Today, a lot of students are using artificial intelligence for their studies, which is a problem in a lot of colleges, high schools and universities. Teachers are wondering “How do you know if a work is authentic or was generated by AI?”. OpenAI could soon help address this issue, with the company considering releasing a tool to detect text generated by ChatGPT.
A tool that could be widely used by teachers
OpenAI recently developed a tool that can detect in seconds whether a text has been generated by ChatGPT. However, this innovation is still in the thinking phase as to whether it will be made available to the public. The company is hesitant to make this tool accessible, fearing that it will dissuade students and professionals from using ChatGPT. A decrease in the number of users could slow down the evolution of AI and impact the company’s revenue. Indeed, an AI requires frequent use to refine its algorithms and continually improve.
Historically, attempts at AI-generated text detection have been largely ineffective. OpenAI even scrapped its previous AI text detector last year due to its low accuracy rate. However, this time around, OpenAI may have managed to come up with something reliable. An OpenAI spokesperson confirmed that the company is actively exploring a method of watermarking text.
OpenAI’s text watermarking is aimed exclusively at detecting text from ChatGPT. The method involves making slight adjustments to ChatGPT’s word selection, creating an invisible watermark. According to an update to a May blog post from OpenAI, the technique is “highly accurate” and “effective against local forgery.”
However, this watermark has limitations. It is considered “less robust against large-scale tampering” when a user uses translation systems, rewording with another model (such as Gemini for example) or inserting and removing special characters. As a result, OpenAI acknowledges that this method can be “easily circumvented by individuals experienced in using AI.”
