I finished reading and grading the last big project from my “Digital Writing” class this semester, an assignment that was about the emergence of openai.com’s artificial intelligence technologies GPT-3 and DALL-E. It was interesting and I’ll probably write more about it later, but the short version for now is my students and I have spent the last month or so noodling around with software and reading about both the potentials and dangers of rapidly improving AI, especially when it comes to writing.
So the timing of of Stephen Marche’s recently published commentary with the clickbaity title “The College Essay Is Dead” in The Atlantic could not be better– or worse? It’s not the first article I’ve read this semester along these lines, that GPT-3 is going to make cheating on college writing so easy that there simply will not be any point in assigning it anymore. Heck, it’s not even the only one in The Atlantic this week! Daniel Herman’s “The End of High-School English” takes a similar tact. In both cases, they claim, GPT-3 will make the “essay assignment” irrelevant.
That’s nonsense, though it might not be nonsense in the not so distant future. Eventually, whatever comes after GPT-3 and ChatGPT might really mean teachers can’t get away with only assigning writing. But I think we’ve got a ways to go before that happens.
Both Marche and Herman (and just about every other mainstream media article I’ve read about AI) make it sound like GPT-3, DALL-E, and similar AIs are as easy as working the computer on the Starship Enterprise: ask the software for an essay about some topic (Marche’s essay begins with a paragraph about “learning styles” written by GPT-3), and boom! you’ve got a finished and complete essay, just like asking the replicator for Earl Grey tea (hot). That’s just not true.
In my brief and amateurish experience, using GPT-3 and DALL-E is all about entering a carefully worded prompt. Figuring out how to come up with a good prompt involved trial and error, and I thought it took a surprising amount of time. In that sense, I found the process of experimenting with prompts similar to the kind of invention/pre-writing activities I teach to my students and that I use in my own writing practices all the time. None of my prompts produced more than about two paragraphs of useful text at a time, and that was the case for my students as well. Instead, what my students and I both ended up doing was entering in several different prompts based on the output we were hoping to generate. And my students and I still had to edit the different pieces together, write transitions between AI generated chunks of texts, and so forth.
In their essays, some students reflected on the usefulness of GPT-3 as a brainstorming tool. These students saw the AI as a sort of “collaborator” or “coach,” and some wrote about how GPT-3 made suggestions they hadn’t thought of themselves. In that sense, GPT-3 stood in for the feedback students might get from peer review, a visit to the writing center, or just talking with others about ideas. Other students did not think GPT-3 was useful, writing that while they thought the technology was interesting and fun, it was far more work to try to get it to “help” with writing an essay than it was for the student to just write the thing themselves.
These reactions square with the results in more academic/less clickbaity articles about GPT-3. This is especially true about Paul Fyfe’s “How to cheat on your final paper: Assigning AI for student writing.” The assignment I gave my students was very similar to what Fyfe did and wrote about– that is, we both asked students to write (“cheat”) with AI (GPT-2 in the case of Fyfe’s article) and then reflect on the experience. And if you are a writing teacher reading this because you are curious about experimenting with this technology, go and read Fyfe’s article right away.
Oh yeah, one of the other major limitations of GPT-3’s usefulness as an academic writing/cheating tool: it cannot do even basic “research.” If you ask GPT-3 to write something that incorporates research and evidence, it either doesn’t comply or it completely makes stuff up, citing articles that do not exist. Let me share a long quote from a recent article at The Verge by James Vincent on this:
This is one of several well-known failings of AI text generation models, otherwise known as large language models or LLMs. These systems are trained by analyzing patterns in huge reams of text scraped from the web. They look for statistical regularities in this data and use these to predict what words should come next in any given sentence. This means, though, that they lack hard-coded rules for how certain systems in the world operate, leading to their propensity to generate “fluent bullshit.”
