OpenAI creates watermarks to recognize ChatGPT texts; The possibility of reducing users with the release of this feature

According to the Wall Street Journal, OpenAI has developed a system for adding watermarks to texts produced with ChatGPT and a tool to recognize these watermarks about a year ago; But he is skeptical about the public release of the tools. On the one hand, the release of the tool in question seems to be a responsible action, and on the other hand, it may end up harming his business.

The watermark that OpenAI applies to text generated with ChatGPT creates a recognizable pattern by changing the way it predicts subsequent words and phrases.

Any tool that can recognize AI-generated text can be very useful for teachers who want to prevent students from cheating using AI. The company has found that adding a watermark does not harm the quality of the chatbot’s output. Also, according to the poll, people around the world support the idea of ​​creating a tool to recognize AI-generated content by a ratio of four to one.

OpenAI confirmed the development of the watermark technology by publishing a blog post. This company says that it uses a very accurate method for watermark detection, which according to some documents, its accuracy level reaches 99.9% and is resistant to possible manipulations; However, OpenAI says its watermarking technology can be circumvented by techniques such as text rewriting by other AI models.

Apparently, OpenAI is concerned that the use of watermarks will reduce the satisfaction of ChatGPT users. Almost 30% of the users participating in the survey have declared that they will reduce the use of this software if the watermark system is implemented.

Despite these concerns, some OpenAI employees believe that watermarking is an effective method. However, the Wall Street Journal reports that due to negative user comments, some employees have suggested trying other, less controversial methods. In its blog post, OpenAI says that it is in the early stages of the feasibility of embedding metadata in the texts produced by artificial intelligence, and it is not yet possible to comment on the efficiency of this method.