OpenAI’s text watermarking method for detecting AI-generated text

An analysis of OpenAI’s proposed text watermarking approach, including provenance value, circumvention risk, and possible effects on non-English writers.

• The text watermarking method involves making small changes to how ChatGPT selects words and creating an invisible watermark in the writing that can be detected later by a separate tool.
• OpenAI is cautious about the potential negative impacts of text watermarking, including susceptibility to circumvention by bad actors and the potential to disproportionately impact groups like non-English speakers.

OUR TAKE
OpenAI’s text watermarking method plays a significant role in addressing this issue by focusing on detecting writing specifically from ChatGPT, offering a targeted approach to identifying potentially fraudulent content. However, this method also raises important considerations regarding its susceptibility to circumvention by malicious actors and the potential impact on certain groups, such as non-English speakers.
-Lia XU, BTW reporter

What happened

According to The Wall Street Journal, OpenAI has developed a text watermarking method to detect AI-generated text, specifically focusing on catching students who cheat by using ChatGPT to write their assignments. This method involves making subtle changes to how ChatGPT selects words, creating an invisible watermark in the writing that can be detected later by a separate tool.

However, an OpenAI spokesperson said, “It’s taking a ‘deliberate approach’ to due to the complexities involved and its likely impact on the broader ecosystem beyond OpenAI. Because it has important risks we’re weighing while we research alternatives, including susceptibility to circumvention by bad actors and the potential to disproportionately impact groups like non-English speakers”.

OpenAI also updated its blog to say that while its text watermarking can detect some AI-generated content well, it can be easily bypassed by bad actors, and it may discourage non-native English speakers from using AI as a helpful writing tool.

Also read: OpenAI’s next model to undergo safety checks by the U.S. Government

Also read: OpenAI supports legislation to shape the future of AI

Why it’s important

This text watermarking method can help educational institutions maintain academic integrity by identifying instances of students using AI to cheat on assignments. It can deter academic dishonesty and promote fair evaluation practices. Focusing solely on detecting writing from ChatGPT sets this method apart from others, ensuring a more precise identification of AI-generated text. This targeted approach enhances the tool’s effectiveness in catching potential instances of cheating.

However, it’s also crucial for ethical deployment to consider the risks of circumvention by bad actors and potential impacts on specific user groups like non-English speakers. They need to seek a balanced approach to its implementation. OpenAI’s deliberate strategy for researching and weighing the risks associated with the text watermarking method demonstrates a commitment to developing responsible AI technologies. This influence ensures that the tool’s deployment aligns with ethical standards and minimises negative consequences on various user groups and the ecosystem.