OpenAI Adds Text Watermarks to ChatGPT in EU
Original: OpenAI will start watermarking ChatGPT’s text in the EU
Why This Matters
EU AI Act enforcement is pushing watermarking from optional to mandatory across major AI platforms.
OpenAI will embed invisible watermarks in ChatGPT and Codex text outputs for EU users over the coming weeks, complying with EU AI Act transparency rules that took effect August 2, 2026.
OpenAI announced Monday it will roll out text watermarking for ChatGPT and Codex across all plans in the EU, with API access for developers worldwide available immediately—though off by default globally. The system, called textGrain and co-developed with researchers from the University of Pennsylvania and Yale, works by subtly steering the model's word choices to embed a hidden pattern. No visible symbol is added; the signal lives in the word selection itself, so it persists when text is copied. A detector using the text and a secret key can identify AI-generated content.
OpenAI acknowledges meaningful limits. In testing, replacing just 10% of words with synonyms dropped detection accuracy from roughly 92% to 66%. Short passages, math answers, and translated text are harder to flag. Detector access is currently restricted to approved researchers and expert organizations. The company stressed that a missing watermark does not confirm human authorship—text might be too short, too edited, or generated by a rival system. OpenAI also confirmed the watermark does not identify individual users, and no measurable performance drop was observed with it enabled.
The move follows Anthropic's announcement two months ago that it would watermark Claude outputs globally—a decision that sparked user complaints. OpenAI had developed a watermark earlier but delayed release over concerns about user behavior.