Hot off the presses: How students use ChatGPT to prepare and write their essays. 1/n
My TAs and I recently had a case of an essay by a student that Turnitin identified as being largely (71%) AI-generated.
Actually, one of the TAs, Frank Onuh, had flagged the essay as being potentially AI-generated before we got the Turnitin score, largely on the rhetoric — he is working on a project about chatbot rhetoric and saw some common flags (e.g. a lot of “nuance,” “delve,” and adjective-heavy prose that begins with summation/identity statements: “George Elliot is a well-respected writer of the 19th century whose highly nuanced novels delve into the contested meaning of relationships and gender”).
Earlier this year, there were various warnings circulating on the internet indicating that such bot-detectors were not necessarily very accurate — and indeed that they reported frequent false positives. I don’t know if this means that the technology has improved, but Turnitin is now claiming that there is a high possibility of false positives on their detector when the amount detected is relatively low (say 20% or less), but not once it gets above that.
Personally, I wonder if this is because student paragraphs and essays often begin quite stereotypically, and, of course, stereotypical writing is what chatbots do: just as, in my experience, it is common to see Turnitin’s similarity checker flag language more heavily at the beginning and end of essays and paragraphs — I have always assumed because students also often begin with identity statements “Frankenstein is a novel by Mary Shelley” — I imagine the AI-detector produces false positives in those same places because even genuine student writing is often machine-like in its anodyneinity (to coin a word that I bet ChatGPT has never used).
I then had a conversation with the student about the use of ChatGPT as a research tool. As part of this conversation, I asked them exactly how they used the bot — i.e. what did they type in the chat box, what did they do with the results, and so on.
Asking students for details of their technical workflow is something I’ve done for over a decade, as, again in my experience, this tends to change generation by generation. About fifteen years ago, for example, I was finding some strange kinds of plagiarism (and arguments) in various student papers: plagiarism in the choice and arrangement of quotations, for example, but not in the content in between; students quoting too much rather than not enough, etc. I uncovered what was going on at the time by asking a few students to sit down with me and show me how they worked: how they wrote their notes, how/if they wrote outlines, how they searched for and found material, and so on.
What I discovered then was that students were thinking (and composing) their essays largely in the way bloggers worked. As I wrote then:
[T]he point of blogs, increasingly, is less to digest facts and arguments than to accumulate and react to them. Political blogs—like the Ed Morrisey’s Captain’s Quarters (now at Hot Air, or Dan Froomkin’s Whitehouse Watch)—tend to consist of collections of material from other on-line sources interspaced with opinion. The skill an accomplished blogger brings to this type of material lies in the ability to select and organise these quotations. A good blog, unlike a good essay, builds its argument and topic through the artful arrangement and excerpting of usually verbatim material passages from other people’s work—in much the same way that some types of music are based on the original use and combination of digitised sound samples from earlier recordings.
When I interviewed my students, I was completely surprised by how they worked:
On the basis of my interviews, it appears to me that most of my first year students now conduct their research and compile their notes primarily by searching the Internet, and, when they find an interesting site, copying and pasting large sections of verbatim quotation into their word processor. Often they include the URL of this material with the quotations; but because you can always find the source of a passage you are quoting from the Internet, it is easy for them to get sloppy. Once this accumulation of material is complete, they then start to add their own contribution to the collection, moving the passages they have collected around and interspacing them with their opinions, arguments, and transitions.
Since I grew up in the pre-copy-and-paste era, this was completely beyond my imagination: I had never thought of an essay as a series of quotations separated by comment, nor thought of my job in writing one as being the arranging and commenting on of quotations. Nor did any of the colleagues of whom I asked the same kind of workflow questions.
As my example blogs (and the example of blogs) suggests, this approach was more than a few student-generations ago. I doubt that many students today follow blogs particularly — newsletters and social media have largely taken over the space, methinks. But you can still learn a lot by asking students questions like “and so what did you do next? What did you type in the search bar? What did you name your file?”
In this case, the student told me that they wrote their essays by giving ChatGPT a quotation and asking it to interpret it for them. So, for example, “Interpret the sentence ‘No one who had ever seen Catherine Morland in her infancy, would have supposed her born to be an heroine’.” If the bot produced bullet points, then they might specify “… in paragraph form” and sometimes they would give a context for the interpretation “… in light of the rise of the novel.” They would then do this for other quotations they thought might be important and then, on the basis of these responses, “put things into their own words.”
You can see how this might go wrong, even if we don’t assume that there was some cutting and pasting going on: one is almost certainly going to reflect the (anodyne) language of ChatGPT in one’s own answer, raising the likelihood that the language will be flagged by the bot-checker even if it was your own. In addition, because the essay is being built out of independent analyses of individual passages, there may not be much in the way of overarching argument in the outputs one gets.
Since this is a very new technology and the essay was a formative exercise, I decided that this was a teaching rather than disciplinary moment, and so I discussed with the student ways that you can use a chatbot to help you discover something you want to write about, rather than what you should write about the thing you chose. I showed them how you can ask it to help you develop research questions in fields you don’t know much about. And I showed them how you can use it to find potential sources as a starting point… warning them that book and article titles are places where ChatGPT seems particularly prone to hallucinating.
And indeed it did: the first article we took from it and then searched for in the library was a complete hallucination. Although ChatGPT gave us the full title, the full author name, and the journal in which was supposed to have appeared, neither Google Scholar, nor Google, nor the Journal, nor the supposed author’s Google Scholar profile seemed to know about it: a search for the article title in quotations marks produced no results, even in shortened form.
This was a first look at how students are using ChatGPT but it certainly won’t be the last: one of the things the student told me was how their siblings and siblings friends in high school were trying to disguise the use of ChatGPT — mostly it seemed by copying the text into a word processor and reformatting.
But I’ll delve into the nuances of that some other day.
