The Watermarking Dilemma: OpenAI’s Internal Struggle Over ChatGPT Text Detection

OpenAI has reportedly developed a system for watermarking ChatGPT-generated text, which has been ready for about a year, according to The Wall Street Journal. However, internal debates have stalled its release. The company faces a dilemma: deciding between doing the responsible thing and potentially harming its bottom line.

The watermarking technique involves adjusting the model’s predictions to create a detectable pattern in the text—a concept similar to Google’s Gemini’s text watermarking. This could be a significant development for educators aiming to prevent students from using AI for assignments. According to The Journal, watermarking doesn’t compromise the chatbot’s text quality. Moreover, a company survey revealed significant public support for an AI detection tool.

Following The Journal’s report, OpenAI confirmed its watermarking efforts in a blog post highlighted by TechCrunch. The company claims its method is “99.9% effective” and resistant to paraphrasing tampering. However, OpenAI also expressed concerns about potentially stigmatising AI tools for non-native speakers.

Interestingly, nearly 30% of surveyed ChatGPT users indicated they’d use the tool-less if watermarking were implemented, a significant figure that adds to OpenAI’s hesitation. Despite some employees advocating for the watermarking’s effectiveness, the company is considering less controversial, albeit unproven, methods. In the recent blog update, OpenAI mentioned it’s exploring embedding metadata—a cryptographically signed method promising no false positives. However, it remains “too early” to determine its efficacy.