OpenAI – the organisation behind ChatGPT – made the eye-opening decision to discontinue a tool meant to detect AI-generated writing, due to its insufficient accuracy. The company says it is now refining more reliable AI detection methods while creating new mechanisms to help users identify AI-created audio or visual content.
While the failure of OpenAI’s technology to detect generative writing tools like ChatGPT has obvious implications for educators, some experts say studies show AI detection does in fact work.
“OpenAI and the other detection tools out there are using the vast database that is the internet; the model wasn’t specifically designed to detect AI writing within student papers,” Annie Chechitelli, chief product officer of Turnitin said.
“It’s not that AI writing detection technology doesn’t work; It’s that their AI writing detection doesn't work.”
What people often forget about, says Chechitelli, is that AI is that it is only as accurate as the data on which it was trained.
“AI used in academic settings needs to be trained on academic material. The same factor applies to AI writing detection. The AI used to detect AI used in student writing has to be trained on student writing. At Turnitin, we have 25 years of experience with student writing,” she said.
“There have been several research tests of our systems, and those of others. For example, a recent study in Europe, in collaboration with the European Center for Academic Integrity, tested 14 AI detectors, including Turnitin’s.”
According to their research, not only did Turnitin’s system have the highest overall accuracy, it was one of just two with perfect accuracy on human-written and computer-translated text. The study found, “Turnitin received the highest score using all approaches to accuracy classification.”
A separate published study found, “Turnitin's ability to detect 91% of the generated submissions as containing some AI-generated content is promising, despite the deployment of adversarial techniques to evade detection by the research team”
“So, the research shows that not only do AI detectors work, ours works quite well.”
Another program that has been trained to detect programs like ChatGPT and Claude is AICheatCheck, which verifies student work authenticity with an accuracy rate of 99.7%.
The company’s co-founder, Aaron Shikhule believes that while more solutions are emerging to combat plagiarism using AI, technologies like ChatGPT will pose a “significant challenge” to academic integrity in Australian schools in 2023 and beyond.
“With the ability to generate humanlike text, students may be tempted to use these chatbots to cheat on assignments,” Shikhule told The Educator.
“Additionally, educators may find it difficult to detect and prevent the use of these chatbots, as they can be easily accessed via the internet, and many are free to use.”
Shikhule said it is therefore important for schools to stay informed about the development of such technology and implement measures, such as AICheatCheck, to detect and prevent its misuse in academic settings.
Giving teachers peace of mind
Shikhule AICheatCheck has been trained to detect AI-generated text, unlike some other anti-plagiarism software in the market.
“The AICheatCheck model has been trained in a way that the chance of detecting human writing as AI has been minimised. Evasive methods will be added to our model updates soon and the team is working hard to keep the model up to date,” he said.
“With AICheatCheck, school leaders and teachers can have peace of mind that they are effectively detecting and preventing AI cheating through the use of AI technology.”
Arend Groot Bleumink, a Netherlands-based AI engineer and co-founder at AICheatCheck, said while ChatGPT is an opportunity for students to learn more in depth on certain topics, it should be used responsibly.
“It’s like using a performance enhancing drug… so you need to notify people that you’re using it without claiming that it’s all just your own hard work.”