OpenAI hides ChatGPT text recognition technology – it could scare users

OpenAI has technology that can recognize text written by ChatGPT with 99.9% accuracy, but the developer did not make it available to the public, writes The Wall Street Journal. The company has concerns that the placement of a tool to determine whether ChatGPT is being used may lead to the refusal of part of the audience to use its AI solutions.

Image source: Growtika/Unsplash.com

The discussion on this matter lasted at OpenAI for about two years, a source told the WSJ. OpenAI employees were oscillating between the company’s stated commitment to transparency and its desire to attract and retain users. There are also concerns that use of the tool may disproportionately impact user groups that are not native English speakers.

One of the company’s surveys of loyal ChatGPT users found that nearly a third would stop using an AI bot or use it less frequently if the tool were offered. The proposed technology allows you to leave watermarks in the text that are invisible to the human eye, which can be recognized using OpenAI detection technology. This technology allows you to determine that all or part of the text was written by ChatGPT.

However, employees expressed concern that watermarks can be erased using fairly simple methods, for example, by translating the text using Google Translate into another language and vice versa, or by adding emoji to the text and then manually removing them.

OpenAI CEO Sam Altman and CTO Mira Murati took part in discussions about the possibility of using the tool to combat AI fraud, but no final decision was made.

According to WSJ sources, OpenAI top managers and researchers met in early June to discuss the project. Meeting participants agreed on the effectiveness of watermarking technology to determine the authorship of text, but no final decision was made because the results of a survey of ChatGPT users last year raised concerns about loss of audience.

Google has also developed a watermarking technology called SynthID, which allows text generated by Gemini AI to be detected. The technology is currently in beta testing and is not yet available to the public.