OpenAI has announced the introduction of invisible text watermarks for ChatGPT and Codex results. The new technology, called textGrain, is designed to help identify content created by users from the European Union using artificial intelligence.
OpenAI has announced the introduction of invisible text watermarks for ChatGPT and Codex results. The new technology, called textGrain, is designed to help identify content created by users from the European Union using artificial intelligence.
This is stated in the OpenAI blog. The decision was made in response to the requirements of the EU Artificial Intelligence Act (EU AI Act), which obliges providers of generative technologies to ensure the possibility of machine-readable detection of generated text.
The new textGrain system works by adding an invisible statistical signal to the model’s word selection. The company has created a special detector to check for this signal. Initially, only approved researchers and expert organizations will have access to it, and later the technology is planned to be released as open source.
Watermarks will be automatically added to ChatGPT and Codex responses for users in the European Union in the coming weeks. For developers using the OpenAI API worldwide, the feature will be available on an optional basis, with the option to manually enable it.
OpenAI assures that the implementation of watermarks does not affect the quality of generated responses and the performance of models, in particular Astra. At the same time, the company notes the existing technical limitations:
Short and highly specialized texts (e.g., mathematics) are much more difficult to detect due to the limited choice of synonyms.
Manual editing significantly weakens the sign: replacing 25% of words with synonyms reduces the detection rate from 92% to 17%.
"A text watermark does not measure human contribution, establish ownership, identify the user, or confirm the authenticity of information," the company emphasizes, urging that the technology be considered only as one of the additional verification tools.
OpenAI has previously implemented tagging for multimedia content. The company uses the open C2PA standard and SynthID technology for images and audio, and also provides special file verification tools on the website and through the Content Provenance API.
