Stories
OpenAI Introduces Invisible Text Watermarking in EU and Opens Detector Access
- •OpenAI is rolling out an invisible watermark for ChatGPT and Codex text outputs in the European Union to comply with the EU AI Act.
- •Global API customers can optionally enable text watermarking for select models, with the feature switched off by default.
- •The textGrain technology adds a statistical signal to word choices, though detection effectiveness drops with short texts, edits, or mathematics content.
- •OpenAI emphasizes that a watermark does not prove AI authorship, establish legal responsibility, or verify factual accuracy of the text.
- •The detector tool is initially available only to approved researchers and expert organisations through an application process.
Key Questions Answered
OpenAI is introducing an invisible watermark called textGrain for AI-generated text in the European Union to comply with the EU AI Act. The technology adds an invisible statistical signal to the model's word choices in eligible ChatGPT and Codex outputs.
Detection performance drops significantly with shorter text lengths and heavy editing. At a 1% false-positive rate, the detector identified watermarks in about 95% of 400-token passages compared to 80% of 200-token passages, while replacing 25% of words with synonyms reduced detection down to 17%.
No, a detected watermark only indicates that an OpenAI system generated or processed part of a passage. It does not establish ownership, legal responsibility, human judgment levels, accuracy, or identify the specific user and their prompts.
