Recently, OpenAI has been embroiled in internal discussions over watermarking technology. According to a report by The Wall Street Journal, the company has already developed a watermarking technology to identify text generated by ChatGPT and has also prepared a detection tool. However, whether to bring this technology to market remains a point of contention within the company.
From a certain perspective, introducing watermarking technology seems like a responsible move. This technology can create a detectable pattern by adjusting the model's predictions of words and phrases. Although this sounds somewhat complex, its purpose is clear—to help teachers prevent students from plagiarizing AI-generated assignments. The report indicates that this watermark does not affect the quality of the chatbot's text. Moreover, a survey by OpenAI found that globally, the number of people supporting AI detection tools is four times that of opponents.
However, the implementation of watermarking technology is not straightforward. OpenAI's blog update confirms that the company has developed watermarking technology and states that this method is highly accurate, claiming to be "99.9% effective." Additionally, this watermark has a certain resistance to "tampering," such as simple rephrasing. However, OpenAI also points out that rewriting with other models can easily bypass this watermark, which causes them concern.
Furthermore, OpenAI considers that the watermark might stigmatize some users, especially non-native speakers. Therefore, despite some employees believing the watermark is effective, a survey shows that nearly 30% of ChatGPT users say they might reduce their usage if the watermark is implemented. Hence, some employees suggest exploring "relatively less controversial" methods, although the effectiveness of these methods has not yet been validated.
In today's blog update, OpenAI mentions they are in the "early stages" of exploring methods to embed metadata and states that it is still too early to judge its effectiveness. However, they point out that since this metadata is cryptographically signed, it can avoid misjudgments.
Key Points:
✅ OpenAI is divided on the rollout of watermarking technology; whether to release it is still under discussion.
📊 A global survey shows that most people support AI detection tools, but users are concerned about the impact of watermarks on usage.
🔒 OpenAI is considering using embedded metadata methods to strive for a balance between technology and user experience.