OpenAI to add invisible watermarks to ChatGPT text in the EU

OpenAI announced that to meet the EU AI Act's transparency rules, it will add invisible, machine-readable watermarks to text output from ChatGPT and Codex in the EU over the coming weeks. The technology, called textGrain, hides statistical signals while the model picks its words, so a detector can tell if text carries an OpenAI watermark.

The feature will first roll out only in the EU and will not be a global default at launch. Starting today, API customers worldwide can choose to enable watermarked output for some models and decide how watermarks fit their transparency duties. OpenAI is also working with cloud partners on this.

Plus, approved researchers and expert groups can apply for detector access from today. The tool only reports whether an OpenAI watermark is found. It never reveals user identity, prompts, or chat content.