OpenAI to watermark eligible ChatGPT and Codex text in the EU
OpenAI says it will add an invisible statistical watermark to eligible ChatGPT and Codex text generated in the European Union. Access to its text detector will initially be limited to approved researchers and expert organizations because short or edited passages can produce unreliable results.
Summary
OpenAI has announced a phased rollout of an invisible statistical watermark for eligible text generated by ChatGPT and Codex in the European Union. The company says the signal, called textGrain, is intended to make its models’ involvement detectable and address transparency obligations under the EU AI Act, which has applied since August 2, 2026.
In practice
The watermark is not a visible label. It is embedded in word choices during generation and analysed by a detector. OpenAI expects to begin adding it to eligible ChatGPT and Codex outputs in the EU over the weeks following its October 5 announcement. The company does not say that every response, model or use case will be covered.
For the API, customers worldwide will be able to opt in to watermarking for select models; it will remain off by default. The text detector will not be publicly available at launch. OpenAI is accepting applications from researchers and expert organizations, with access granted case by case.
Context
Article 50 of the EU AI Act provides for machine-readable marks that enable synthetic or manipulated content to be detected. The European Commission clarifies that this provider obligation is separate from the rules for people publishing AI-generated text on matters of public interest, which include an exception for human review or editorial responsibility. A technical watermark, by itself, is not a visible disclosure to readers.
OpenAI says its tests found detection more effective on longer passages with more flexibility in how they could be worded. For psychology passages, it reports detecting about 80% of watermarks in 200-token texts and 95% in 400-token texts, at a set false-positive rate of 1%. According to the company, replacing 10% of words with synonyms reduced detection from about 92% to 66%; replacing 25% reduced it to 17%. These are results reported by OpenAI, not a guarantee of performance on everyday text.
Why it matters
- A statistical watermark may help identify some text generated by OpenAI models, but it cannot show who prompted the model or how much human work went into the passage.
- A positive result does not prove that text is accurate, lawful or responsible; a negative result does not prove human authorship.
- Restricted detector access acknowledges that short, highly constrained or edited passages can produce false negatives—and that a detector can also flag unwatermarked text incorrectly.