After a report The Wall Street Journal The company, which claims OpenAI is sitting on a tool that can spot essays written by ChatGPT with high accuracy, has shared a bit about its research into text watermarking — and why it didn’t release its detection method. according to The Wall Street Journal‘s report, debate over whether or not to release the tool prevented it from seeing the light of day, even though it was “ready”. In an update published on Sunday through May blog postnoted by TechCrunchOpenAI said, “Our teams have developed a text watermarking method that we continue to review as we explore alternatives.”
The company said watermarking is one of many solutions, including classifiers and metadata, that it is considering as part of a “wider investigation into text provenance.” According to OpenAI, it was “highly accurate” in some cases, but not so good when faced with some forms of falsification “such as using translation systems, recreating a word with another generative model, or asking the model to input a specific model.” does not perform. character between each word and then delete that character.” OpenAI writes that text watermarking “may disproportionately affect some groups.” “For example, it could detract from the use of AI as a useful writing tool for native English speakers.”
OpenAI measures these risks, according to a blog post. The company also wrote that it prefers to release authentication tools for audiovisual content. in his statement to TechCrunchan OpenAI spokesperson said the company is taking a “deliberate approach” to text provenance because of the “complexities and its likely impact on the broader ecosystem beyond OpenAI.”
