OpenAI Could Watermark AI Text With “99.9% Accuracy” but That Won’t Happen (Yet)
According to the Wall Street Journal , ChatGPT’s parent company, OpenAI, has apparently been watermarking AI text for over a year. Such technology would be a boon to teachers , researchers and just about anyone who wants to be sure what they’re reading comes from a real person , but the company is concerned that it could be circumvented and could hurt business.
“Technology that can detect text written by artificial intelligence has been discussed internally for two years with 99.9% confidence,” the WSJ writes about OpenAI, claiming that the technology is essentially ready for use and will work by adjusting the model so that it follows a detectable pattern. The publication said the company does not believe the watermarking tool will reduce the quality of ChatGPT’s output, but is concerned that detecting and watermarking text generated by ChatGPT could turn off nearly 30 percent of users who supposedly told OpenAI it would less likely. use ChatGPT if the generated text had a watermark.
In response to the WSJ report, OpenAI published a blog confirming that it studies watermarking internally, but while it “was very accurate and even effective against local interference such as paraphrasing,” it is less useful for text that has been translated or reworded using an external model. It is also vulnerable to hacks such as adding unnecessary characters and then removing them, which OpenAI says makes it “trivial for attackers to bypass.” As an alternative, the company is considering using metadata to label AI-generated text, which it says has the benefit of “no false positives.” The company already uses metadata to label AI-generated images, which it detailed in a blog post.
Whatever method a company uses, the need for it to emerge sooner rather than later is clear. According to a poll commissioned by the WSJ, “people around the world supported the idea of an artificial intelligence detection tool by a margin of four to one.”
Google already uses watermarks for AI-generated text , as explained during this year’s Google I/O conference .