OpenAI has made an important decision that allows European users of ChatGPT and Codex to identify their generated content more transparently. This action is in response to new EU regulations for transparency in AI-generated content.
Watermarking: OpenAI's New Tool
OpenAI plans to activate watermarking for its texts in the coming weeks. This watermark is designed in such a way that the user cannot see it and is hidden as very subtle changes in word choice. This action is taken to comply with the "EU Artificial Intelligence Act," which requires companies operating in this field to mark content generated by AI systems.
Read more: Alireza Bozorgmehr: The Strait of Hormuz Crisis Could Disrupt the Internet
OpenAI's new watermark, unlike typical markers, does not add a visible sign or phrase to the text but instead is subtly hidden in the text by selecting specific word patterns. This technology is known as "textGrain" and a technical report has also been published alongside it. This technology was developed in collaboration with researchers from the University of Pennsylvania and Yale University.
Functionality and Challenges of Watermarking
OpenAI has emphasized that this watermark can remain with the text when copied and transferred, but changes to parts of the writing may reduce the system's ability to identify its origin. Tests have shown that replacing about 10% of the words in the text with synonyms can reduce identification accuracy from about 92% to 66%. Additionally, short texts and mathematical responses make watermark identification more difficult.
Despite this technology, OpenAI has stressed that the presence or absence of a watermark alone cannot prove that a text was written by a human. In fact, very short or edited texts may complicate their identification, and this watermark does not specify the user's identity.
The Future of Watermarking in AI
Initially, access to the watermark detection tool will be limited to verified researchers and specialized organizations to assess its accuracy and responsible applications. API developers worldwide can also use this feature for some OpenAI models, but by default, this option will be turned off.
It is worth mentioning that OpenAI had previously developed watermarking technology but refrained from releasing it due to concerns about user migration to competing services. Now, with this new decision, the company aims to create greater transparency in its generated content in the European Union. This action is in line with a similar decision by Anthropic for watermarking texts generated with Claude.
Read more: Samsung has Transformed the User Experience of Phones with New Innovations · Seyed Mohsen Haji Saeed: The Future of Tourism is in Human-AI Interaction
