OpenAI to Watermark Eligible EU ChatGPT and Codex Text

OpenAI plans to begin adding invisible watermarks to eligible text generated by ChatGPT and Codex in the European Union over the coming weeks, the company said, as it responds to transparency requirements under the EU AI Act. Outside the bloc, watermarking will remain off by default, though customers using OpenAI’s API can enable it for select models worldwide, according to the company. The system, called textGrain, changes statistical patterns in a model’s word choices rather than adding a visible label or metadata. A detector analyzes a passage for the resulting signal and estimates whether it carries an OpenAI watermark. OpenAI said the feature is intended to help identify text generated by its systems, but has acknowledged limitations in how reliably the signal can be detected. OpenAI’s reported tests show that editing can reduce detection. For 400-token passages, the company said its detector identified watermarked text about 92% of the time under the test conditions. After 10 synonym replacements, the reported rate fell to 66%; after 25 replacements, it fell to 17%.