OpenAI to add invisible text watermarking in ChatGPT and Codex in the EU
In short
OpenAI said it will add an invisible watermark, called textGrain, to text generated by ChatGPT and Codex in the European Union, rolling out over the coming weeks to eligible users on all plans. API customers worldwide can opt in for select models starting today, off by default, while detector access goes initially to approved researchers and expert organizations. OpenAI said the move is to comply with the EU AI Act's transparency rules; Anthropic announced a similar system earlier.
Read the full story 2 min read
OpenAI said it will add an invisible, machine-readable watermark to text generated by ChatGPT and Codex in the European Union, a move the company described as complying with the EU AI Act's transparency rules. The company said the watermark will roll out over the coming weeks to eligible ChatGPT and Codex users on all plans, but only in the EU, and that it is not making text watermarking a global default at launch. [ 1 , 2 , 3 ]
The watermark, called textGrain, works by subtly shaping the model's word choices to leave a pattern that readers cannot see but a detector can pick up, according to the reports. OpenAI said API customers anywhere in the world can opt in to watermarked outputs for select models starting today, with the feature off by default. Detector access will initially be granted to approved researchers and expert organizations on a case-by-case basis; OpenAI said the tool reports whether it detects an OpenAI watermark without identifying the user or revealing prompts or conversations. [ 1 , 2 , 3 , 4 ]
OpenAI published a technical report alongside the announcement, co-written with researchers from the University of Pennsylvania and Yale, TechCrunch reported. OpenAI said it saw no meaningful change in model performance with the watermark switched on, and that because the signal lives in the words themselves it travels with text when copied and pasted. [ 2 , 3 ]
OpenAI also set out limits. According to TechCrunch, one test found that replacing 10 percent of words with synonyms dropped detection from about 92 percent to 66 percent; Engadget reported around an 80 percent success rate for shorter text. OpenAI said short passages, mathematics and translated text are harder to detect, and that the watermark “does not guarantee reliable detection,” nor does it verify accuracy, establish ownership, measure human contribution or prove human authorship. [ 1 , 2 , 3 ]
The rules behind the change are the EU AI Act's transparency requirements, which Engadget said took effect on August 2. Article 50 requires providers of generative AI systems to make text outputs identifiable in a machine-readable way; new entrants are already covered, while pre-existing providers such as OpenAI, Anthropic, Microsoft, Google and Meta have until December 2 to comply. Anthropic announced a similar text watermarking move earlier, also citing the EU rules, and said at the time that other major providers were implementing marking systems as well. [ 1 , 2 , 4 ]
Why it matters
The EU AI Act's transparency rules now push major AI providers to mark machine-generated text, and OpenAI's rollout sets up textGrain as an alternative to Anthropic's and Google DeepMind's approaches. Because detector access is limited at first and OpenAI itself says detection is not reliable, the practical effect for readers and institutions outside the EU is likely to be gradual.
Key facts
- OpenAI said it will add an invisible text watermark to ChatGPT and Codex output for eligible users in the European Union over the coming weeks, on all plans. [ 1 , 3 ]
- API customers globally can opt in to watermarked outputs for select models starting today; the feature is off by default. [ 1 , 3 , 4 ]
- OpenAI said it is not making text watermarking a global default at launch. [ 1 , 3 ]
- The system is called textGrain and works by subtly shaping the model's word choices so a detector can identify a pattern. [ 2 , 3 ]
- Detector access will initially be limited to approved researchers and expert organizations; the detector does not identify users or reveal prompts or conversations. [ 1 , 2 , 3 ]
- OpenAI said short passages, math answers and translated text are harder to detect, and that its watermark “does not guarantee reliable detection.” [ 1 , 2 , 3 ]
- The EU AI Act's transparency rules took effect on August 2, with pre-existing providers required to comply by December 2. [ 2 ]
Confirmed by several sources
- OpenAI announced it will watermark ChatGPT and Codex text output in the EU to comply with the EU AI Act. [ 1 , 2 , 3 , 4 ]
- OpenAI's watermarking system is called textGrain and works by subtly shaping word choices to leave a machine-detectable pattern. [ 2 , 3 ]
- API customers worldwide can opt in to watermarking for select models, with the feature off by default. [ 1 , 3 , 4 ]
- Detector access is initially limited to approved researchers and expert organizations, and does not identify users or reveal prompts. [ 1 , 2 , 3 ]
- Anthropic previously announced a comparable text watermarking move tied to the EU AI Act. [ 1 , 2 , 4 ]
- OpenAI said the watermark does not verify accuracy, ownership, or human authorship and does not guarantee reliable detection. [ 1 , 2 , 3 ]
Still unclear
- Which models API customers will be able to watermark. Engadget reported OpenAI did not specify which models would be available to the general global public.
- Whether the watermark will eventually be applied outside the EU. OpenAI said it is not a global default “at launch” and described the EU rollout as a regional approach to learn from real-world use, leaving the longer-term scope unstated.
- How reliable detection will be in everyday use. OpenAI published benchmark scores and an example where replacing 10 percent of words with synonyms cut detection from about 92 percent to 66 percent, but said performance under ideal conditions does not guarantee reliable detection.
- The relationship between OpenAI's textGrain and Google DeepMind's SynthID for text. One outlet said textGrain matched or exceeded other approaches such as SynthID, which it described as the basis for Anthropic's watermarking; other outlets did not make that link.
What local media are saying
Timeline, local time
- The Verge reports OpenAI is rolling out its textGrain watermark to ChatGPT and Codex in the EU and opening opt-in watermarking for API customers. [ 1 ]
- Engadget reports OpenAI will add a digital watermark to text and code generated in the EU to comply with new transparency rules. [ 2 ]
- TechCrunch reports OpenAI will begin watermarking ChatGPT text in the EU, citing the company's blog post. [ 3 ]
- 9to5Mac reports OpenAI detailed its watermarking system for ChatGPT, Codex and the API, including detector applications opening for approved researchers. [ 4 ]