Maybe like some of you, the first time I used GAI, I was completely shocked. I asked it to correct mistakes in a very problematic document and then witnessed it magically rewrite the text in error-free prose. Later, I showed it to tutors of the writing service I direct. As they watched it effortlessly adjust a paper to North American academic standards, one of them blurted out, “We’re screwed.”

Like me, back in those early days, you probably heard predictions that students no longer needed to learn to write. With ease, GAI could compose their class assignments, research articles, and even their dissertation. Meanwhile, teachers were complaining that using GAI was lazy and cheating, leading to stupider students who were unable to gain skills and benefits from the writing process.

But after teaching my first iteration of Strategies for Writing with AI, this is what I learned: Writing with GAI is difficult.

For fun, let’s take a hypothetical journey. Imagine a first-year undergrad who is in a mandatory writing class that they are not very interested in, but would still like to get a decent grade. When the teacher asks them to write an essay comparing different themes of Romeo and Juliet, they feed the teacher’s prompt into ChatGPT, which writes the essay in a few minutes.

Now what?

If the student wants a good grade, they are going to have to read it and make some assessments. Is the style too GAI? If so, how will they change it to convince the teacher they wrote it? What makes GAI writing different from their own? Vocabulary, phrasing, syntax, verbosity? What to change? How much to change? Furthermore, if they really want an A, they’ll have to consider the content. What argument is made? Is it valid? Do they actually support it? Answering these questions entails using higher-level skills—perhaps even more so than if the student just sat down and wrote the essay themselves.

Now consider a graduate student writing a high-stakes article about their original research that they intend to submit to a top-tier journal. They can’t just have GAI write the article for them and send it off to the editor. The stakes are too high. Instead, they may use GAI to help them with steps in the process. They could use something like Copilot to help them find sources, suggest topic sentences, even write an early draft based on an outline. Later, when at the editing and proofing stages, they may ask Copilot to check for errors, rephrase sentences, even assess the logical order.

All of these tasks take skill. The student has to recognize what could be done and then decide what should be done. If they use GAI to find references, they’ll need to check each one to make sure they are accurate. If they ask for content ideas, they need to consider audience and purpose to determine what is appropriate. If they want GAI to write an early draft, they will need to be familiar with the writing process to understand what they are doing and what editing steps will follow. Even proofreading is not a straightforward task. Having GAI rewrite to produce error-free text is risky. It may inadvertently make mistakes, or the student’s voice might be lost. The student must carefully assess the GAI-rewritten text and decide what adjustments should be made. This task could be done with intuition only, but intuition only goes so far. At some point, they are going to need some advanced rhetorical knowledge, such as when to use guiding language—and when to avoid it. Alternatively, the student could ask GAI to list potential mistakes and then go through the list, only changing what they deem appropriate. This too takes advanced skills. For example, what is an error? When is a mistake actually just style? What needs to be changed, and what is just part of a student’s voice?

In all of these tasks, the student must come up with suitable prompts. Asking Claude, or any chatbot, to “Make my writing better” produces much different results than “Correct my mistakes.” Prompts even work differently for different users. GAI learns what a user wants and adapts to their particular preferences. So, just giving a student a list of prompts is not enough. The student must teach GAI what they want. This can involve a lot of back and forth, which again takes skills.

In sum, to effectively use GAI for writing, students must have some rhetorical knowledge to know what to do and prompt engineering skills to get a chatbot to do it. This knowledge must be acquired, and the skills must be honed.

Using GAI to write is different than traditional methods, but it’s neither cheating nor being lazy.

Now, consider the future of writing. Students are already using GAI on a daily basis as an intricate part of their writing process. New generations are going to grow up only knowing how to write with GAI.

Instead of becoming stupider or worse writers, they will become more efficient and effective, able to write faster at a more consistently higher quality.

GAI has raised the bar.

Any kind of error will be unforgivable. There are just too many resources to address them. Even problems with logical skips or flawed structure may also become inexcusable. In short, more will be expected of students.

If more is expected from students, instead of making writing easier, it’s making it more difficult.

This blog entry has two purposes: The first is to show some cool things students can do with GAI. The second is to bring up the depressing truth that GAI is one more way we divide the world by wealth.

GAI is really good at mundane tasks, ones that humans might consider boring and tedious. Consequently, it can quickly answer simple questions, which can lead to a deeper understanding of text.

