OpenAI will watermark ChatGPT text in the EU but makes it optional for API users worldwide
OpenAI is adding invisible textGrain watermarks to ChatGPT in the EU, but unlike Anthropic, it will let API customers worldwide opt out. Tests show detection rates as high as 95 percent, dropping to 17 percent when a quarter of the words are replaced.
OpenAI is implementing invisible textGrain watermarks in ChatGPT for EU users, while allowing API customers globally to opt out, in response to EU regulations.
What happened
OpenAI is introducing invisible textGrain watermarks for ChatGPT in the EU to comply with the EU AI Act. The watermarks are optional for API users worldwide, and detection rates depend on factors like text length and editing. OpenAI also plans to open-source the technology and limit detector access initially.
Why it matters
This development is significant for AI developers and businesses as it addresses regulatory compliance in the EU while offering flexibility for global API users. The effectiveness of watermarks against edits and the open-source release could influence future AI transparency and security measures.
- OpenAI's watermarking approach balances EU compliance with global API flexibility.
- Watermark detection effectiveness varies, highlighting the need for robust verification methods.
- Open-source release of textGrain could foster innovation and transparency in AI watermarking.
The bigger picture
The implementation of textGrain watermarks may set a precedent for AI transparency and regulatory compliance. The optional nature for API users could affect adoption rates, while the open-source release might encourage broader development and scrutiny of watermarking technologies.
OpenAI is introducing textGrain watermarks for ChatGPT in the EU, offering optional API access globally. The technology's effectiveness depends on various factors, and its open-source release may influence future AI transparency efforts.
Technical details
- textGrain: An invisible statistical signal embedded in the model's word choices to watermark text.
- Detection rates: Vary based on text length, content type (e.g., psychology vs. math), and editing (e.g., replacing 10% or 25% of words).
- Open-source release: OpenAI plans to make textGrain available for others to build upon.
- Detector access: Initially restricted to selected researchers and organizations under the EU Code of Practice.
FAQ
What is textGrain, and how does it work?
TextGrain is an invisible watermarking technology that embeds a statistical signal in the model's word choices. It works by altering the model's output to include a detectable signal without affecting the text's readability or quality.
Why is OpenAI making watermarking optional for API users?
OpenAI allows API users worldwide to opt out of watermarking to provide flexibility, unlike Anthropic's mandatory approach. This may cater to different user needs and regulatory environments.
How effective are textGrain watermarks against edits?
Editing text, such as replacing 10% or 25% of words with synonyms, significantly reduces detection rates, making the watermark easier to evade. This could impact the reliability of watermarking as a verification method.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by The Decoder on Oct 5, 2026.
Comments
Sign in to join the discussion.