OpenAI will begin automatically watermarking ChatGPT outputs in the European Union starting in the coming weeks — the first time the company has enforced AI content labeling by default in any region. The move is a direct response to the EU AI Act, which took effect in August and mandates detectable AI disclosure for high-impact generative systems.
Outside the EU, watermarking remains optional: users elsewhere won’t see it unless they manually enable it, and developers using the OpenAI API won’t get it turned on automatically at all. The watermark, called textGrain, is proprietary and works by subtly biasing word choices, patterns invisible to readers but detectable with a key and a specialized tool.
OpenAI says it’s sharing that detector with a limited group of researchers and institutions, with others needing to apply for access. In testing, textGrain correctly identified watermarked text 92% of the time under ideal conditions. Long, untranslated prompts. But tampering breaks it fast: swapping just 10% of the output drops detection to ~63%.
Be the first to comment
Join the argument. No password, just your email or a passkey.