What OpenAI announced
On October 5, OpenAI described a phased approach to text provenance. The company says it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union over the coming weeks. Separately, API customers worldwide can opt in for selected models; watermarking remains off by default in the API. OpenAI also opened applications for detector access, initially limiting it to approved researchers and expert organisations.
The announcement is about a signal embedded in text generation, not a visible label attached to every response. OpenAI calls the method textGrain. Its stated purpose is to help assess whether a passage contains a signal associated with its models. The EU deployment is planned, not described as complete, and the API choice is distinct from the regional ChatGPT and Codex rollout.
How a statistical watermark differs from metadata
A text watermark cannot simply be stored inside a document in the way image metadata can. OpenAI says textGrain influences word choices to create an invisible statistical pattern. A detector then evaluates a passage for that pattern. The method depends on the text containing enough flexible language for the signal to be measurable, so it is not a universal identifier that works equally well on every sentence or subject.
OpenAI reports that shorter or tightly constrained passages are harder to classify. Its published evaluation also found that editing can weaken the signal: replacing words with synonyms reduced detection rates in the tested passages. Those results are the company’s own evaluation, not a guarantee about every language, model, document type or real-world editing process.
What a detection result cannot prove
OpenAI explicitly warns that a watermark does not measure how much a person contributed, establish ownership or responsibility, identify a user, or verify whether a passage is accurate. A positive result is therefore a limited provenance clue. It should not be treated as proof that a person copied text, that an organisation used a model improperly, or that a claim in the text is true.
A negative result is equally limited. The signal may be absent because the passage is short, constrained, edited, translated, created by an unsupported model, or produced before watermarking was available. OpenAI says its detector is not public at launch because false positives and missed detections remain concerns. Businesses should avoid using an unvalidated detector as an employment, academic, legal or compliance decision-maker.
The EU context requires careful reading
OpenAI frames its approach as a response to the EU AI Act. Article 50 includes transparency obligations for providers of generative AI systems, while the Commission’s guidance and the precise obligations depend on role, system and use. A vendor’s implementation announcement is not a substitute for reading the regulation or assessing a particular organisation’s duties. Watermarking is one technical approach; it does not by itself resolve every disclosure or transparency question.
For businesses that publish or process generated material, the practical step is to document how content is created, reviewed and disclosed. Teams should identify which tools and plans they use, whether an API option is enabled, and what the relevant product says about regional availability. Legal or compliance teams should use the official EU materials for interpretation rather than infer that every AI-assisted paragraph must carry the same signal.
A practical policy for content teams
Publishers and marketing teams can treat provenance signals as one possible input in a broader process. Keep human review, source checking and editorial accountability in place. If a piece of text matters, retain the working history and source material instead of asking a detector to reconstruct who did what. Do not promise customers that a watermark will make generated content tamper-proof or reliably distinguishable after normal editing.
Developers integrating models should check the specific API documentation, model eligibility and opt-in behaviour before designing a workflow around watermarking. For EU-facing products, explain the feature accurately and make uncertainty visible. The announcement is a meaningful move toward machine-readable provenance, but OpenAI’s own limitations show why it should complement—not replace—clear disclosure, records and human judgment.
Sources & further reading
Have a factual correction or a source to suggest? Contact the editorial desk.
