OpenAI is adding text watermarking in ChatGPT and Codex

OpenAI is rolling out an invisible, machine-readable watermark called textGrain in text outputs from ChatGPT and Codex for eligible users in the European Union. The company says the regional launch will occur over the coming weeks and is not a global default at this time; OpenAI framed the phased approach as a way to gather real-world feedback before wider deployment.

OpenAI compared textGrain to other watermarking approaches and reported that it "matched or exceeded" systems such as Google DeepMind's SynthID for text, which is the method Anthropic used when it announced watermarking in August. The company also released benchmark scores indicating similar AI performance on watermarked and unwatermarked text, while cautioning that the watermarking does not assure detection or verify other attributes of content.

In its announcement OpenAI warned of limitations: textGrain "does not guarantee reliable detection," and watermarks should not be used to prove accuracy, assign ownership, measure the extent of human contribution, or establish human authorship. Given those risks, OpenAI said detector access will be restricted — approved researchers and expert organizations can apply for access on a case-by-case basis, and the tool will only report whether it detects an OpenAI watermark without revealing user prompts or conversations.

Outside the EU rollout, OpenAI is enabling API customers worldwide to opt in to watermarked outputs for select models immediately, allowing developers and organizations to decide how watermarking fits their transparency needs. The company also said it is working with cloud partners to make watermarking available for OpenAI model outputs accessed through their services in the coming weeks.