OpenAI is introducing an invisible text-watermarking technology named textGrain to comply with machine-readable text labeling rules under the European Union AI Act. The technology embeds statistical signals into model word selections. OpenAI plans to activate textGrain for ChatGPT and Codex users within the EU over the coming weeks, while offering optional opt-in watermarking for API customers worldwide and through cloud partners such as Microsoft Azure.
Differentiated EU and Global Rollout
Unlike Anthropic's mandatory global watermarking for Claude, OpenAI is separating consumer deployment in Europe from developer access elsewhere. EU users accessing ChatGPT and Codex will see watermarking enabled by default, but global API developers can choose whether to apply the hidden mark. OpenAI also plans to release textGrain as open-source software so third parties can build upon the implementation.
Evasion Risks and Technical Accuracy
The reliability of textGrain varies according to passage length and subject matter. Internal tests using a 1 percent target false-positive rate yielded a 95 percent detection rate on 400-token psychology passages, dropping to 80 percent on 200-token samples. Detection rates were significantly lower for mathematics content due to restricted vocabulary options. Text modifications also diminish detection performance: swapping 10 percent of words with synonyms reduced detection on 400-token passages from 92 percent to 66 percent, while replacing 25 percent lowered detection to 17 percent. OpenAI notes that detection signals do not establish ownership, user identity, human authorship, or text accuracy.
Controlled Access and Benchmark Testing
OpenAI is limiting access to its textGrain detector tool during the initial phase. Selected researchers and institutions must apply case by case under the EU Code of Practice, as false positives and missed watermarks remain technical risks. Meanwhile, existing image and audio verification tools on openai.com/verify remain available to the public. Internal evaluations on OpenAI's frontier model Astra showed no significant benchmark score variations with watermarking enabled across tests such as GPQA Diamond, BrowseComp, and DeepSWE.
AI-generated editorial illustration; not a documentary photograph.
Source: OpenAI will watermark ChatGPT text in the EU but makes it optional for API users worldwide
SIGNAL TITANS
Leave a Reply