In class, I gave students a genre analysis handout, which mostly consists of a list of questions that can be used to explore a paper. One example is “What is the average length of sentences in a paper? What is the longest sentence and why is it so long? What is the shortest sentence and why is it so short?”

Knowing a text’s mean sentence length is very useful. Superficially, it quickly answers how long sentences usually are in a particular genre. But it can also lead to a deeper rhetorical understanding. One of my students noted that in the text he was analyzing, the beginning and end of sections tended to have shorter summarizing sentences, while the middle sections had longer ones, containing lots of details. He concluded that the shorter sentences were for skimming readers, who might only look over the first and last parts of sections, while the middle, longer ones explained details for readers very interested in the subject.

But there remains a problem: Calculating the length of sentences is tedious and pretty much a waste of time. Fortunately, GAI can do the exercise in minutes with a simple prompt such as, “I will upload a paper. Tell me how many sentences there are. List the average length of the sentences, the shortest sentence, and the longest sentence.” Voila! All the tedious work is done for you, and you can spend valuable time exploring patterns.

Well, almost. Unfortunately, GAI has trouble figuring out what a sentence is, and you might need to train it first.

I had this problem when I was trying to figure out what was going on with the ends of paragraphs. Like many of my students, at one time I was taught that paragraphs should end with a link to the next one. But when writing, I found this link often repetitive and unnecessary. So, I decided to see how paragraphs really end. I told GAI, “I have uploaded my text. List the last sentence of each paragraph in my text. Tersely, describe the main purpose of the last sentence in 1 of 3 ways: (1) it summarizes the paragraph, (2) it links to the next paragraph, or (3) something else.”

This is where the classism part comes in.

I have access to our university’s subscription to Copilot, Claude, and Microsoft’s Azure (a version of ChatGPT). Also, I subscribe to ChatGPT, which I pay for with my research funds. I started my exercise by giving my prompt to my university subscription to Copilot. It appeared to follow my instructions. However, when I investigated, it was obvious it wasn’t always sure what a paragraph is. Four times it identified the end of a paragraph when it was actually in the middle.

Azure did worse. It was only able to identify seven of the 23 paragraphs. Claude completely failed, stating it couldn’t read the Microsoft Word file I uploaded (though it had no problems in the past).

Then I tried my subscription to ChatGPT 4.o., it recognized all the paragraphs, indicating that three ended with links to the next paragraph, 19 ended with summaries, and one was something else. (It was the conclusion.)

One major difference among the chatbots is that I’ve trained my subscription of ChatGPT and now it knows me. It’s learned what I like (e.g., always be terse) and don’t like (e.g., don’t rewrite anything ever unless I specifically ask you to). In fact, I’ve taught it what a paragraph is*.

The other chatbots were set up by the university for optimal privacy, which means they don’t save submitted text to their repositories. The folks in IT told me that this set up restricts how much they can learn, and thus they start each session with more of a clean slate. Though I could probably teach them what a paragraph is and how to answer my prompt, to some degree, I would have to do this every time I use them.

A list of caveats: Maybe I’m missing something—some setting that will get university subscribed chatbots to remember stuff. Or maybe IT will figure out how to customize these subscriptions to know me better. Also, I have not experimented with the free versions. Basically, I have already become an elitist GAI snob, complaining that the Ivy League university I work for hasn’t addressed all my idiosyncratic needs.

But in this exercise and many others, I, and my students, have come to the same conclusion: To get the best results, you’re gonna have to pay.

I don’t imagine this will get much better. Companies that make GAI need to justify subscription fees. It will probably just get worse.

In competitive research environments, GAI is just another dividing line between the haves and have nots.

*To teach ChatGPT 4.o what a paragraph is, I counted the paragraphs in a research article, and then uploaded the article. I asked ChatGPT how many paragraphs it contained, and it answered wrongly. I asked it to list the first three words of each paragraph, checked for accuracy, and negotiated any errors. I then told it how many paragraphs there were and asked it to count the paragraphs again. I repeated this a few times, and it seemed learn what paragraphs are.

Note: You may have noticed I’ve changed from AI to GAI. My friend Jane Freeman pointed out that the two are not the same, and I should stop pretending they are.

AI is really good at making things average. It can take someone’s written idiosyncrasies and standardize them to create a bland but very acceptable voice (i.e., style), one that’s very useful for students who use English as an additional language (EAL).

