OpenAI is introducing an invisible watermark to text generated by ChatGPT and Codex for eligible users across all plans within the European Union. The rollout, occurring over the coming weeks, responds to the EU AI Act, which mandates that AI-generated text be machine-readable. To comply, OpenAI is implementing a system called textGrain.
API customers worldwide can opt in to activate watermarked text outputs for select models starting today, though the feature remains off by default. OpenAI is not making text watermarking a global default at launch, keeping the mandatory rollout limited to eligible users in the EU. The company stated that this regional approach allows it to learn from feedback and real-world use.
OpenAI is also coordinating with cloud partners to make watermarking available for model outputs accessed through their services in the coming weeks.
TextGrain Embeds Statistical Signals
TextGrain embeds a statistical signal directly into the model’s word choices during the generation process, rather than using hidden characters or spaces. This creates a hidden pattern that a separate detector can identify. OpenAI reported that this method matched or outperformed other tested approaches, including Google’s SynthID for text.
The system does not noticeably impact output quality. On GPQA Diamond, benchmark scores for the GPT-6 Astra model were 94.44% without the watermark versus 93.94% with it.
OpenAI acknowledged limitations regarding text length and editing. At a target false positive rate of 1%, the detector identified watermarks in approximately 80% of 200-token passages (roughly 150 words) and about 95% of 400-token passages (roughly 300 words) for subjects like psychology. Detection rates were substantially lower for mathematics due to less flexible word choices.
- Editing and Synonyms: Substituting words with synonyms weakens the signal. In tests with 400-token English passages, replacing 10% of words with synonyms dropped detection rates from approximately 92% to 66%. Replacing 25% of the words further reduced detection to 17%.
A detected watermark does not measure human contribution, establish ownership, verify accuracy, or identify the specific user, prompt, or account. OpenAI cautioned that the absence of a watermark does not prove a human wrote the text, as content may be too short, translated, or heavily edited.
:quality(82)/cbfbd69459c94116b66e3e7bc35677cd "Photo: nl.beincrypto.com")
OpenAI Limits Detector Access
Approved researchers and expert organizations can apply for access starting today. OpenAI is not releasing the detector to the general public due to the risk of false positives and missed watermarks. These applications will be granted on a case-by-case basis in accordance with the EU’s Code of Practice to help evaluate the tool’s reliability.
The detector will only report if an OpenAI watermark is present without revealing user identities or conversations.
Anthropic began marking text from supported Claude models worldwide in August, and Google has utilized SynthID for text and visual markings. While OpenAI’s tools for verifying audio and images remain public, the text detector’s limited release reflects the company’s view that the technology is not yet fully mature.
OpenAI plans to expand detector access once it believes results can be interpreted responsibly
.
