Generative Artificial Intelligence and Writing Assignments

This July, Harvard undergraduate Maya Bodnick published a piece in The Chronicle of Higher Education
under the splashy headline “GPT-4 Can Already Pass Freshman Year at Harvard.” In it, Bodnick chronicled
how she had fed the essay prompts from a number of popular humanities and social science
courses—including EC 10, GOV 1295, GENED 1033, and a First Year Seminar on Proust—into GPT-4 and
submitted the A.I.-generated papers verbatim to their respective instructors, all of whom apparently found
them indistinguishable from A- and B-level work produced by real Harvard undergraduates. Though her
article was more stunt than study, arguably Bodnick succeeded in proving at least the most moderate form
of her argument: i.e., that Harvard instructors teaching writing-based humanities and social science courses
“need to adapt to their students’ new reality—fast.”

But how? It’s far from clear what it means in practice to “adapt” to the rise of generative A.I., and much of
the advice that’s been made available—including Bodnick’s (viz. that instructors move most, if not all, of
their assessment into the classroom, swapping all of their paper assignments for modalities akin to blue
book exams)—is unrealistic, pedagogically dubious, or both. In this document, we try to provide you with
some guidance about how you might revise or rethink your writing assignments to account for the
capacities and affordances of generative A.I., with the understanding that individual instructors are best
positioned to think through the details of exactly how this guidance can be best applied to their own
courses.

If you would like to talk through your plans or get feedback on a proposed assessment, please email
[email protected] to book an appointment with a member of our senior staff.

1. Three High-Level Principles

In the next section of this document, below, we offer specific guidance and examples as to how you might
adapt some of the most commonly-assigned genres of writing assignments to preserve the learning
objectives that you likely had in mind when first developing them. First, however, we suggest three
near-”universal” approaches that any instructor assigning writing may find useful in discouraging students
from outsourcing their thinking and writing to A.I. in ways that contravene the purposes of your assignments.

1. Talk directly and specifically with students about how your assignments are meant to work. Our
students are not, by and large, looking for opportunities to cheat or take shortcuts. The vast
majority, in fact, are just as concerned to determine the ethical and responsible use of A.I. as are
their instructors. The primary challenge posed by generative A.I. is not that, in making cheating easy
it will, therefore, make it rampant, but rather that its utility will blur the lines for even our most
scrupulous students between “seeking help” or “brainstorming ideas,” on the one hand, and
“soliciting an unacceptable degree of assistance,” on the other. Talking to students about what you
want them to learn, how completing your assignments on their own will help them practice and
learn it, and whether or not employing various kinds of assistance from A.I. is consonant with those
goals, will go a long way toward narrowing the problematic use cases you might be worried about.

2. Disaggregate process from product, and render it visible. Even prior to the advent of generative
A.I., most of the writing that students did in our courses was, unfortunately, a black box—instructors

1
 distributed prompts, and a few weeks later, students submitted final products, with little to no
evidence of how they arrived at their conclusions. Now more than ever, we would encourage
instructors to open up that black box by asking students to share early stages of their research and
writing, in the form of preliminary assignments like project proposals, lists of analytical questions,
annotated bibliographies, brief source analysis exercises, draft introductions, etc. Asking students
to share their work in progress makes it considerably harder, not to mention less appealing, for
students to outsource their thinking and writing to a large language model, as it would require them
to forge, convincingly, not one but multiple phases of thinking and drafting.
○ This approach may come with a bonus: Whatever it costs in labor to review so many
preliminary submissions—e.g. In the sense that you may not be able to assign as many final
products—will probably be more than compensated by the improved coaching that students
receive on their preliminary work.