I think this limitation (along with the limitation that GPT-3 and ChatGPT are not capable of searching the internet) makes using GPT-3 as a plagiarism tool in any kind of research writing class kind of a deal-breaker. It certainly would not get students far in most sections of freshman comp where they’re expected to quote from other sources.
Anyway, the point I’m trying to make here (and this is something that I think most people who teach writing regularly take as a given) is that there is a big difference between assigning students to write a “college essay” and teaching students how to write essays or any other genre. Perhaps when Marche was still teaching Shakespeare (before he was a novelist/cultural commentator, Marche earned a PhD specializing in early English drama), he assigned his students to write an essay about one of Shakespeare’s plays. Perhaps he gave his students some basic requirements about the number of words and some other mechanics, but that was about it. This is what I mean by only assigning writing: there’s no discussion of audience or purpose, there are no opportunities for peer review or drafts, there is no discussion of revision.
Teaching writing is a process. It starts by making writing assignments that are specific and that require an investment in things like prewriting and a series of assignments and activities that are “scaffolding” for a larger writing assignment. And ideally, teaching writing includes things like peer reviews and other interventions in the drafting process, and there is at least an acknowledgment that revision is a part of writing.
Most poorly designed assigned writing tasks are good examples of prompts that you enter into GPT-3. The results are definitely impressive, but I don’t think it’s quite useful enough to produce work a would-be cheater can pass off as their own. For example, I asked ChatGPT (twice) to “write a 1000 word college essay about the theme of insanity in Hamlet” and it came up with this and this essay. ChatGPT produced some impressive results, sure, but besides the fact that both of these essays are significantly shorter than 1000 word requirement, they both kind of read like… well, like a robot wrote them. I think that most instructors who received this essay from a student– particularly in an introductory class– would suspect that the student cheated. When I asked ChatGPT to write a well researched essay about the theme of insanity in Hamlet, it managed to produce an essay that quoted from the play, but not any research about Hamlet.
Interestingly, I do think ChatGPT has some potential for helping students revise. I’m not going to share the example here (because it was based on actual student writing), but I asked ChatGPT to “revise the following paragraph so it is grammatically correct” and I then added a particularly pronounced example of “basic” (developmental, grammatically incorrect, etc.) writing. The results didn’t improve the ideas in the writing and it changed only a few words. But it did transform the paragraph into a series of grammatically correct (albeit not terribly interesting) sentences.
In any event, if I were a student intent on cheating on this hypothetical assignment, I think I’d just do a Google search for papers on Hamlet instead. And that’s one of the other things Marche and these other commentators have left out: if a student wants to complete a badly designed “college essay” assignment by cheating, there are much much better and easier ways to do that right now.
Marche does eventually move on from “the college essay is dead” argument by the end of his commentary, and he discusses how GPT-3 and similar natural language processing technologies will have a lot of value to humanities scholars. Academics studying Shakespeare now have a reason to talk to computer science-types to figure out how to make use of this technology to analyze the playwright’s origins and early plays. Academics studying computer science and other fields connected to AI will now have a reason to maybe talk with the English-types as to how well their tools actually can write. As Marche says at the end, “Before that space for collaboration can exist, both sides will have to take the most difficult leaps for highly educated people: Understand that they need the other side, and admit their basic ignorance.”
Plus I have to acknowledge that I have only spent so much time experimenting with my openai.com account because I still only have the free version. That was enough access for my students and me to noodle around enough to complete a short essay composed with the assistance of GPT-3 and to generate an accompanying image with GPT-3. But that was about it. Had I signed up for openai.com’s “pay as you go” payment plan, I might learn more about how to work this thing, and maybe I would have figured out better prompts for that Hamlet assignment. Besides all that, this technology is getting better alarmingly fast. We all know whatever comes after ChatGPT is going to be even more impressive.
But we’re not there yet. And when it is actually as good as Marche fears it might be, and if that makes teachers rethink how they might teach rather than assign writing, that would be a very good thing.