New Turnitin AI Writing Detector Claims 98% Accuracy
Academic plagiarism-spotting software provider Turnitin will alert educators to AI-written text. Turnitin claims the new tool can identify when ChatGPT or other generative AI is the author with 98% accuracy.
Turnitin has been developing the new technology for two years, anticipating the release of ChatGPT and other synthetic text generators and the demand for AI writing checks among academic institutions. The new tool is available to more than 2.1 million teachers at more than 10,700 schools and educational institutions through the standard Turnitin platform. Turnitin’s AI detector tests submitted text for signs of AI assistance, marking sentences that trip its AI alarm. It then measures how much of the whole essay is AI-generated. The model is trained on academic writing rather than public content to make it better at spotting issues in what students submit. The software errs on the side of avoiding false positives and only marks an issue when it has more than 98% confidence.
“For 25 years, Turnitin has been a trusted partner of educators, giving insights into student writing and helping to uphold the standards of academic integrity for more than 16,000 institutions in 140 countries worldwide,” Turnitin CEO Chris Caren said. “Educators told us that being able to accurately detect AI-written text is their first priority right now. They need to be able to detect AI with very high certainty to assess the authenticity of a student’s work and determine how to best engage with them. It is equally important that detection technology becomes a seamless part of their existing workflow, which we have done by integrating AI detection capabilities into Turnitin solutions.”
There are a growing number of attempts to spot AI-generated writing. Companies and independent developers are experimenting with how best to trace text, including GPTZero, which has had some success in spotting AI authorship based on elements like randomness and complexity of the language. OpenAI released its own tool for detecting text written by generative AI models. The ‘classifier’ is still under development and not yet always able to accurately distinguish when a text is written by AI, garnering a 26% accuracy rating and 9% false positives.