Stories
OpenAI Introduces Invisible Text Watermarks to Comply With EU Regulations
- •OpenAI is rolling out invisible text watermarking for ChatGPT and Codex in the European Union over the coming weeks to meet EU AI Act requirements.
- •The technology, named textGrain, embeds an invisible statistical signal into word choices that can be analyzed using a dedicated detector.
- •API customers globally can opt in to use the text watermarking feature for select models, though it remains disabled by default.
- •Detection reliability decreases with shorter text passages, edited content, and mathematical text, with synonym replacements significantly weakening the signal.
- •Detector access will initially be limited to approved researchers and expert organizations, and OpenAI plans to eventually make textGrain open source.
Key Questions Answered
OpenAI's technology, called textGrain, adds an invisible statistical signal to the model's word choices without visibly changing the output. A dedicated detector can then analyze a passage to determine if it contains the watermark.
OpenAI will initially introduce invisible watermarks to eligible ChatGPT and Codex text outputs in the European Union over the coming weeks to comply with the EU AI Act's requirements. The feature will be limited to the EU for users across all plans, while global API customers can opt in for select models.
Detection becomes less reliable with shorter passages, constrained content, and edited text. For instance, replacing 10% of words with synonyms in a 400-token passage reduced detection from about 92% to 66%, while replacing 25% of words dropped it to 17%.
