In response to a report suggesting OpenAI possesses a highly accurate tool to detect ChatGPT-generated text, the company has provided insights into its ongoing research on text watermarking. While the company has developed a promising method, it has refrained from public release due to concerns about its effectiveness and potential biases.
OpenAI acknowledged that text watermarking is one of several techniques they’ve explored to determine the origin of text. While the method has shown high accuracy in certain scenarios, it faces challenges when dealing with manipulated text or when used by non-native English speakers. The company expressed concerns that such a tool could inadvertently stigmatize AI usage among certain groups.
Given these risks, OpenAI has prioritized the development of authentication tools for audiovisual content. The company emphasized its cautious approach to text provenance, citing the complexities involved and the potential implications for the broader AI ecosystem.