OpenAI makes text watermarks available via the API and will restrict detector access to experts
Text watermarks from OpenAI are now available as an option via the API worldwide for selected models. For now, only approved experts will receive access to the detector. The rollout in ChatGPT and Codex will begin in the EU in the coming weeks.
OpenAI has made optional watermarking of text outputs available to API customers worldwide for selected models; it is disabled by default. Researchers and specialist organizations can apply for access to the detector, with the company initially approving applications individually. The detector will not be publicly available at launch because of the risk of missed watermarks and false positives.
In ChatGPT and Codex, the rollout is expected to take place over the coming weeks for eligible users on all plans, exclusively in the EU. According to OpenAI, this step is intended to ensure compliance with the transparency requirements of the EU AI Act. The textGrain method creates an invisible statistical signal through word choice, so the watermark remains part of the text when it is copied.
According to the company, detection reliability has significant limitations. In an internal test of passages 400 tokens long, replacing 10 % of words with synonyms reduced the detection success rate from approximately 92 % to 66 %; replacing a quarter of the words reduced it to 17 %. The company says short passages, mathematical answers and translated text are also harder to detect. The absence of a watermark does not prove human authorship, and a detected watermark does not establish the extent of human involvement, ownership or the accuracy of the text.
According to benchmarks from OpenAI, outputs with and without watermarks perform similarly; however, these results alone do not demonstrate equal writing quality. The company has published a technical report on the textGrain method and plans to release the technology as open source. Watermarking is expected to become available through cloud partner services in the coming weeks.
Why it matters
A watermark may help identify that text was generated or processed by a system from OpenAI. However, detection limitations mean it does not provide certainty about authorship. Organizations using the API can now enable watermarking according to their transparency needs, while access to verification remains limited.
What was added since the original report
Verified updates
-
The introduction of watermarks is intended to ensure compliance with the transparency requirements of the EU AI Act.; According to a test by OpenAI, replacing 10 % of words with synonyms reduced detection from approximately 92 % to 66 %.; According to OpenAI, the absence of a watermark does not prove human authorship.
- The introduction of watermarks is intended to ensure compliance with the transparency requirements of the EU AI Act.
- According to a test by OpenAI, replacing 10 % of words with synonyms reduced detection from approximately 92 % to 66 %.
- According to OpenAI, the absence of a watermark does not prove human authorship.
-
Optional watermarks are already available to API customers worldwide for selected models.; The EU rollout will include eligible users on all plans.; Access to the detector will initially be approved individually for researchers and specialist organizations.; According to benchmarks from OpenAI, text with and without watermarks performs similarly.
- Optional watermarks are already available to API customers worldwide for selected models.
- The EU rollout will include eligible users on all plans.
- Access to the detector will initially be approved individually for researchers and specialist organizations.
- According to benchmarks from OpenAI, text with and without watermarks performs similarly.
Two audiences, two different impacts
What this means
For individuals
Users of ChatGPT and Codex in the EU must expect that their outputs may carry a watermark even after being copied. Its presence alone does not indicate how much of their own work or creativity they contributed to the text.
For a business
Companies generating text via the API can incorporate optional watermarking into their content publishing process for supported models. However, they will not have access to a generally available detector at launch, limiting their ability to verify watermarks independently.
Risks and complianceCheck the original
Event sources
confirmed by 2 independent sources · 2 publishers, 2 independent. We count feeds from the same owner only once.