OpenAI has responded to a recent Wall Street Journal report claiming the company has developed a tool for accurately identifying essays written by ChatGPT. Although this tool is fully developed, its release has been delayed due to ongoing internal discussions about its potential consequences.
In an update to a previous blog post, OpenAI revealed that it has been working on text watermarking and other methods to trace the origins of text but has not yet finalized or deployed these technologies.
OpenAI has been exploring various methods to trace the origin of text, including watermarking, classifiers, and metadata. The company reported that while watermarking has shown high accuracy in some contexts, it struggles with certain types of tampering.
OpenAI Delays Release of Essay Detection Tool Amid Internal Debates Over Ethical Implications
Techniques such as translation, rewording, or inserting and removing special characters can undermine its effectiveness. Additionally, the potential negative impact of watermarking on some groups, particularly non-native English speakers, is a concern.
The company has been cautious about releasing the watermarking tool, reflecting on the broader implications of such technology. OpenAI has also prioritized developing tools for verifying audiovisual content over text authentication. The deliberate approach is driven by the complexities and potential ecosystem-wide effects that text provenance tools could have.
In its statement, OpenAI emphasized the importance of considering the broader impact of text detection technologies. The company aims to balance innovation with responsibility, ensuring that any tool it develops does not disproportionately disadvantage or stigmatize certain users.
OpenAI’s careful and measured approach to text detection reflects its commitment to addressing both technical challenges and ethical considerations in the deployment of AI tools. The company continues to research and evaluate various solutions while being mindful of their broader consequences.
