Skip to the content
AI and Cybersecurity Learning and Practice Center Test automation Open the Learning Center

NewsPolicy

OpenAI starts watermarking ChatGPT and Codex text in the EU, and publishes how easily the mark is lost

OpenAI said on October 5 that it is adding an invisible watermark it calls textGrain to text ChatGPT and Codex generate for users in the European Union, rolling out over the coming weeks, because the EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way. The watermark slightly changes the model's word choices to leave a statistical pattern a detector can find later; API developers anywhere can opt in. OpenAI published the limits: at a 1% false positive rate its detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages, and replacing 10% of the words with synonyms cut detection from about 92% to 66%. Access to the detector is limited at first to approved researchers and expert organizations, and OpenAI notes that the absence of a detected watermark does not prove a human wrote the text.

Why it matters

Labeling AI text is now law in the EU, and the provider's own numbers show how fragile the label is: a light paraphrase removes much of the signal. Treat a detected watermark as a strong hint and a missing one as no evidence at all, in a phishing email or anywhere else.

Sources

Read it with the rest of the news in the Learning Center