together The Word Made Digital, which I’ll be teaching again this Fall. His reading list overlaps with mine a bit and includes a nice article on appropriation in writing – I may just rip that right off. I won’t manage to be in New York to hear students read their programs’ output, but I hope the conclusion to the class goes well and that I’ll be able to read and run some things that will give me a sense of the event.