Plagiarism detection software that is already screening the work of more than 62 million students worldwide is getting a major upgrade to detect AI-generated writing 鈥� though some universities say they don鈥檛 want it. This comes several months after the release of the AI chatbot ChatGPT that can generate entire essays prompted major US school districts and universities around the world to ban its use.
It is unclear if AI-generated writing can be reliably detected. Academic researchers have shown that it is mathematically impossible for detectors to keep up. Even OpenAI, the company that developed ChatGPT, has struggled with this challenge 鈥� its only correctly identified 26 per cent of AI-generated writing as 鈥渓ikely AI-written鈥� when tested on sample English texts.
鈥淲e did not expect that OpenAI would release ChatGPT to the world as quickly as it did,鈥� says a spokesperson at Turnitin, the company behind the plagiarism detection software. However, the spokesperson also says Turnitin already had a team in place ready to work on adding tools for detecting AI-based plagiarism to its software.
Advertisement
The Turnitin AI writing detector was initially trained to analyse English-language prose sentences generated by OpenAI鈥檚 GPT-3 and GPT-3.5 AI technology, which is used by the free version of ChatGPT. That currently leaves out text generated by GPT-4, which powers the premium version of ChatGPT and Microsoft鈥檚 Bing Chat, along with writing generated by Google鈥檚 Bard AI chatbot.
One preview video shows the Turnitin software highlighting suspected AI-generated sentences in green and then displaying the number of AI-generated sentences out of the total. A Turnitin slideshow presentation shows another format where a percentage appears below a blue box titled 鈥淎I鈥�, which the slide says is an indicator of how much of a document may be AI-generated.
The detector supposedly only flags sentences when it has 98 per cent confidence. False positive situations that involve falsely flagging human writing as AI writing are supposed to occur just 1 per cent of the time 鈥� Turnitin says it trained its detector to avoid such mistakes in exchange for potentially failing to catch some AI-generated sentences.
, a librarian at the British Columbia Institute of Technology in Canada, expressed doubt that the Turnitin detector鈥檚 false positive rate will be so low in practice, given that OpenAI鈥檚 own AI detector had a higher false positive rate of 9 per cent. He and others who spoke to 快猫短视频 also warned that the AI detector is a 鈥渂lack box model鈥� that does not explain the reasoning behind its results 鈥� the original plagiarism software at least showed its work by pointing to the original material that was being plagiarised.
Even the advertised false positive error rate could lead to 200 or more cases being flagged by Turnitin鈥檚 software out of the 20,000 student writing submissions in an average semester, says at the University of Michigan-Dearborn. 鈥淲e don鈥檛 want to be falsely accusing hundreds of students,鈥� says Casey.
This has led some educators to raise concerns about whether Turnitin is rushing to roll out the update scheduled for 4 April without giving teachers the chance to even try out the AI writing detector beforehand.
Certain educational institutions asked to be exempted from the update after Turnitin initially described the detector as being the new software default. Beyond individual US universities, including the University of Michigan-Dearborn, 鈥渁 significant majority of UK institutions have chosen not to deploy at this stage鈥�, says at the Universities and Colleges Information Systems Association, a nonprofit representing digital practitioners in higher education in the UK.
鈥淚t really is a shame that Turnitin hasn鈥檛 provided the ability to opt out to everyone or hasn鈥檛 communicated it to everyone,鈥� says Linkletter.
