OpenAI Begins Text Watermarking in the EU Under AI Act, Offers API Option Worldwide

Key Info

OpenAI is expanding content provenance to cover text, watermarking eligible ChatGPT and Codex output in the EU to comply with the EU AI Act, with an opt-in watermarking option for API customers worldwide.

Highlights

• Watermarking embeds an invisible statistical signal in generated text with no visible marks or special characters
• It only answers "was this likely generated by an OpenAI model?" and does not identify author, account, conversation, or prompt
• OpenAI says testing shows no impact on model capability, speed, or response quality
• Watermarks can be removed by rewriting or translating, and are often undetectable in short passages
• Detector access is currently limited to approved researchers to help evaluate and improve the tech