OpenAI has been developing a system to watermark text written by ChatGPT, but it has struggled to agree internally on whether to make it public. The Wall Street Journal has reported that OpenAI has confirmed its efforts to create a system to watermark text, following the report. 

The company has stated that its system is precise and can even detect localized manipulations like paraphrasing, but it is less effective against more widespread manipulations. However, the company has chosen to delay its release due to worries about the negative perception of using AI as a tool for writing in non-native English.

The company mentioned that watermarking is just one of several strategies, including the use of classifiers and metadata, that it has explored in its extensive study of text provenance. OpenAI has noted that its system has been very accurate in some cases but has not been as successful in dealing with certain types of manipulations, such as translations, rephrasing with another AI model, or asking the model to insert a special character between each word and then removing it.