3. Create opportunities for students to reflect on/talk about their work. So long as students imagine
that they are submitting their final written work to a single reader (i.e. the instructor), and that said
reader will never ask them to elaborate on, defend, or recapitulate their ideas in further
conversation, leaning on generative A.I. might seem like a relatively safe (even victimless)
indiscretion. If, however, students realize that they may have many readers—and, moreover, that
those readers will ask them many questions about their writing (including how they arrived at their
conclusions, how they would respond to various counterarguments, what they would have included
had they had more space or time, … )—the value proposition of outsourcing all of those decisions to
a large language model that won’t be able to help them respond to their readers in the moment
becomes much less appealing. Students in courses that employ peer feedback opportunities (like
Expos, and many department tutorials) consistently report that reading and interrogating other
students’ writing is one of the most rewarding parts of a Harvard education. This seems like an
ideal time to lean into that feedback, and ask students to think of the writing that they submit to
their courses as the first, rather than the last, word on their thinking.

2. Specific Approaches Tailored to your Objectives

Jump to … Response paper | Single source analysis | Article summary | Annotated bibliography |
Research paper | Final exam | Inhabit a voice | Creative writing

Response paper

Objective Identify a key issue in a reading

How to Use a large language model (LLM, like ChatGPT) to develop the ideas that emerge
incorporate A.I. from the student’s initial insight. In the sense that this kind of response paper is often
used to gauge students’ intuitions about a reading or measure comprehension,
ChatGPT’s role might be to put that intuition into dialogue before class discussion. For
example, asking students to list the 3–5 main issues of a reading and rank them in
importance before asking an AI to do the same could allow for both authentic
“prediction” by the student along with a corroborating or competing interpretive
response that could be discussed further in class. Students could perhaps ask the AI
what kinds of evidence or what kinds of scholars typically address these issues
and/or what some of the pressing debates around those issues are or have been.

2
Other ways to In-class, small group discussion. Have students create the same ranked list of 3–5
assess issues and then work in small groups to compare their responses. This kind of
small-group work requires facilitation by instructors to make sure students are on the
right track (in the same way that facilitating input from an AI would be necessary).

Scaffolding Students need to understand how the “prediction effect” works, namely, that there’s
required inherent value to guessing “the answer” to a problem on one’s own before being given
the answer or exploring it too much further through experimentation or research. This
initial, flexible “guess” or hypothesis both keeps us invested in learning and makes us
reflective about where we were at the start of that learning process. If students are
motivated in this way, the role of AI or its absence is less material to the learning
experience than the students’ investment in what that experience is helping them
learn.

Objective Pose an analytical question

How to Perhaps students could develop the question and then pose it to the AI assistant to
incorporate A.I. see if it yields an interesting response.

Other ways to In-class, small group discussion. Have students create a handful of analytical
assess questions and then use a rubric to workshop them together in class. This kind of
small-group work requires facilitation by instructors to make sure students are on the
right track (in the same way that facilitating input from an AI would be necessary).

Scaffolding Framing and models. Students need to understand the role of asking
required questions—certain kinds of questions—in moving from readings and lectures to
developing a thesis and working with evidence. This kind of framing, along with
models of more/less productive analytical questions will help students here, whether
or not an AI is part of the activity.

Objective Prepare for discussion

How to Students might ask an LLM to generate a handful of questions that might provoke
incorporate A.I. insightful academic discussions of a given text, then practice responding to these as
preparation for class.

Other ways to Class discussion. The quality of the class discussion is the best evidence of the
assess students’ preparation, but if one intends to put more emphasis on the quality of
classroom discussion this year, then transparency about evaluation criteria will be
essential. LEARN MORE

Scaffolding If greater emphasis is put on the quality of the discussion itself (rather than on
required documents like response papers), giving the students both practice opportunities and
feedback early and often will be essential.

3
Single source analysis*

Objective Close read a text for language / formal aspects

How to While the ultimate perspective, interpretation or argument should be the student’s
incorporate A.I. own, the student could use the LLM as an assistant in analyzing the text, potentially
by getting the LLM to analyze short passages of the text, then working to critique and
synthesize the ideas the A.I. offers with the student’s own.

