Link: OpenAI developed a watermarking method for detecting text written by ChatGPT with 99.9% reliability (Wall Street Journal)

For nearly two years, OpenAI has debated the release of a project that has been ready for about a year. The company has struggled between its commitment to transparency and the desire to retain users, with internal surveys showing nearly a third of loyal ChatGPT users would be deterred by anti-cheating technology. Despite this, supporters argue that the benefits of the tool, which creates a detectable watermark in AI-generated text, outweigh the risks.

Concerns have been raised about the potential negative impacts on non-native English speakers and the possibility of simple techniques to erase the watermarks. Determining who can use the detector is also a challenge, as too few users would render it ineffective, while too many could lead to bad actors deciphering the technology. OpenAI has considered providing the detector to educators or companies that help identify AI-written papers.

Discussions about the watermarking tool began before ChatGPT's launch in November 2022 and have been a source of tension. Surveys have shown strong global support for an AI detection tool, but also concerns about false accusations and reduced usage. Tests have shown watermarking doesn't impair ChatGPT’s performance, but employees have recommended exploring less controversial approaches and emphasized the need for a plan to influence public opinion and new laws on AI transparency.#

--

Yoooo, this is a quick note on a link that made me go, WTF? Find all past links here.