OpenAI will bring invisible watermarks to ChatGPT and Codex in the EU

OpenAI expects to add invisible watermarks to text generated by ChatGPT and Codex for eligible users in the European Union over the coming weeks. The company presented the measure as part of its response to transparency requirements associated with European artificial intelligence regulations, although the technical signal alone is not conclusive proof that a person or a model wrote a text.

The technology is called textGrain. Rather than adding visible characters or hidden elements to the formatting, it incorporates a statistical signal into the model’s word selection. A detector can look for this pattern in the text and estimate whether it contains an OpenAI watermark.

The rollout will differ between products and the API

In ChatGPT and Codex, the watermark will be added to the outputs of eligible EU users, according to OpenAI’s announcement. The company has not specified in the available information which criteria determine eligibility.

For certain models in its API, by contrast, customers worldwide will be able to enable watermarking voluntarily. The feature will remain disabled by default. OpenAI has also said it is working to offer the watermark through cloud service providers, but has not detailed the scope of that availability here.

Access to the detector will initially be restricted: approved researchers and expert organizations will be able to request it. Therefore, the fact that an output is watermarked does not mean that any recipient can verify it directly with a public tool.

Detection depends on the text and how it has been edited

OpenAI describes textGrain as a technology that is still at an early stage and warns that it can produce false positives—detecting a watermark where there is none—and false negatives—not detecting a watermark that is present. Its performance also varies with the length and type of content. Short and technical texts may leave less room for the model to incorporate a recognizable statistical pattern.

In evaluations reported by the company, the detector identified around 80% of 200-token passages and 95% of 400-token passages in tests involving psychology content, with a target false-positive rate of 1%. These are evaluation results under specific conditions, not a guarantee of accuracy for every text or everyday use.

Editing can significantly weaken the signal. In a test cited by OpenAI, replacing 25% of the words in 400-token passages with synonyms reduced detection from 92% to 17%. The result demonstrates a limitation of the system when text is changed; it does not establish that every type of editing will produce the same decline.

What a watermark can—and cannot—indicate

A positive detection may indicate that the text displays the pattern associated with an OpenAI watermark. It does not identify the user, reveal their instructions, determine how much human work was involved, or establish who owns the rights to the content. Nor does it verify that the information is correct.

Conversely, the detector’s failure to find the signal does not prove that a person wrote the text. It may fail to detect it because of the text’s brevity or type, later modifications, or because it was generated using a model that does not incorporate textGrain.

OpenAI places the rollout in the context of the transparency obligations in the EU AI Act. This connection should be understood as the framework the company cites to explain its decision: the available information does not support claiming that the regulation requires every generated text to be watermarked in this specific way. The watermark is a technical identification tool, with limitations acknowledged by the company itself, and does not replace other ways of assessing the origin, responsibility, or reliability of a publication.