OpenAI Already Has Tools To Detect ChatGPT-Generated Text ...
OpenAI has confirmed the development of a watermarking ability and detection tool for ChatGPT-generated text. These tools have been in development for about a year but are yet to be rolled out due to internal conflicts within the company.
According to a recent blog post by OpenAI, the detection tool has a high accuracy rate in identifying text generated by ChatGPT, thanks to the incorporation of watermarks within the text. However, the effectiveness of the system is limited in cases where the text has been altered by other generative models or translation tools.
Challenges and Considerations
OpenAI acknowledges that the detection tools have not been released over concerns of reduced usage by surveyed ChatGPT users. There are also worries that such tools could stigmatize the use of AI as a writing tool for non-native English speakers. Despite these concerns, some employees at OpenAI believe that watermarking is the way forward.
Aside from watermarking, OpenAI is exploring other potential solutions to identify ChatGPT-generated text. These include embedding classifiers and metadata, which are still in the early stages of development.
Sources: WSJ / OpenAI official blog