EAL students often have written accents. Just as we speak with an accent, we also write with one. For EAL students, this might mean using odd phrasing/vocabulary, missing an article (i.e., a, the), or using an inappropriate preposition (e.g., give in, give up, give over). Now, it would be odd to insist someone speak with a particular accent, but it is commonplace to insist students write with one, specifically, a North American academic one.

If EAL students don’t adopt a North American academic accent, their writing can be judged substandard. I’ve seen many EAL students’ papers with editor comments such as, “Clean up the language” or even the dreaded, “Have this checked by a native speaker.”

A few years ago, students told me that overseas their departments would spend up to $1,500 to have an academic journal article checked by a native speaker.

Then, AI turned everything on its head.

With a simple prompt (e.g., “Correct my mistakes”), AI can rewrite a paper, adjusting accents to North American academic standards.

The problem is that by eliminating accents, it also eliminates a student’s voice, leaving them with an acceptable but very bland piece of writing.

In class, we discussed written accents, and students generally agreed that they did not want to eviscerate theirs. Accents, after all, are part of our personalities. Consider if you had to change your spoken accent and talk like someone from a different part of your home country. You could probably do it, but you’d feel stupid, like you weren’t yourself.

The question is how to ensure that your written accent is adjusted enough to be acceptable, but not so much that your voice is eliminated.

In class, one student had the idea of having AI describe her written accent. Once she saw the defining characteristics, she thought she could learn to control it. 

Based on her idea, I made this lesson plan, which we all followed: 

1. Freewrite a series of emails inviting various people to a dinner party. The idea is to write in your own voice. Turn your font to the color white, so there is no way you can make changes and correct errors.

2. Upload your emails to AI and give it this prompt: “Based on my writing, can you tell what my first language is?”

3. Then investigate further: “What characteristics indicate my first language?”

Nine students came that day, and AI identified their four native tongues correctly, though for Chinese speakers it deduced an “east Asian language.” Based on my use of colloquial language (e.g., “hey, dude,” “how’s it going?”), ChatGPT 4.o identified me as a native speaker of English who has far too many typos.

For defining first-language characteristics, AI identified common errors, phrasing that was most like translated from a first language, style—such as overly polite, and features (e.g., no contractions). All of this was useful to know, but nothing earthshattering.

Then came the creepy part.

The next step was to ask if AI could use the freewritten emails to compose a new message with the same voice: “Use my style to write an email inviting my friend to the same dinner party. His name is George. I want to know if he’ll bring a dessert.”

AI did disturbingly well.

It produced a message that I would have no problem signing.

Then, for fun, I asked it to write the same email with a Chinese accent. The results were cringy: Broken English with missing verbs and articles, odd phrasing, and nonconventional vocabulary. The class and I agreed it was way too close to being racist.

So, I asked it to write it again with a Chinese accent, but no mistakes.

It worked. Really well.

We concluded that here’s what could happen:

1. An EAL student uploads their unedited writing and asks AI to “remember it”.

2. Then the student uploads a detailed outline and asks AI to write a paper, retaining the author’s voice (minus the mistakes).

3. The student reads through the paper, checking for mistakes and ensuring they feel comfortable with the voice.

For a student who has brilliant ideas but not so great English, readers are suddenly given flawless prose, and the student has saved many hours, if not days, of work.

But then we asked, “Is this okay?”

None of us were sure. Some of us felt the student would not learn English well. However, for the student, English is the medium, not the focus. We also got this feeling that the student was cheating.

I ended the class by pointing out that even if we wouldn’t write this way, others already are. I know because I’ve interviewed some of them for my research.

Is this where academic writing is headed? Without a doubt, for some kinds of writing (e.g., emails, grant proposals, lab reports), it’s already happening. But for dissertations, research articles, statements of purpose? Hm…

My fourth blog ended with “Next week: What happens if AI rewrites text for you? Then what? Honestly, I’m not sure, but we’re going to find out.”

But in the proceeding blog, I didn’t address the issue.

We experimented in class, but honestly, our conclusions were underwhelming. I thought I’d just skip the subject and move on to something more successful. However, a realization woke me up in the middle of the night: Research is not about success. It’s about reporting findings—both good and not so good.

So, in the true spirit of research, here’s what we did and found. 

