OpenAI launches textGrain watermarking to meet EU AI Act requirements

The launch of textGrain signals a shift from voluntary safety commitments to mandatory technical compliance under the EU AI Act. By implementing invisible watermarking, OpenAI is attempting to balance regulatory transparency with user experience, though the system's vulnerability to simple editing highlights the ongoing difficulty of definitive AI attribution. This move follows a similar implementation by Anthropic, suggesting that statistical watermarking is becoming the industry standard for text provenance. Streaming and media professionals should monitor the upcoming open-source release of textGrain to see if it becomes a cross-platform tool for content verification.

OpenAI developed textGrain in-house rather than licensing Google DeepMind's SynthID or Meta's TextSeal, citing the need for finer control over the tradeoff between watermark detectability and response diversity. The company's FAQ states that textGrain matched or exceeded SynthID's detection performance in internal testing, and OpenAI plans to release the technology as open source so external developers can build on it. The system works by subtly biasing token selection during generation, creating a statistical pattern detectable by a verification tool without inserting hidden characters or unusual punctuation.

OpenAI's deployment strategy differs materially from Anthropic's approach to the same EU AI Act requirement. Anthropic applies watermarking globally to all supported Claude models at the model level, meaning the signal is present regardless of which product surface delivers the text, and the company has not described an equivalent opt-out for API developers. OpenAI, by contrast, limits automatic watermarking to EU-based ChatGPT and Codex users and offers API customers worldwide an opt-in toggle at the project or organization level, giving developers explicit control over whether their outputs carry the signal. OpenAI also signed the EU Code of Practice on Transparency of AI-Generated Content alongside several other major AI providers, committing to a multi-layered provenance approach that combines watermarks with C2PA Content Credentials metadata.

On the technical side, OpenAI published a detailed textGrain report specifying generation and detection procedures, including how to set a budget for the sampling entropy lost in exchange for watermark signal strength. The Verge reported that OpenAI included AI benchmark scores showing similar performance between watermarked and unwatermarked text, though the company acknowledges that textGrain does not guarantee reliable detection and that substantial paraphrasing or translation can render the watermark undetectable. OpenAI is providing detector access to approved researchers and expert organizations on a case-by-case basis, and is working with cloud partners to make watermarking available for model outputs accessed through their services.