OpenAI has developed a tool that can identify assignments or any other text written with ChatGPT and requested by students. According to the Wall Street Journal, the company is considering whether to actually release it.
An OpenAI spokesperson has confirmed that the company is researching text watermarking; But due to the complexities involved and its possible impact on the wider ecosystem beyond OpenAI itself, it has taken a cautious approach.
He added: “The text watermarking method that we are developing has a good technical performance; But there are problems and threats in this way that we are investigating. “Text watermarking can be problematic for users who use a language other than English.”
The new text watermarking method will be different from most previous attempts to recognize AI-generated text. Even OpenAI itself retired its previous AI text tracker last year due to its lack of accuracy.
With the new text watermarking method, OpenAI will focus solely on recognizing text written with ChatGPT, not other language models. This is done by making small changes to the way ChatGPT selects words, essentially creating an invisible watermark in the text that can later be detected by a separate tool.
OpenAI recently updated its approach to AI-generated content recognition in a post. According to OpenAI, the new watermarking method is effective against local manipulations such as paragraph changes.
OpenAI claims that the new text watermarking method can even recognize AI-generated text despite manipulations such as translation and rewriting systems with other generative models.
