• Source:JND
Add as a preferred

OpenAI's ChatGPT is among the popular tools for text generation, summarisation and other text-related queries. While it may prove difficult to differentiate these texts from human-written alternatives, the company has reportedly developed a system to watermark the text generated using ChatGPT and a tool to detect these AI writings for about a year. However, the company is currently debating the release of these functionalities for a while now.

Citing The Wall Street Journal, The Verge reported the OpenAI team is internally divided on whether this will be right for the user who may wish to remain anonymous about AI use or whether they should pick a responsible and transparent approach and hit the interests of these users. The watermarking can predict words to detect patterns of AI writing of the model without affecting text quality.

Also Read: OpenAI Starts Rolling Out 'Advanced Voice Mode' To Select ChatGPT Plus Users; Details

OpenAI's survey hinted people across the globe supported an AI detection tool by a "margin of four to one". Spotted by TechCrunch, the company has also confirmed in an updated May blog post that OpenAI has developed a watermarking tool. "Our teams have developed a text watermarking method that we continue to consider as we research alternatives," OpenAI noted.

The ChatGPT maker also said it is "highly accurate" and can detect localised "paraphrasing" but faces challenges while detecting translation systems, rewording via a different model or asking AI to add special characters between words and then delete them. This makes "it trivial to circumvention by bad actors." It could disproportionately affect non-native English speakers.

Also Read: OpenAI Will Now Help Users Identify If Images Are Made By AI

As per documents accessed by the WSJ, this tool is 99.9 per cent effective in watermarking texts and offers a score of how likely the text was AI-generated. The OpenAI survey also revealed that 30 per cent of users would use OpenAI's ChatGPT less if the watermarking system was deployed. The company may consider other methods which are "potentially less controversial among users but unproven".