OpenAI launches textGrain watermark to ID AI-generated text in EU
OpenAI rolled out textGrain watermarking to detect AI-generated text for EU compliance.
Why it matters: Legal professionals and tech firms must understand how watermarking affects content verification and legal compliance, especially around intellectual property and misinformation risks.
- textGrain embeds an invisible statistical signal in AI-generated text to mark OpenAI content.
- Detector finds watermarks in about 80% of 200-token and 95% of 400-token texts at 1% false positive rate.
- Editing text by replacing words with synonyms weakens watermark detection significantly.
- OpenAI will deploy textGrain to ChatGPT and Codex users in the EU soon; API users globally can opt in now.
OpenAI has introduced textGrain, a watermarking technique that embeds an invisible statistical signal into the word choices of AI-generated text. This innovation responds to the European Union's AI Act, effective August 2, 2026, which requires AI companies to make AI-generated content identifiable in a machine-readable format. The watermark helps detectors assess whether a passage contains OpenAI's watermark, aiding transparency and authenticity verification.
In controlled evaluations, OpenAI's detector identified watermarks in about 80% of 200-token passages and approximately 95% of 400-token passages at a low false positive rate of 1% — meaning very few cases falsely flagged human-written content as AI-generated. However, the watermark signal is fragile; editing the text by replacing words with synonyms can reduce detection rates sharply, from 92% to 66% when 10% of words are replaced and down to 17% at 25% replacement.
The watermark feature is set to roll out over the coming weeks to eligible ChatGPT and Codex users in the European Union. API customers outside the EU can already opt-in to watermarking on select models. While textGrain enhances the ability to detect AI-generated content for compliance and intellectual property purposes, its effectiveness diminishes with text modifications, posing challenges for comprehensive enforcement.
This launch is part of OpenAI's broader efforts to meet upcoming regulatory demands and address concerns about misinformation and content provenance. Legal teams and tech companies should monitor how watermarking evolves as both a compliance tool and a factor in content authenticity assessments.
By the numbers:
- 80% detection rate — for 200-token passages at 1% false positive rate
- 95% detection rate — for 400-token passages at 1% false positive rate
- 17% detection rate — when 25% of words replaced with synonyms, reducing watermark signal
Yes, but: The watermark can be weakened or bypassed by editing the text, which significantly reduces detection accuracy.
What's next: OpenAI plans EU-wide rollout of watermarking for ChatGPT and Codex users over the next weeks; global API opt-in already available.