Other ways to Annotation. Ask students to show you their close reading skills in action by having
assess them annotate a passage. You can have them do this by hand, by distributing a printed
passage with wide margins, or let students do it online (either in class or at home) by
uploading your readings into a Harvard-supported online annotation tool like Perusall.
LEARN MORE

Scaffolding Students would benefit from practicing close reading in class, with instructor
required feedback, and/or from watching how the instructor performs a close reading.

Objective Close read an object for form / function / conservation

How to In cases where students are developing the ability to describe objects with precision,
incorporate A.I. they may practice doing this “for” an LLM in order to test their skills. Early in the term,
they could describe the object but not name it, then see if LLM can guess the object.

Other ways to Show and tell. An in-class (or video-recorded) “show-and-tell” project, in which
assess students need to give the audience a “guided tour” of the object could be a way to
ensure that they grasp the ways the ideas they’ve developed relate to various features
of the object.

Scaffolding Students would benefit from an opportunity to perform an in-class object-analysis in

Objective Make field / ethnographic observations

How to After students collect their observations, you might ask them to input them into a LLM
incorporate A.I. and prompt it to sift through fragmentary notes in search of patterns on which the
student researcher may wish to reflect.

Other ways to Multimedia observations. Ask students to make and compile their observations in
assess forms that students cannot outsource to an LLM (e.g. annotated video observations,
annotated images, photo essays).

Scaffolding Students would benefit from some low stakes “show and tell” assignments that ask
required them to capture and share raw audiovisual data in class in order to practice analyzing
it.

4
Article summary

Objective Parse another writer’s logic / argument

How to You might ask students to begin by producing their own summary, after which they
incorporate A.I. would be tasked with prompting an LLM to do the same. Students could share both
their own summary and the LLM’s, and spend time in class comparing their results
with the goal of developing intuitions not only about the particular article assigned,
but also more generally about what makes a “summary” more or less insightful /
useful.

Other ways to Concept mapping. Students produce a visual diagram of the ways in which an author’s
assess premises, evidence, and arguments connect and flow. LEARN MORE

Scaffolding Students would benefit from some in-class time devoted to introducing concept
required mapping and its conventions.

Objective Characterize another writer’s use / interpretation of evidence

How to Depending on the nature of the evidence cited in the article, students might be able to
incorporate A.I. ask an LLM to generate feedback on how the author interpreted the evidence (e.g.
“Does [the cited source] support the claim that [the article’s author makes]?”) or to
find out whether other researchers have ever cited the same or similar evidence (and
to what purpose).

Other ways to Scavenger hunt. Rather than asking students to produce a prose summary, have them
assess spend some time in the library drilling down on one (or a handful) of the author’s
footnotes. You could prepare a “scavenger hunt” worksheet that students complete
before class, where they are asked to ferret out information about the material cited in
the footnote—who wrote it? For what purpose? Does the cited passage in fact support
the claim that the article’s author is making about it? Etc.

Scaffolding Students would benefit from a session with a research librarian to orient them to
required library catalogues, databases, etc. that can help them complete the “scavenger hunt.”

Annotated bibliography

Objective Form a picture of the literature within a field

How to In addition to getting help with the low-level formatting technicalities that AI tools
incorporate A.I. have mastered, students may get feedback on the sampling of titles that they have put
together (that they must of course evaluate critically).

Other ways to Source map. Instead of a list of sources, students produce a diagram of the ways that
assess various sources relate to each other and to the primary work the student is analyzing

5
 (some might function as comparisons, some as contestable claims, some as
applicable theories, etc).

Scaffolding Students would benefit from framing that motivates (the otherwise potentially tedious
required seeming) annotated bibliographies, including the elements of summary and citation.
Students will also likely need models for summary and guidance on how to locate and
evaluate the kinds of sources that are best suited to the annotated bibliography at
hand.

Objective Assess the utility of various sources

How to Practice by having the students critique an LLM’s own attempts to create a
incorporate A.I. bibliography of sources in response to an analytical question or topic.

