Stories
OpenAI Introduces Invisible Text Watermarks for ChatGPT in EU
- •OpenAI is adding invisible statistical watermarks to ChatGPT and Codex text in the EU to comply with the EU AI Act.
- •The system uses textGrain to alter word patterns invisibly, making machine-generated text machine-readable.
- •The watermark detector will initially be available only to approved researchers and expert organizations.
- •API customers globally can choose to opt in to the watermarking feature for selected models.
- •The watermark has limitations, as it cannot measure human contribution, prove ownership, or determine factual accuracy.
Key Questions Answered
OpenAI is adding invisible watermarks to eligible text outputs from ChatGPT and Codex in the EU to comply with the EU AI Act, which requires providers of generative AI systems to make AI-generated text identifiable in a machine-readable way.
ChatGPT uses a system called textGrain, which changes the statistical pattern of words selected by the AI model without making the watermark visible to readers. The technology embeds an invisible statistical signal into word choices, which a detector can then scan for.
OpenAI initially makes its text watermark detector available exclusively to approved researchers and expert organisations on a case-by-case basis. The detector signals whether an OpenAI watermark is present but does not reveal user identities, prompts, or conversations.
