OpenAI announced in a blog post on Monday that it will begin adding invisible watermarks to texts generated by ChatGPT and Codex in the European Union. The decision aims to comply with the transparency rules of the EU Artificial Intelligence Act, which entered into force on August 2 and requires AI content to be marked in a way that is identifiable by other systems.
The watermark will be offered in the coming weeks only to eligible ChatGPT and Codex users in the EU, across all plans. API developers anywhere in the world, meanwhile, can enable it for certain models starting today; the feature is off by default and is not being made the global default.
The watermark is not an actual symbol; it leaves a pattern that the reader cannot see but a detector can catch by subtly shaping the model’s word choices. Because it resides within the words, it travels along with copy-paste. OpenAI stated that the watermark does not identify the user and that it has not observed a meaningful change in model performance. The company also published the technical report of the “textGrain” method, co-authored with researchers from the University of Pennsylvania and Yale University, which ranks next-word predictions with a secret key.
OpenAI’s tests show that the watermark can be removed through editing: replacing 10% of the words with synonyms dropped detection from about 92% to 66%. Those that are harder to detect:
- Short passages
- Math answers
- Translated texts
Because of these limitations, detector access will be given only to approved researchers and expert organizations. The company also warned that the absence of a watermark “does not prove human authorship.”
The announcement came two months after Anthropic announced that it would watermark Claude texts worldwide; that decision had drawn backlash from some Claude users. OpenAI had previously developed a text watermark as well, but shelved it over concerns that users would switch to competitors (The Wall Street Journal, 2024). Anthropic, Google, Meta, Microsoft, and OpenAI are among the companies that have pledged to comply with the EU’s implementation rules on AI content.
Background
OpenAI is not a new name in the FikirPilot archive: we have published 105 news stories mentioning this name in the last 90 days; the most recent is dated October 4, 2026.