OpenAI announced on October 5, 2026, that it has begun deploying a new invisible watermarking system for text generated by ChatGPT and Codex. The technology, which is being rolled out specifically for users in the European Union, is designed to make AI-generated content identifiable in a machine-readable format while remaining undetectable to the human eye.
The system, titled “textGrain,” is a response to the regulatory requirements established by Article 50 of the EU AI Act. This legislation mandates that providers of AI systems ensure their outputs are marked as artificially generated. While the watermarking is mandatory for EU-based ChatGPT and Codex users, OpenAI is making the feature optional for API customers worldwide, where it remains disabled by default.
The Mechanics of textGrain
Unlike some digital watermarking methods that rely on hidden Unicode characters or metadata, textGrain is a statistical approach. It influences the selection of specific words—or “tokens”—during the text generation process. By creating a predictable pattern in the probability of word choices, OpenAI can later analyze a block of text to determine if it originated from its models.
The rollout follows a specific compliance timeline set by the European Union. While the EU AI Act officially came into force earlier this year, systems that were already on the market before August 2, 2026, have been granted until December 2, 2026, to meet these machine-readable marking obligations. OpenAI’s October rollout ensures its primary consumer and developer tools meet this deadline ahead of schedule.
Technical Limitations and Fragility
Despite its utility for provenance, OpenAI has acknowledged significant limitations regarding the robustness of textGrain. Because the watermark is embedded in the statistical distribution of words rather than the file structure itself, it is highly susceptible to modification. According to data provided by the company, the ability to detect the watermark drops sharply if the text is edited or rewritten.
In testing, a 400-token passage (roughly 300 words) showed a 92% detection rate. However, if a user replaces just 25% of the words with synonyms, the detection rate plummets to 17%. This fragility means that simple paraphrasing, translating the text into another language, or using another AI model to “clean” the output can effectively strip the watermark away.
Because of these vulnerabilities and the potential for false positives, OpenAI is restricting access to the textGrain detector. The tool is not available to the general public, students, or teachers for plagiarism checking. Instead, access is limited to approved researchers and expert organizations focused on information integrity and AI safety.
While OpenAI claims the watermarking process has minimal impact on the overall quality and creativity of the writing, the company is continuing to monitor how the constraints on token selection affect specialized tasks. For now, the mandatory implementation remains a regional strategy focused on the EU’s strict transparency standards, while the rest of the global market retains the choice to opt in via the API.
