Like countless faculty this post-ChatGPT academic year, I had the eerie sensation of occasionally reading papers that sounded off.
To be sure, they weren’t off in the ways a professor’s grading brain is trained to scan for: concepts misconstrued, non sequitur paragraphs, typos carelessly lurking. Quite the opposite. These were competent, articulate, polished, and yet not-quite-human papers.
Generative AI style, if it can be said there is such a thing, feels Wikipedian: Early on, it was called “beige”; a “generic cake mix.” Like the similarly crowdsourced Wikipedia, it sounds like everyone and no one at the same time. And, says Naomi Baron, a linguist who studies computer-mediated communication, just as when Wikipedia first came along and “People said, ‘Wait a second, do you trust what is there? Because who knows who actually wrote it?’” AI-generated text sparks suspicion.
As text that’s probabilistically formulated, AI’s prose is a product of a series of safe bets. Human writing, in contrast, is defined by inconsistency: be it the lyrical variation of a savvy stylist or the sloppy haste of an undergrad fumbling their and there.
What troubled me in my students’ essays was not the plagiarism. That might not even be the right word for this offense. Indeed, says Anna Mills, a California-based community college teacher who advocates for AI literacy in college classrooms, “Plagiarism doesn’t quite fit, because AI is not a human author you can steal from.” The better analogy might be to those online mills that have long hawked term papers.
No, what troubled me was the feeling that we’re arriving at what AI scholar Karl MacDorman calls the “uncanny valley of the mind,” not unlike highly realistic robots that creep us out.
As a child, Japanese roboticist Masahiro Mori got that uneasy tingle when he visited wax museums or saw prosthetic limbs. In 1970, the Tokyo Institute of Technology professor published in an obscure Japanese journal an essay that nonetheless became one of the most influential ideas in the field of human-robot interaction. His concept of the uncanny valley holds that the more human a robot appears, the more likable it is — but only up to a point. Too human-seeming and a robot triggers the same revulsion we might feel when seeing a corpse or a zombie. That feeling may stem from some confusion buried deep in our consciousness about whether an entity is animate or lifeless.
When I query ChatGPT about the uncanny valley, it knows where I’m coming from: “If an AI generates text that is highly convincing in terms of content and coherence, but lacks emotional depth or genuine understanding . . . the text may seem almost human-like, but the absence of authentic emotion or empathy can create a sense of dissonance,” it tells me, its tone radiating HAL 9000-like calm.
Chatbots have been around since the 1960s, but they weren’t widespread. Yoshua Bengio, the so-called godfather of AI, has predicted that one day AI text will become “almost indistinguishable” from human-written text.
At that point, the perceptual eeriness might wear off. The existential eeriness will likely remain.
Universities have by and large responded to the advent of generative AI tools with the policy equivalent of an emoji shrug. According to some surveys, only about 15 to 20 percent of schools and instructors have developed guidelines on AI use.
There are faculty who think we should integrate and encourage the use of ChatGPT in academia, and of course there are just as many who think the opposite. But most are just struggling to figure out how to address questions like the one we have just considered: Is using ChatGPT plagiarism?
While university administrations wrangle with this question, 70 percent of college-bound seniors have an answer: Yes, using AI is a form of cheating and plagiarism. Yet one quarter of them are already using AI to complete essays and assignments.
An undergraduate-authored piece in the Chronicle of Higher Education with a clickbait headline tailor-made for professors, “You Have No Idea How Much We’re Using ChatGPT,” was the Chronicle’s most-read article last year. GPTZero, a popular AI detector, scans text for “perplexity,” or random word choice; and “burstiness,” or sentence complexity and variation — structural features that seem to separate human from robot prose.
One trick to make written text appear human is to sprinkle some typos into a piece of AI-generated writing, gumming up the evidence with a bit of sloppiness. Another tactic, more subtle, involves inserting a special character, a zero-width space, before some letters: The space scans like an error to the software — and therefore the product of a flawed human — but is invisible to human eyes.
Newer iterations of GPT are getting subtler. One literature professor at Berkeley had apparently not come across words like “delve” and “multifaceted” in 30 years of teaching Dostoevsky before seeing them suddenly populate his students’ compositions last fall. Such genericisms contribute to the lifelessness of the prose. They are a dead giveaway that a student has had an AI assist. But the savvier, spicier output of newer generative AI systems may be able to emulate the human voice more effectively.
We have been here before: The internet would “weaken the traditional major institutions of learning,” one economist announced in Science in 1995. “Many of the traditional functions of universities will be superseded . . . their role in intellectual inquiry reduced.”
The Silicon Valley hype machine would certainly love to convince us that our era is the AI equivalent of the early 1990s internet. Faculty don’t begrudge or police students using Google in academic research and spell-check is a rudimentary AI we expect they’ll employ. Something is surely lost, though, when the tools to fix style are allowed to formulate substance as well. This, then, is the challenge of using generative AI in higher education: How do we teach and practice thinking and creativity — precisely the skills these tools encourage us to outsource for the sake of efficiency?
Writing is not just a means to an end: It is a means of self-discovery. Finding one’s voice is not just about what ends up on the page — the journey matters more than the destination.
Michael Serazio is a professor of communication at Boston College and the author of, most recently, “The Authenticity Industries: Keeping It ‘Real’ in Media, Culture, and Politics.”