Other ways to In-class lightning round. In some cases you can achieve the same learning objective
assess in a quicker, lower stakes way than through academic writing. For instance, students
could simply rank five sources in terms of utility, then be called on in class to defend
their rankings or grades.

Scaffolding Students would benefit from guidance on how scholars assess and articulate the
required quality/veracity of sources.

Research paper*

Objective Define a research question

How to Ask a LLM (much in the same vein as the “identify a key issue” activity above) to
incorporate A.I. formulate a handful of questions based on a series of themes/key terms and a set of
criteria for the assignment. Then, ask students to evaluate the strengths and/or
feasibility of the responses.

Other ways to Question drafting peer workshop. Have students identify the themes/key terms
assess they’re most interested in from the course and then use the criteria for the assignment
to practice drafting research questions (perhaps by swapping questions with peers
and offering each other feedback on those questions’ alignment with the assignment
prompt).

Scaffolding As with all assignments—but especially ones that include a synthesis of several
required smaller steps and a have longer-term focus on process—going over (and coming back
to) the prompt will help clarify for students the role of any given step and how it
relates to other steps. In the case of developing research ideas, this is likely to
include helping students develop research questions (which is a specific instance of
formulating analytical questions en route to a provisional thesis. In any writing
context, including the use of A.I., it is important to make sure students know that a
provisional thesis isn’t a commitment: it’s a heuristic for framing the early stages of
writing and thinking about structure, and/but it’s likely to change as the process of

6
 inquiry unfolds and evidence + counterevidence get worked through. That is:
throughout the writing process, one’s thesis and one’s evidence/analysis mutually
shape each other.

Objective Identify relevant research material

How to You might ask students to use an LLM to help them generate lists of sources, and
incorporate A.I. take note of where it succeeds and fails in order to see whether this yields any
insights into how well different subfields are represented in the datasets on which the
LLM was trained. You might also encourage students to use an LLM to generate lists
of most frequently cited materials in a subset of literature within a field.

Other ways to You might try a variation on the scavenger hunt described above, asking students to
assess use HOLLIS and other library databases (i.e. rather than A.I.) to assemble a collective
course bibliography of sources during a class field trip to the library.

Scaffolding Students would benefit from a session with a research librarian to orient them to
required library catalogues, databases, etc. that can help them complete the “scavenger hunt.”

Objective Enter an existing conversation in a field

How to After a student has identified the main “schools,” perspectives, or approaches that
incorporate A.I. characterize their field of research, have them ask an LLM to roleplay one (or more) of
them in a conversation. You might ask the student to submit a screen recording of
their conversation.

Other ways to Role play. Students could be assigned to embody the various schools or approaches
assess they have identified in the field, and to perform a live discussion or debate in class
with instructor or peer feedback.
Card game. Building on the “dinner party” framework of putting scholars in
conversation with each other, turn it into a card game: have each writer / reading /
perspective discussed in class on a separate index card. Divide class into small
groups and distribute cards (randomly or have them pick their favorites). The first
player plays a card to start a conversation; the next student plays a card adding a new
perspective or argument to the conversation, using their card/idea to build on the
conversation that the previous card/idea started. This can be expanded by
incorporating students’ own ideas or arguments (in a “they say / I say” framework, as
mentioned below).

Scaffolding Students would benefit from instruction in the key rhetorical moves of academic
required argument, as in the book They Say, I Say, as well as some practice in public speaking
(which can be arranged through the Bok Center).

Final exam*

7
 Objective Synthesize course readings

How to While it’s unlikely that you would invite students to use A.I. during a final exam, they
incorporate A.I. might be encouraged to use it to generate practice essay or short answer prompts
that you could practice in a review session.

Other ways to N/A?

Scaffolding N/A?
required

Objective Apply course concepts to new problems / scenarios

How to See above.

Other ways to N/A?

