- Messages
- 4,002
- Likes
- 1,873
- Points
- 2,130
OpenAI announced the implementation of hidden text watermarking technology in its ChatGPT and Codex services in the European Union.
This innovation is related to the entry into force of the European Artificial Intelligence Act, which requires creators of generative models to make generated text recognizable to algorithms. In the coming weeks, invisible watermarks will appear in responses from users of all plans in EU countries.
This option will be made available as an option for API clients worldwide, and access to the watermark detector will initially be limited to vetted researchers and specialized organizations. The company plans to publish the source code for the technology later.
The proprietary textGrain method injects a subtle statistical pattern into the neural network's word selection process. According to internal benchmarks on the flagship Astra model, this intervention has virtually no impact on the quality and accuracy of generation.
However, the developers acknowledge significant technical limitations of the method. The detector is less effective at recognizing short fragments and precise disciplines like mathematics, where the choice of wording is severely limited.
Furthermore, the watermark is destroyed by manual editing: replacing 10% of the words with synonyms in a 400-token text drops recognition accuracy from 92% to 66%, while replacing a quarter of the words reduces the rate to 17%. OpenAI also emphasized that the watermark does not contain user data, does not reveal the entered prompts, and does not confirm the factual authenticity of the text.