OpenAI has activated textGrain — its new text watermarking system, for ChatGPT and Codex in the European Union, beginning in the coming weeks. The move fulfils a requirement under the EU AI Act, which mandates that generative AI providers mark synthetic text so it can be identified as such.
Unlike earlier approaches that altered formatting or added visible tags, textGrain embeds an invisible statistical signal into the model’s word selection process. It does this by biasing certain token choices. Among semantically interchangeable options, according to a secret cryptographic key. The result is undetectable to readers but identifiable by OpenAI’s detector, which looks for the pattern across longer passages.
The rollout is split by channel and geography. For ChatGPT and Codex, watermarking applies only to eligible users in the EU. No global or opt-out provision is mentioned. For the OpenAI API, watermarking is available globally but remains off by default. Developers must explicitly opt in for select models.
OpenAI has opened applications for access to its watermark detector, though initial access is restricted to approved researchers and expert organisations. The company stresses the detector is still experimental: false positives and false negatives remain common, especially in short texts (under 400 tokens) or domains with rigid phrasing, like mathematics.
Editing degrades detection reliability. Replacing just 10% of words in a 400-token passage with synonyms drops detection accuracy from ~92% to 66%. At 25% synonym replacement, detection falls to 17%.
OpenAI also clarifies what the watermark does not do. It does not measure human contribution. It does not establish ownership, responsibility, or user identity. It does not verify factual accuracy. And: the absence of a detected watermark does not prove human authorship.
In benchmark testing, watermarked outputs scored slightly higher on some metrics than unwatermarked ones. Though OpenAI says it observed no meaningful difference in quality or readability.
The company plans to publish further technical details in an updated report and intends to open-source textGrain in the future. It also says it will revisit the approach as standards evolve, particularly around how watermarks hold up under translation, summarisation, and editing, and whether they can better distinguish AI assistance from full AI authorship.
Anthropic’s earlier implementation for Claude. Which also uses word-choice biasing, prompted industry-wide anticipation of similar moves. OpenAI’s version arrives with narrower scope, tighter geographic limits, and more explicit caveats about fragility and interpretive limits.
No timeline is given for expansion beyond the EU. No pricing, hardware requirements, or model-specific eligibility thresholds are disclosed. The watermark applies only to text output. Not images, audio, or video generated by OpenAI systems.
OpenAI expects to refine textGrain based on real-world use, researcher feedback, and regulatory developments, but for now, it marks the first live deployment of its watermarking tech in a consumer-facing product.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.