OpenAI would not watermark ChatGPT texts because it would discourage users.

According to the Wall Street Journal, OpenAI has had a system for watermarking content created by ChatGPT for a year, as well as a tool for detecting the watermark. A boon for teachers who do not want to use ChatGPT, but a deterrent for users whose exam papers are written by ChatGPT.

This puts the company in a difficult position: take responsibility or risk damaging profits. According to a study by OpenAI, 30 percent of users would consult ChatGPT less if it had a watermark because they would feel caught out.

In response to the Journal’s story, OpenAI updated a blog post spotted by TechCrunch, saying the company is in the early stages of a text watermarking method and continues to explore alternatives.

Text watermark

OpenAI’s text watermarking would only focus on detecting texts created by ChatGPT and not texts created by other models. It does this by making slight changes to the way ChatGPT selects words. This creates an invisible watermark in the text that can be retrieved using a separate tool.

“The text watermarking method we developed is technically promising, but it carries significant risks that we are weighing as we explore alternatives, including its vulnerability to circumvention by malicious actors and the potential to disproportionately impact groups such as non-English speakers,” an OpenAI spokesperson told TechCrunch.

Not rentable?

OpenAI is also concerned about the impact of the detection tool on the use of its software. According to its own research, almost 30 percent would use the software less if a watermark were implemented. This can create a feeling of being “caught” especially in an academic environment.

According to The Journal, OpenAI found that the watermarks had no impact on the quality of ChatGPT’s output. “People around the world support the idea of ​​an AI recognition tool by a margin of four to one,” The Journal reports.

After The Journal broke the story about the existence of OpenAI’s text watermarks, the company updated a May blog post spotted by TechCrunch, saying that the company is “in the early stages” of exploring metadata embedding and that it’s too early to say how well it will work.