OpenAI Considers Adding Watermarks to ChatGPT-Generated Text

According to The Wall Street Journal, OpenAI has had a system for watermarking text generated by ChatGPT and a tool for detecting watermarks ready for about a year. But the company is internally debating whether to release it. On the one hand, it seems like the responsible thing to do; on the other, it could hurt the company’s bottom line.

OpenAI's watermarking is described as adjusting how the model predicts the most likely words and phrases after previous words and phrases, creating a detectable pattern.

Providing any way to detect AI-written documents is a potential boon for teachers trying to discourage students from using AI to do their homework. The Journal reports that the company found that watermarks did not affect the quality of the chatbot’s text output. In a survey the company commissioned, “people around the world supported the idea of ​​AI detection tools by a 4:1 margin,” the Journal writes.

After the Journal published its story, OpenAI confirmed that it had pushed for the text watermark in an updated blog post today spotted by TechCrunch. In it, the company said its method was accurate (“99.9% effective,” according to documents seen by the Journal) and resistant to “spoofing, such as paraphrasing.” The company also said it was concerned about stigmatizing the usefulness of its AI tools for non-native speakers.

But it seems OpenAI is also worried that using watermarks could alienate ChatGPT users, who according to a survey, nearly 30% of them told the company they would use the software less if the watermark feature was implemented.

Still, some employees feel that the watermarking feature will work. But in response to user frustration, the Journal said that some of the proposed testing methods “may be less controversial among users but are unproven.” In an update to today’s blog post, the company said it is “in the early stages” of exploring metadata embedding, and that it’s “too early” to know how well the new feature will work, but that because it’s cryptographically signed, there should be no false positives.