OpenAI, the research lab behind the AI program ChatGPT, has released a tool designed to detect whether text has been typed by artificial intelligence, but cautions that it’s still not completely reliable.
In a blog post on Tuesday, OpenAI linked to a new classifier tool that has been trained to distinguish between human-written text and that written by a variety of AIs, not just ChatGPT.
The Open AI researchers said that while it was “impossible to reliably detect all text written by AI,” good classifiers could detect signs that the text was written by AI. The tool could be useful in cases where AI was used for “academic dishonesty” and where AI chatbots were positioned as human, they said.
But they admitted that the classifier is “not entirely reliable” and only correctly identified 26% of AI-written English texts. It also incorrectly labeled human-written texts as likely written by AI tools 9% of the time.
“The reliability of our classifier generally improves as the length of the input text increases. Compared to our previously released classifier, this new classifier is significantly more reliable on the text of the latest AI systems.”
Ever since ChatGPT was opened to public access, it has caused a wave of concern among educational institutions around the world that it could lead to cheating on exams or assessments.
UK teachers are being urged to review the way their courses have been assessed, while some universities have banned the technology altogether and reverted to paper-and-pencil exams to prevent students from using AI.
A professor at Australia’s Deakin University said about one in five of the assessments she was grading during the Australian summer term had used AI assistance.
Several scientific journals have also banned the use of ChatGPT in the text of articles.
OpenAI said the classification tool had several limitations, including its unreliability in text under 1,000 characters, as well as misidentifying some human-written text as AI-written. The researchers also said it should only be used for English text, as it performs “significantly worse” in other languages and is not reliable for checking code.
“It should not be used as a primary decision-making tool, but rather as a complement to other methods of determining the source of a text,” OpenAI said.
OpenAI has now asked educational institutions to share their experiences using ChatGPT in classrooms.
While most have responded to AI with bans, some have embraced the AI wave. The three main universities in South Australia last month updated their policies to say that the use of AI such as ChatGPT is allowed as long as it is disclosed.
