OpenAI to watermark ChatGPT output in EU, offers API opt-out

VERA Brief

AI-generated. Grounded in the article and its cited sources.

OpenAI will use invisible watermarks on ChatGPT output in the European Union to identify AI-generated content. API users globally can opt out of this feature. The watermarks have shown varying detection rates depending on text modifications.

Key facts

• OpenAI will implement invisible textGrain watermarks for ChatGPT content in the European Union.
• API customers worldwide will have an option to disable the watermarking feature.
• Initial tests showed a detection rate as high as 95 percent for the watermarks.
• The detection rate drops to 17 percent when approximately a quarter of the original words are substituted.
• The implementation aims to provide a verifiable signal for AI-generated content.