It’s a reality: OpenAI has a system that allows generating watermarks on writings created through ChatGPT, as well as a tool to detect it. However, the company is unsure if this should be released to the public, according to The Wall Street Journal.
Apparently, the company is facing a dilemma: should they update ChatGPT to include watermarks in the generated texts? While it seems like a responsible decision, how would it ultimately affect their business?
In a nutshell, the watermark proposed by OpenAI for ChatGPT would work like this: the words and phrases used would create a kind of easily detectable pattern. Furthermore, this method does not affect the quality of the text generated by the chatbot, according to TWSJ. OpenAI ensures that their tool is very reliable and capable of achieving a 99.9% effectiveness. In fact, it is so good that it can detect alterations, ‘such as paraphrasing.’ However, it is not as good at detecting other types of reformulations, such as those that may result from combining the texts of ChatGPT with other generative models.
For many, this is good news. Since its release, many users have requested a tool to help them distinguish texts generated by ChatGPT. However, the implementation of watermarks also has a downside for OpenAI: a decrease in users.
According to what we can read on The Verge, a number of surveyed ChatGPT users (almost 30%) told OpenAI that they would use the chatbot less if these watermarks were finally implemented. Additionally, OpenAI states in their blog that their detection tool could ‘stigmatize the use of AI as a useful writing tool for non-native English speakers.’
Currently, the company is studying other alternative detection methods, such as the use of metadata. OpenAI clarifies that, although they are ‘in the early stages of exploration,’ it would be a more effective approach. ‘Unlike watermarks, metadata is cryptographically signed, which means there are no false positives.’





