Large language models have gained traction since the commercial release of OpenAI’s GPT-3 in 2020. Now multiple companies have built their own rival machine learning systems, kickstarting a new wave of startups developing products powered by generative AI. These models operate like general-purpose chatbots. Users type instructions, and they will respond with passages of coherent, convincing text.
Students are increasingly turning to AI tools to complete assignments, while teachers are only beginning to consider their impact and role in education. Opinions are divided. Some believe the technology can hone writing skills, while others see it as cheating. Schools in California, New York, Virginia, and Alabama have blocked pupils from accessing the latest ChatGPT model on public networks, according to Forbes.
Education departments aren’t quite sure what academic policies should be introduced to regulate the use of AI text generators. Besides, all rules would be difficult to enforce anyway considering there is currently no effective way to detect machine-written work. Enter Turnitin. Founded in 1998, the US company sells software that calculates how similar a particular essay is compared to content from a large database of papers, webpages, and books to look for signs of plagiarism.
Turnitin was acquired by media giant Advanced Publications for $1.75 billion in 2019, and its software has been used by 15,000 institutions across 140 countries. With over two decades of experience, Turnitin has a broad reach in education and has amassed a huge repository of student writing, making it the ideal company to develop an academic AI text detector.
Turnitin has been quietly building the software for years ever since the release of GPT-3, Annie Chechitelli, chief product officer, told The Register. The rush to give educators the capability to identify text written by humans and computers has become more intense with the launch of its more powerful successor, ChatGPT. As AI continues to progress, universities and schools need to be able to protect academic integrity now more than ever.
“Speed matters. We’re hearing from teachers just give us something,” Chechitelli said. Turnitin hopes to launch its software in the first half of this year. “It’s going to be pretty basic detection at first, and then we’ll throw out subsequent quick releases that will create a workflow that’s more actionable for teachers.” The plan is to make the prototype free for its existing customers as the company collects data and user feedback.
“At the beginning, we really just want to help the industry and help educators get their legs under them and feel more confident. And to get as much usage as we can early on; that’s important to make a successful tool. Later on, we ‘ll determine how we’re going to produce it,” she said.
Patterns in AI writing
Although text generated by AI is convincing, there are telltale signs that reveal an algorithm’s handiwork. The writing is usually bland and unoriginal; tools like ChatGPT regurgitate existing ideas and viewpoints and don’t have a distinct voice. Humans can sometimes spot AI-generated text, but machines are much better at the job.
Turnitin’s VP of AI, Eric Wang, said there are obvious patterns in AI writing that computers can detect. “Even though it feels human-like to us, [machines write using] a fundamentally different mechanism. It’s picking the most probable word in the most probable location, and that’s a very different way of constructing language [compared] to you and I,” he told The Register.
“We read by jumping back and forth our eyes without even knowing it, or flitting back and forth between words, between paragraphs, and sometimes between pages. We’ll flip back and forth. We also tend to write with a future state of mind I might be writing, and I’m thinking about something, a paragraph, a sentence, a chapter; the end of the essay is linked in my mind to the sentence I’m writing even though the sentences between now and then have yet to be written.”
ChatGPT, however, doesn’t have this kind of flexibility and can only generate new words based on previous sentences, he explained. Turnitin’s detector works by predicting what words AI is more likely to generate in a given text snippet. “It’s very bland statistically. Humans don’t tend to consistently use a high probability word in high probability places, but GPT-3 does so our detector really cues in on that,” he said.
Wang said Turnitin’s detector is based on the same architecture as GPT-3 and described it as a miniature version of the model. “We are in many ways I would [say] fighting fire with fire. There’s a detector component attached to it instead of a generate component. So what it’s doing is it’s reading language in the exact same way GPT-3 reads language, but instead of spitting out more language, it gives us a prediction of whether we think this passage looks like [it’s from] GPT-3.”
The company is still deciding how best to present its detector’s results to teachers using the tool. “It’s a difficult challenge. How do you tell an instructor in a small amount of space what they want to see?” Chechitelli said. They might want to see a percentage that shows how much of an essay seems to be AI-written, or they might want confidence levels showing whether the detector’s prediction confidence is low, medium, or high to assess accuracy.
The software isn’t designed with the goal of getting ChatGPT banned in academia. Although it could deter students from using these types of tools, Turnitin believes its detector will instead enable teachers and students to trust each other and the technology.
“I think there is a major shift in the way we create content and the way we work,” Wang said. “Certainly that extends to the way we learn. We need to be thinking long term about how we teach. How do we learn in a world where this technology exists? I think there is no putting the genie back in the bottle. Any tool that gives visibility to the use of these technologies is going to be valuable because those are the foundational building blocks of trust and transparency.” ®