With the rise of text-generation tools such as OpenAI’s ChatGPT, the race is on to create software that professors can use to detect the use of such tools by students.
On April 4, Turnitin CEO Chris Caren announced in a blog post on the company’s website that an AI writing detector was now available as a part of their existing product. Turnitin, a plagiarism detection software, is integrated into the Canvas learning management system at St. Joe’s.
While Turnitin boasts a 98% accuracy rate for its AI detection software, critics point out that those numbers have not been independently tested and stories have circulated about “false positives” in which students have been accused of using AI to write text when they insist they did not.
Richard Haslam, Ph.D, associate professor of English, is one faculty member who is wary of the accuracy of these detection tools.
“Not long after ChatGPT came out, there was something called GPTZero, which was created by an undergraduate Princeton student,” said Haslam. “I tried that out, but it didn’t seem to have high reliability.”
Existing programs to detect AI-generated writing, like GPTZero, rely on identifying patterns and phrases that ChatGPT and other programs often use.
“It’s a supervised learning problem,” said Abolfazl Saghafi, Ph.D., professor of mathematics and an expert in machine learning. “We give it some inputs and see whether the output is human generated or machine generated. … it’s a binary classification problem.”
Because of the way AI detection software is created, it cannot identify text that might be given to another program to remove some of the characteristic patterns. Other generative pre-trained transformer (GPT) programs from OpenAI, for example, can be trained on a student’s assignments to recreate their personal writing style. Additional programs exist that can entirely rewrite the material, making them undetectable. As anti-plagiarism software becomes more commonplace, these programs are likely to grow in popularity.
Marcello Balduccini, Ph.D., assistant professor in the department of decision and system sciences, said programs that could rewrite AI-generated text to avoid detection are definitely feasible.
“That would absolutely be possible,” said Balduccini. “[OpenAI’s software] can already sort of do that.”
Without any alteration, however, ChatGPT and other programs might not be advanced enough to pass for authentic student writing. Chase Davis ’24, who works as a tutor in the SJU Writing Center, said he is confident many professors would be able to recognize something written by text-generation AI.
“They’re pretty in-tune with the way a student writes,” Davis said. “ChatGPT has a formalized structure that it will always follow.”
In a recent Primary Research Group study of faculty views on AI technology, though, only 13.62% of surveyed faculty thought their institutions had provided them with adequate guidelines for how to deal with the possibility of AI-generated text in the classroom.
“Faculty will have lots of different views on this,” Haslam said. “[We are] definitely in an era of flux and turmoil and uncertainty.”