My students all use English as an additional language. Sometimes, their writing contains odd phrasing or unconventional syntax that a primary speaker of English would not include (e.g., missing articles a and the, inappropriate prepositions, verb-subject-object syntax). Scholars call this a written accent—just as we speak with an accent, we write with one. Unfortunately, North American academia has a very low tolerance for written deviations, particularly when it comes to published work. Many of my students have had their research articles sent back with instructions such as, “Have a native speaker proofread.” In fact, students have told me that they’ve spent $1,000 or more to have their written accents adjusted.

Then came AI, and all that was instantly changed. Just submit your text and a simple prompt, such as, “Improve my writing,” and in minutes, AI flawlessly adjusts a written accent. Furthermore, it does it for free and without judgment.

The question is what’s next?  

No serious graduate student would blindly accept AI’s revisions. There has to be some assessment to make sure the meaning is the same, but also that the tone or voice is appropriate.

So in class, this is what we did:

1. Freewrite for 15 minutes an introduction to a subject you are very familiar with. Do not correct any errors.

2. Ask AI to improve your text.

3. Now what? Come up with strategies for continuing.

To address the third point, some students had AI list changes it made, while others had it bold them. Then students scrutinized each one. The process was time-consuming and often not very satisfying.

Changes AI made were often more stylistic choices. Sometimes they were random, such as replacing phrases or words. And sometimes they were just plain wrong. Even when students asked AI to focus only on discrepancies caused by first language interference, the results were unclear.

Granted, the process of scrutinizing AI’s changes does raise linguistic awareness and could lead to better writing. But it’s very inefficient and not practical for long papers.

A pragmatic approach is to have AI revise a text and then read through it to make sure there are no inaccuracies. I feel like this is what a lot of students are doing.

The problem is that AI doesn’t just address errors; it alters the tone or voice of a paper to the point that it no longer feels like the author wrote it. How do you make sure your voice is maintained? That’s what we’re working with this afternoon, and, in the spirit of scientific research, I promise to report results—good or not so good.

Next: What if AI could write with your voice—something as intriguing as it is creepy.

In class, one approach I’ve been using is to introduce a writing concept, have students work with it, and then we all see what AI can do. A few weeks ago, we went over rhetorical moves, which focus on purpose, not content. (See blog 04.) Then last week, we went over guiding language, which is any language that guides the reader through the text. Examples include “moreover,” “see figure 3,” and “in sum.” (see Hyland, 2005).

Before class, I experimented with AI, and quickly found that its definition of guiding language was not the same as mine. To get us on the same page, I wrote this prompt:  

I define “guiding language” as language that does the following: (1) shows relationships (e.g., “however,” “so”, “part of”), (2) begins a new rhetorical move (e.g., “past studies show”),  (3) indicates order (e.g., “first”, “next”), (4) gives direction (“see Figure 1”), (5) summarizes (e.g., titles, headings, topic sentences), and (6) maps structure (e.g., “in this paper I will first…then…”)

I gave the prompt to students and asked them to see what additional prompts they could come up with.

One student constructed a terrific prompt and agreed to let me share it on this blog. His prompt creates a table that lists rhetorical moves and the language that introduces them. Try it out. It’s exciting to watch the table magically appear.  

For my paper, generate a table. The first column should list all rhetorical moves. The second column should list guiding language that introduces a rhetorical move. The third column should explain the purpose of the rhetorical move. The fourth column should state the sentence number that the guiding language comes from (e.g., sentence 4). The fifth column assesses if the sentence delivers what it's supposed to deliver.

We tried it in class, and we all loved it.

But I felt like we needed to pause and ask, “Yeah, it’s really cool, but is it worth it?”  

This is a question we have to keep asking. AI can be shockingly amazing, but sometimes it’s much more flash than substance. (You ever try to make a podcast with NotebookLM?)

To answer the question, I go back to our first day of class, when we decided we needed to use AI not just to make better papers but to become better writers. Does the prompt do either?

Yes and no. First, yes: It does raise awareness of guiding language and rhetorical moves, which can lead to more mastery of writing. Second, if the prompt is used on one’s own paper and then on a published, reputable paper, the two can be compared and contrasted to make similarities and differences salient. Finally, using the prompt on several published, reputable papers can help the reader identify common acceptable patterns.

