Watermarking for ChatGPT Coming Soon

Recently, it was revealed that OpenAI has been working for a long time on a tool that can detect text written by ChatGPT with high accuracy. However, the company has shared some insights on why they have not yet released their research and detection methods regarding text watermarking.

According to reports, the AI watermark in question has actually been ready for a year but has not been unveiled until now. OpenAI stated, “Our teams have developed a text watermarking method that we continue to consider while exploring alternatives.” The company mentioned that they have been working on numerous solutions, including classifiers and metadata, and have achieved very accurate results (99.9%).

However, this level of performance cannot always be achieved in every situation. Performance drops due to specific distortions, such as using translation systems on the text, rephrasing it with another generative model, or asking the model to insert a special character between each word and then removing it. While OpenAI is trying to improve performance in these scenarios, they have also explained why they have not yet released the watermarking tool. The company is concerned that text watermarking might disproportionately affect certain groups.

According to a blog post, OpenAI is assessing these risks. The company also mentioned that they will focus on authentication tools for visual-audio content. Once the developments for text watermarking are complete, the tool could be made available, although no timeline has been set yet. Besides the risk of “disproportionately affecting certain groups,” the company is likely also considering the potential for revenue loss.