OpenAI announced on October 5, 2026, that it will introduce text watermarking for eligible ChatGPT and Codex outputs within the European Union in the coming weeks. This initiative is a direct response to Article 50(2) of the EU AI Act, which requires generative AI providers to ensure AI-generated or manipulated content is machine-readable and detectable as artificial.
The company's watermarking technology, named textGrain, embeds a statistical signal into the model's word choices. This signal is imperceptible to human readers but can be detected by a specialized tool. OpenAI stated that textGrain's performance "matched or exceeded" other methods, including Google DeepMind's SynthID for text.
For API customers globally, the option to opt in to text watermarking for select models is available starting October 5, 2026, though it remains off by default.
OpenAI acknowledges several limitations of its text watermarking technology. The company's detector can produce both false negatives and false positives, particularly with shorter texts or content in domains with limited word choice, such as mathematics. For instance, in psychology-related content, the detector identified watermarks in approximately 80% of 200-token passages and about 95% of 400-token passages, at a target false-positive rate of 1%. However, detection rates were significantly lower for mathematics content.
Editing also substantially weakens the watermark signal. OpenAI's evaluations of 400-token passages showed that replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words further decreased detection to approximately 17%.
OpenAI emphasized that a watermark does not measure human contribution, establish ownership or responsibility, identify the user, or verify accuracy. The absence of a detected watermark does not prove human authorship, as text can be too short, edited, translated, or originate from an unsupported model or another company's tools.
The company plans to make textGrain available as open-source technology to allow others to build upon and improve the approach. OpenAI is also opening applications for access to its text watermark detector, initially limiting access to approved researchers and expert organizations to evaluate and refine the technology. This phased approach reflects both the EU AI Act's requirements and the current limitations of the technology.