And now no: How often is a writer in a situation when the advantages are actually worth the time invested? For graduate students writing research papers, it makes sense. But other writers need to answer according to their own situations.

In our class, the prompt worked well to help identify text features and stimulate conversation.  

However, I think while we’re being dazzled by AI, we need to always ask if it’s making us better papers and better writers. In sum, is it worth it?

References

Hyland, K. (2005). Stance and engagement: A model of interaction in academic discourse. Discourse Studies, 7(2), 173-192.

Starting to write in a new genre can be challenging, particularly when trying to determine a suitable structure. How are ideas organized to make a coherent narrative, particularly with research articles, which often have common structures.

What if a writer could get some authentic and trustworthy examples of structure that could be used as options or ideas for how they might structure their own paper?

Such examples can be created by making an outline of a successful paper that comes from the same genre (e.g., a published article in a target journal). Since this outline is created after a text is finished, it’s called a reverse outline. (See my previous blog.) Essentially, it reduces a paper to its main ideas, which can be analyzed and studied in a small, manageable space.

Here’s an example of a reverse outline for the introduction of a research paper. The “P” stands for paragraph:

P1: the background of Industry 4.0

P2: the development and advantages of Cobots

P3: The status of Vietnamese garment manufacturing and Cobots in Vietnam

P4: The integration of Industry 5.0

P5: The technical challenges in the implementation of Cobots

P6: The labor challenges in the implementation of Cobots

P7: The cost challenges in the implementation of Cobots

P8: The gap in the literature review of Cobots

P9: The purpose of study

The problem with using this reverse outline is that it’s focused on content and not easily transferred to a different study.

What if instead the reverse outline was focused on purpose, such as P9 above: “The purpose of this study”? Such a point is called a rhetorical move. When the purpose changes, a new rhetorical move is created.

Here’s an example of a rhetorical-move, paragraph-level reverse outline of a research article introduction:

P1: broad view of subject and background

P2: problem (gap) of current studies

P3: methods used

P4: findings overview

This one is made on the paragraph level, but it could also just be a series of rhetorical moves, which could be anywhere from a phrase to multiple paragraphs.

If the writer has several of these rhetorical-move reverse outlines, they could compare structures and get a better understanding of acceptable options in a genre.

The problem is that making a rhetorical-move reverse outline can be a long, arduous process.

Fortunately, AI can really speed things up.

AI can make really good rhetorical-move reverse outlines, but the AI systems I’ve used (ChatGPT 4.o, Copilot, Claude) tend to start by including too much content, and I’ve had to fight them to focus on purpose.

After some experimenting, my students and I came up with this prompt to make a paragraph-level, rhetorical-move reverse outline of an introduction:

I will upload an introduction, Make a reverse outline of it. Summarize each of the [number of] paragraphs. Be terse. Focus on rhetorical moves, not content. Avoid including content. Be general. For example, you might write “introduces topic”, “narrows to specific topic,” or "summarizes findings”. Here is my introduction:

Some back and forth still might be necessary, but eventually you’ll get there.  

IMPORTANT: When submitting others’ texts to AI, it is crucial to make sure that the system you are using does not add the texts to its database. Otherwise, part of that text could be given to other users without citation. To keep files out of ChatGPT 4.o’s database, go to settings, select “data controls”, and turn off the feature called “improve the model for everyone”. I’m not sure how to change this setting in other systems. Anyone know how?

Once you train AI to make rhetorical-move reverse outlines, you might try to make multiple ones at once and even compare and contrast them.

Here’s a prompt that’s given pretty decent results:

(1) I am going to attach five research articles. (2) Make a rhetorical-move reverse outline for each of the five [name of section, e.g., abstracts, introductions]. (3) Now note similarities and differences.

This prompt is done in three stages, which sometimes has better results than trying to do everything all at once. The files are attached, which seems to work better for larger files. Note “paragraphs” is not stated, so AI will interpret moves no matter how short or long they are.

In minutes, a student can see patterns to gain an understanding of a genre and options they might use.

Next week: What happens if AI rewrites text for you? Then what? Honestly, I’m not sure, but we’re going to find out.

I first learned about reverse outlines from Rachael Cayley’s blog, Explorations of Style, where she explains, “Simply stated, a reverse outline is an outline that we create from an existing text.”  The idea is to reduce a text to its core points, which can then be more conveniently analyzed.

