OpenAI to Implement Invisible Watermarks in ChatGPT Texts for the European Union
The textGrain technology will allow AI-generated content to be identified through word statistics but can be easily bypassed by editing, OpenAI reports.
OpenAI will add invisible watermarks to texts generated by ChatGPT and Codex. The watermarks function by altering the statistical selection of words and are not noticeable to users, but detectors will be able to recognize them.
The watermarks will only appear for users in the European Union due to the requirements of the EU AI Act. This law mandates that AI developers label synthetic content, or face fines of up to 15 million euros or 3% of the company's annual global turnover. In other countries, including Uzbekistan, watermarks can be enabled voluntarily via API.
The detector identifies watermarks in 80% of texts with a length of 200 tokens and in 95% of texts with 400 tokens. If 10% of the words are replaced with synonyms, accuracy drops by 25%, and if 25% of the text is altered, it falls to 7–15%. OpenAI states that the watermarks do not affect the quality of generation and do not prove that the text was entirely created by AI.
