OpenAI begins rolling out invisible text watermarks for EU users
OpenAI will introduce an invisible, machine-readable text watermark called textGrain to ChatGPT and Codex users in the European Union over the coming weeks, in response to the EU's AI Act requirements. The feature will not become a global default for now.
OpenAI has developed a new watermarking technology called textGrain that makes AI-generated text machine-identifiable without leaving visible traces for users. Over the coming weeks, the feature will roll out to ChatGPT and Codex users across all subscription plans, but initially only within the European Union.
The company says the regional rollout is tied to requirements under the EU's AI Act. Anthropic made a similar move in August, basing its own watermarking system on Google DeepMind's SynthID technology. OpenAI claims textGrain's performance "matched or exceeded" other approaches, including SynthID.
Not a global standard yet
OpenAI emphasizes that text watermarking will not become a global default at launch, saying the regional approach gives it room to learn from real-world use and feedback. The company also published benchmark results showing similar performance between watermarked and unwatermarked text.
However, OpenAI cautions that textGrain "does not guarantee reliable detection" and that watermarks do not verify a text's accuracy, determine ownership, measure how much a human contributed, or prove human authorship.
Also available to API customers
Starting today, API customers worldwide can opt in to watermarked outputs for select models, letting them decide how watermarking fits their own transparency obligations. OpenAI is also working with cloud partners to extend availability through their platforms in the coming weeks.
The company is additionally granting approved researchers and expert organizations access to a detection tool, initially on a case-by-case basis. The tool will only indicate whether text contains an OpenAI watermark, without revealing user identity or conversation content. Given the risk of missed watermarks and false positives, the tool will not be made publicly available at launch.
