Overview
- OpenAI announced Monday that it will enable invisible watermarking for eligible ChatGPT and Codex outputs across the European Union over the coming weeks and let API customers worldwide opt in for select models.
- The system, called textGrain, embeds a secret-key statistical signal by slightly nudging the model’s word choices so a detector can identify AI-generated text from the pattern alone.
- OpenAI’s own tests show the watermark is unreliable for short, highly constrained, translated or edited passages—for example swapping 10% of words with synonyms cut detection from about 92% to 66% and swapping 25% cut it to about 17%.
- Access to the watermark detector will be restricted initially to approved researchers and expert organisations, and OpenAI says a positive match does not reveal who wrote or prompted the text while a negative result does not prove human authorship.
- The rollout is aimed at meeting the EU AI Act’s Article 50 transparency requirement ahead of a Dec. 2 compliance deadline for earlier models and follows similar provenance moves by rivals like Anthropic, though enforcement may be limited because watermarks can be removed by editing or combining outputs.