Scaffolding N/A?
required

Inhabit a voice

Objective Identify and be able to reproduce key stylistic patterns in a primary source or key
conceptual features of a thinker’s perspective

How to After drafting their own piece of writing (e.g. a letter from one historical figure to
incorporate A.I. another), students could upload it to a LLM and ask the LLM to respond (e.g. as the
imaginary letter’s recipient), to which the student might then draft a second response.

Other ways to Briefing book. Students may be reluctant to attempt to inhabit the voice of a person
assess whose identity they do not share; in these circumstances, outsourcing the writing to a
LLM may be particularly tempting. Instead of asking students to inhabit a different
voice, you might ask them to compile a briefing book of the background research that
would go into preparing to inhabit that other voice.

Scaffolding Students would benefit from sessions with librarians and curators, as relevant, as well
required as opportunities to read as many ego documents as possible produced by the kind of
person they are meant to channel.

Objective Cultivate empathy for a figure / perspective distant in time, space, or experience

How to See above.

8
 incorporate A.I.

Other ways to See above.

Scaffolding See above.

Creative writing

Objective Produce a piece of original fiction (or creative nonfiction)

How to Students could ask an LLM to produce “background” materials that flesh out the life
incorporate A.I. and experiences of their fictional characters (e.g. a diary kept by someone matching
the description of their short story’s protagonist). They could then read these for
inspiration when drafting their story.

Other ways to N/A. LLMs remain largely ineffective at producing narrative fiction, and the workshop
assess format of most creative writing courses would make it difficult/infeasible to employ
A.I. illicitly.

Scaffolding N/A.
required

3. A Final Note About Generative A.I. and Equity

Just as we used to assume (incorrectly) that our students were all “digital natives,” and would all, uniformly,
understand how to do things online better than we would, we might be tempted to assume that all of our
students will know more about how to access and use generative A.I. than we could ever teach them. That
may be true of some of our students—but it’s likely not true of all of them, meaning that encouraging them to
use A.I. as part of your graded assignments may introduce unforeseen inequities into your assessment
scheme. Insofar as you want to allow, or even encourage, your students to make use of A.I. to enhance their
written work, you’ll want to make sure that they all have equal access to the most useful platform(s), and a
fair chance to develop the requisite amount of proficiency in what is now called “prompt engineering” (i.e.
the ability to compose queries that will produce the richest/highest quality output from an A.I. tool).

This may mean holding a course “hackathon” where students all receive some instruction in effective use of
A.I. and opportunities to practice with more expert users on hand; it could also mean encouraging students
to go through a free online “prompting course” like this one. At the very least it means including explicit
instruction about how you want students to use generative A.I. in the high level principles articulated at the
start of this document.

9
The figure above presents a typology of common writing assignment sequences. On the far left (in vertical
text), the labels for the y-axis indicate a progression of general learning objectives, moving from “show you
understand an author’s argument” to the more complex operations of “formulating your own argument” and
putting that argument into dialogue with other scholars or different audiences. Closer to the y-axis itself are
the groupings of specific learning objectives that tend to belong to each stage of this progression.

The x-axis breaks down the most common genres of writing assignments, and on the plane of the graph
itself are common kinds of assignments, along with blue and yellow arrows. The blue arrows suggest the
most complex objectives for which a kind of assignment is likely to serve as evidence, while the yellow
arrows indicate the ongoing role of less complex assignments as scaffolding for increasingly complex ones.

The figure is not exhaustive, and different instructors or disciplines will have different intuitions about the
exact order of some elements of the graph. That being said, the figure’s argument is relatively simple:
assignments should be sequenced and scaffolded in alignment with their objectives, and learning
experiences tend to be more positive when the sequencing and scaffolding is both intentional and
transparent.

Considerations for the Use of Generative AI

Two Possible Roles

10
● Role 1. The AI “does” part(s) of an assignment for a student, and the student then evaluates the AI’s
work and gives it feedback (before making use of the output)
● Role 2. The student does part(s) of an assignment, and the AI then evaluates the student’s work
and gives it feedback (before making use of the output). In this case, the AI arguably also “does”
part(s) of the assignment for the instructor.

11