To make a common reverse outline, I ask students to summarize each paragraph in about seven words or less. Why seven words? From my experience, if you need more than seven words, the paragraph might not be very focused, though there are many exceptions (e.g., the first and last paragraphs of a paper, which often list a lot of points).

My friend Jane Freeman is the Director of the Graduate Centre for Academic Communication at the University of Toronto, and she has cautioned that if students use AI, they may be missing out on learning valuable skills. Jane has a good point, so with her advice in mind, I started last week’s class by teaching the traditional reverse outline and asking students to make one of a cover letter (a document that is usually short and manageable to analyze quickly but still has focused paragraphs).

I made my own:

1. Who I am and why I fit the company

2. Professional experience—how it applies to job

3. Current Cornell experience—how it applies to job

4. Invitation to possible next step

Next, I asked if AI can make reverse outlines for us.  It turns out that it actually does a very good job, which makes sense. Chatbots, such as ChatGPT and Copilot, are based on large language models. They take enormous amounts of data and extrapolate patterns from it, all crucial skills for summarizing. Consequently, they can make reverse outlines in seconds.

Then I asked students to try to duplicate their reverse outlines by using an AI tool. Quickly, many of them found that AI has a tendency to be verbose, sometimes giving reverse outlines nearly as long as the original text, and even summarizing more paragraphs than were in the original text. With these points in mind, together we created this prompt:

I will upload an introduction. Make a reverse outline of it. Summarize each of the 6 paragraphs. Be terse.

Students then shared their AI-generated reverse outlines so we could go over them together. Though there were variations, generally they were all pretty similar. Each reverse outline recognized that the second paragraph was about the applicant’s professional experiences and the third was about the applicant’s educational experiences.

However, in my reverse outline above, I noted in the body paragraphs that the applicant’s experiences were connected to the skills required for the job. Specifically, I used this phrase: “how it applies to the job.” The AI-generated reverse outlines did not make this connection. I feel like the connection is an essential part of a cover letter—the applicant needs to show how their experience fits with a company’s needs. Why wasn’t AI picking up on this?

We went back to the cover letter and noticed that the connection between the applicant’s experiences to the job was stated near the end of the paragraphs, and it was not made very clearly. Why did I see it so obviously to me? I think that when I made my reverse outline, I was expecting the connection, and thus I included this in my summary.

I was assuming the cover letter was written a certain way, and my assumption was so strong that I included it even though it was not clearly stated in the text. This is a problem for many of us. When you read your own writing, you fill in ambiguities, like the mind filling in for your eye’s blind spot. The result is that even though we feel we’ve stated something clearly, it might not actually be that clear for the reader.

This discovery led to this strategy: (1) ask AI to generate a paragraph-level reverse outline of your writing, (2) make sure you agree with the reverse outline—but if you don’t, refrain from taking action, (3) compare AI’s summaries to the paragraphs in your text, (4) are the points clearly stated and easy to find for the reader? If not, should you make adjustments? (5) if you disagreed with AI’s summary, can you explain why it came to its conclusions, and should you revise something?

Using this strategy, AI can serve as an objective party, summarizing only what is clearly stated. If AI isn’t picking up something that you thought was evident, perhaps it’s not as clearly stated and as prominent as you thought. In sum, creating an AI-generated reverse outline can help you decide if your ideas are as clear to you as they are to the reader.

A while ago, I wrote to Rachael Cayley and asked her what she thought of AI-generated reverse outlines. Her response was that the process of making a reverse outline has value; having AI make it denies a writer of this value. She’s right. In fact, often when making a reverse outline, writers realize faults in their structure, organization, and/or logic. But what happens when a writer summarizes something that is actually not clear or not even stated? Using AI gives more of an objective perspective that’s not meant to replace the traditional process; it’s just a different approach that has its own merits.

Next week’s blog: In some cases AI can do what a writer has difficulty doing or even can’t do. This is particularly true when we use rhetorical-move reverse outlines for ideas. Watch this space.

Oh, and thanks Rachael, whose blog is the inspiration for me writing this one.

The first Strategies for Writing with AI class started well. Of the 15 students, I had taught six before and know all of them to be intelligent, inquisitive, and diligent—just the qualities needed to explore AI writing tools.

The goal of the first class was to determine what we want the course to entail. We started with a survey and then had an open discussion.

