ChatGPT cheating? US professors grapple with viral AI technology as semester starts

While errors are common and its prose can be rigid, ChatGPT responses are routinely substantial, spitting out anything from cookie recipes to relationship advice to computer coding to how Jane Austen deployed literary techniques in her 19th century novels. — Image by DCStudio on Freepik

Plagiarism in college isn’t new, but the arrival of a popular, free language platform powered by artificial intelligence has academics at Duke, UNC and elsewhere abuzz about its implications.

ChatGPT answers written prompts in seconds with language that is often human-like. It burst onto the scene in early December when it surpassed a million users over its first week. Created by the San Francisco startup OpenAI, it’s one of several large language models that generate responses by culling from a vast trove of information and using machine learning to determine what it writes.

“ChatGPT is the best one that’s publicly available,” said Thomas Hofweber, director of the forthcoming AI Project at the University of North Carolina-Chapel Hill. “The hardware used to train these models has advanced tremendously, and the datasets on which they’re trained are huge. It’s just extremely impressive.”

Hofweber explained that ChatGPT’s capacity to explain complex concepts puts it beyond traditional search engines. Microsoft is considering investing US$10bil (RM43.66bil) in OpenAI with the aim of incorporating ChatGPT into its Bing search engine. (Seeing a competitive threat, Google is weighing how to proceed with its own AI platform.)

While errors are common and its prose can be rigid, ChatGPT responses are routinely substantial, spitting out anything from cookie recipes to relationship advice to computer coding to how Jane Austen deployed literary techniques in her 19th century novels.

This latter ability is of particularly interest to Charlotte Sussman, chair of Duke University’s English department, who is teaching an introductory course on Austen this semester. Spring courses at Duke began Wednesday.

Among her colleagues, Sussman said, ChatGPT is “pretty close to being the number one topic.”

“I’m 100% sure that all instructors in the department are thinking about it and worrying about it,” she said. “They’re already sending me links (to ChatGPT answers) like ‘Look at this! We’re all going to be out of a job.’”

ChatGPT responses aren’t perfect, but instructors point out neither is most actual student writing. And so for students who may feel overwhelmed by course work or health issues, or who simply aren’t prioritising certain classes, the prospect of creating a fully-written, passable essay with a few keyboard clicks may be tempting.

And professors are now on the lookout.

Catching a ChatGPT cheat

Darren Hick, a philosophy professor at Furman University in Greenville, South Carolina, is one of the first instructors to publicise catching a student cheating with ChatGPT.

Last semester, a student returned a take-home exam with details on philosopher David Hume that hadn’t been covered in class. Hick asked if the student had used ChatGPT, and the student admitted doing so.

“Based on the responses I’ve gotten, there is a general concern among teachers on how to deal with this new threat,” he said.

Hick noted that catching ChatGPT cheaters presents a unique challenge. While existing plagiarism detection tools can prove if a student pulled language from specific sources, AI bots compose original text that is not easily traceable to specific authors or publications. There are programs that purport to detect bot writing, including ones that sprang up in the wake of ChatGPT, but Hick said these still fall short of definitively showing that a student tried to get away with using a robot’s work.

“It provides a mountain of circumstantial evidence, but that’s all it is,” he said.

In light of his personal cheating incident, Hick made a change to his spring course syllabus: If he suspects students of submitting AI-written work, he’ll give them an impromptu oral exam to test their knowledge.

Hick also believes ChatPGT cheating will be more prevalent in K-12, where assignments can be more straightforward. And last week, New York City schools banned the program on its student devices, a telling decision from the country’s largest school district.

Grading a ChatGPT essay

As college classes restart across the Triangle this week – at Duke, UNC, N.C. State, N.C. Central and several smaller schools – local instructors are deliberating how they’ll contend with the viral chat bot. No university appears to have set a broad policy on ChatGPT, leaving decisions to individual professors.

William Peace University in Raleigh will address ChatGPT next week at a faculty development meeting. Several area department chairs told The News & Observer they’re likely to formally discuss the technology with colleagues this semester.

To thwart any would-be cheating, Sussman said she will hold more in-person exams free of technology. She also intends to tailor questions in a way she doesn’t believe ChatGPT can easily handle like response papers on specific pieces of Austen criticism. But she may also try welcoming ChatGPT into her pedagogy - Sussman has heard of English instructors who asked students to critique the bot’s responses.

“I mean, aside from it being dangerous or whatever, it is pretty interesting,” she said.

After Sussman shared one of her past assignment prompts about Austen’s Pride and Prejudice, the N&O entered it into ChatGPT and showed her the results.

“If a student handed this in, I would give it a bad grade, circling the vague words, like ‘things’ and ‘feelings,’ and explaining that I expect the student to explain what things we see and what specifically we learn about Elizabeth’s feelings for Mr Darcy,” she said in response. “But, if I didn’t suspect ChatGPT, I would probably give the student credit for the assignment, which might be all they hoped for anyway.”

Elizabeth Brown, a PhD student at UNC-Chapel Hill who was a teaching assistant last semester, said she has found ChatGPT to generate an odd combination of flawless paragraph structure and grammar along with vague perspective.

“It’s really good at like topic sentence, first detail, second detail, conclusion,” she said. “And really interestingly, those are some of the skills that a lot of students seem to be struggling with a bit.”

Brown recently entered one of the assignment prompts from the public policy course she helped teach into ChatGPT and found the answer to be “light on details and evidence.” But she did hypothesise she might have given the paper a B-minus.

Like the arrival of Google 25 years ago, Triangle instructors say the ultimate impact of ChatGPT – both positive and negative – will take time to become clearer. But what is obvious to many is that the technology will be transformative.

“Everybody knows it will dramatically change a lot of things, but nobody can really predict what precisely,” Hofweber said. “Anybody who says otherwise is making it up. – The News & Observer (Raleigh)/Tribune News Service

Subscribe now to our Premium Plan for an ad-free and unlimited reading experience!

Next In Tech News

Meta to introduce watermarking feature for some AI products
AMD forecasts $45 billion AI chip market this year
Apple to launch new iPads, M3 MacBook Air to fight weak sales - Bloomberg News
Explainer-What are the EU's landmark AI rules?
Bain Capital last bidder in race for SoftwareOne - sources
New Mexico sues Meta, CEO Zuckerberg over child protection failures
Disney begins integrating Hulu into Disney+ streaming service
Elon Musk says his AI firm xAI is not raising funds 'right now'
Alphabet unveils long-awaited Gemini AI model
Fresenius Medical Care says data on 500,000 people stolen in U.S

Others Also Read