OpenAI cites EU AI Act requirements as the reason for text watermarks in the EU
According to OpenAI, the introduction of text watermarks is intended to ensure compliance with EU AI Act transparency requirements. They will appear in ChatGPT and Codex in the EU over the coming weeks; through the API, they are already optionally available worldwide for selected models.
OpenAI links its planned introduction of text watermarks to compliance with EU AI Act requirements for machine-readable labelling of AI-generated content. Over the coming weeks, it is to enable watermarking for eligible users of ChatGPT and Codex across all plans in the EU. At launch, watermarking will not be the default setting worldwide.
OpenAI API customers can already opt in to watermarking worldwide for selected models; it is disabled by default. Access to the detector will initially be approved on a case-by-case basis for researchers and specialist organisations. The company will not make it available to the public at launch because of the risk of missed watermarks and false positives.
The textGrain technology adjusts word choice so that the text contains an invisible statistical signal that persists even when copied. According to OpenAI, the watermark does not identify the user, and internal benchmarks showed similar model performance with and without watermarking. The company has published a technical report and plans to release the technology as open source.
According to OpenAI, the method does not guarantee reliable detection. In an internal test of passages containing 400 tokens, 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 %. Detection is also more difficult for short passages, mathematical answers and translated text. According to the company, the absence of a watermark does not prove human authorship.
Why it matters
A watermark may indicate that a system from OpenAI generated or processed part of a text. However, it does not verify the accuracy of the content, ownership or the extent of human contribution. It therefore remains a limited signal for assessing the origin of a text, and its detector will also be unavailable to the public at launch.
What was added since the original report
Verified updates
-
OpenAI links the introduction of watermarks in the EU to compliance with EU AI Act requirements.
-
The introduction of watermarks is intended to ensure compliance with EU AI Act transparency requirements.; According to an OpenAI test, 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 EU AI Act transparency requirements.
- According to an OpenAI test, 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 across all plans.; Access to the detector will initially be approved on a case-by-case basis for researchers and specialist organisations.; According to OpenAI benchmarks, texts with and without watermarks show similar performance.
- Optional watermarks are already available to API customers worldwide for selected models.
- The EU rollout will include eligible users across all plans.
- Access to the detector will initially be approved on a case-by-case basis for researchers and specialist organisations.
- According to OpenAI benchmarks, texts with and without watermarks show similar performance.
Two audiences, two different impacts
What this means
For individuals
Texts from eligible users of ChatGPT and Codex in the EU are to carry a watermark even after being copied elsewhere. Detecting it, however, does not reveal how much the user's own decisions or edits contributed to the text.
For a business
Companies using selected models through the OpenAI API can decide whether to label outputs based on their transparency requirements. Enabling watermarks is optional and available worldwide.
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.