Many of the survey results were a standard bell-curve spread. When asked what they think of AI, their answers ranged from pretty good to revolutionary, with the majority agreeing that “It’s just a tool, but an amazing one.” Experience using AI writing tools yielded similar results: A couple have almost never used AI tools, two others “use it all the time for many things,” and the majority use it “sometimes…for a few things.”

There were only a few items that had skewed results. One asked if they wanted to use AI to write papers for them. Students ranked it the least desirable. Perhaps this is reflective of their context. Graduate students’ writing is a large part of their reputation. Letting ChatGPT write their papers is too risky.

I imagine this is very different for undergraduates, who most often write for an audience of one (i.e., the teacher) for the primary purpose of having their work criticized. When I taught first-year composition, there were plenty of students who needed to fulfill a university requirement but were not interested in becoming better writers. Having ChatGPT write their paper for them would have been very tempting.

But in our class, students were primarily interested in becoming more efficient and effective writers—my goals also.

The other skewed answer was their knowledge of policies about writing with AI. Only two felt policies are clear and had no problem understanding them. Four felt they weren’t very clear, but they could understand them well enough. The other seven felt, “Policies seem to be disparate and unclear.”

This is a problem. Here the experts are needed for guidance, but they are suspiciously absent, leaving students to fend for themselves. Hopefully, this is a trend that will shift—or die—soon.

Following the survey, in our open discussion, students talked about wanting to write faster and more like a “native speaker.” Some also said they wanted to learn how to write well.  

The class confirmed what I suspected: Students see AI as a useful tool, and they’d like to know how to use it to become better writers.

Over the weekend, I created a course schedule. Tomorrow, we start with reverse outlines, something AI does a pretty good job with. I’ll tell you how it went.

At Cornell University, I teach international graduate students, who are always overworked and underslept. Since my classes are pass/fail and optional; the only reason a fatigued student would enroll in one is because I can offer them something useful. Merely interesting is not enough. In fact, one time during the first day of an academic writing class, I went over the syllabus, and then a student packed up her things to leave. I asked her if something was wrong. She said no, just that the class was not what she wanted. I never saw her again.

I’ve worked diligently to meet students’ needs, constantly polling them to find out what’s effective and making appropriate adjustments. For seven years, I honed one of my courses, called Revising and Editing, till it was a well choreographed experience.

Then came ChatGPT.

Almost all my students use English as an additional language, and often just as they speak with accents, they write with them, making minor mistakes, such as missing an article (e.g., a, an, the) or using an inappropriate preposition (e.g., research in, research about, research on, research into). Part of my course was helping them adjust their accents to conform to North American standards. It turns out that ChatGPT, and other AI writing tools, are really good at making these adjustments, a fact that only took my students weeks to figure out. Consequently, my enrollment dwindled.

But where there is need, there is opportunity. I subscribed to ChatGPT and immediately realized that it’s the most powerful writing tool I’ve seen in 25 years of teaching.

Meanwhile, colleagues are raising alarms, pointing out rightfully that if students rely on AI writing tools too much, they might not develop critical thinking skills nor fully gain the benefits from the process of writing. Some also pointed out that AI writing tools are based on the dominant language (i.e., white middle to upper class), and have inherent biases.  All good points, for sure. But I get this nagging feeling that part of their beef was just making up excuses because they, like me, didn’t really want to have to figure out a new technology—one that many of their students have already mastered.

Some teachers have required writing courses, where enrollment is guaranteed. I do not have that luxury. This last semester, I had four students enrolled in my Revising and Editing class. I decided I had to refocus the course and rename it Strategies for Writing with AI.

The first class is on Wednesday, Oct. 16, 2024—two days from now.

My goal is to enable students to work with AI writing tools so they can become more efficient and effective while still working ethically and continuing to learn. Furthermore, they should experience all the benefits of writing, such as organizing and developing ideas.

Great goals, but there is a problem. I really don’t know what I’m doing. and, as far as I can see, there are no reliable sources to turn to. Still, I’ve been experimenting, and I have some ideas to start with. Plus, I’ve been learning from some seasoned AI writers: my students. I’m now excited to see what happens.

Already, students have shown interest. As soon as I announced my course change, six more students enrolled. Weeks later, I’m at capacity with 15.  

This blog is meant to document how the course goes and what strategies I develop with my students. Come back next week, and I’ll tell you how things went.