OpenAI to watermark eligible ChatGPT and Codex text in the EU

On 5 October 2026, OpenAI announced its approach to text provenance under the EU AI Act. In the coming weeks, the company will begin watermarking eligible text outputs for eligible ChatGPT and Codex users in the European Union with an invisible textGrain signal. API customers, meanwhile, can enable watermarked outputs worldwide for selected models as an optional feature starting today.

How will the watermark work?

According to OpenAI, textGrain embeds an invisible statistical signal in the model’s word choices. A detector then assesses whether that signal is present in the text being examined. The company describes the method as a limited signal about a text’s origin, not as a definitive verification tool on its own.

The watermark does not show who wrote the text, who the user is, how much human input was involved, or whether the content is accurate. Nor does it make judgments about ownership or responsibility. The absence of a detected signal also does not prove that a person wrote the content. The signal may not be detected in short, edited, translated, or text generated by an unsupported model.

What are the limits of detection results?

OpenAI’s published evaluations show that detection performance varies with text length and how much flexibility the content allows in word choice. With a false-positive target of 1%, the watermark was detected in about 80% of 200-token texts on subjects such as psychology, and around 95% of 400-token texts. Detection rates were substantially lower for content where word choice is more constrained, such as mathematics.

Editing can also weaken the signal. In an evaluation of 400-token texts, replacing 10% of the words with synonyms lowered the detection rate from around 92% to 66%; with 25% changed, the rate fell to 17%. For this reason, the detector will not initially be available to the general public. Researchers and expert organizations can apply, with access granted case by case for evaluation and improvement purposes. Without revealing users or prompts, the tool will report only whether an OpenAI watermark was detected.

Who will get access, and when?

Watermarking is planned to begin in the coming weeks for eligible text outputs from eligible ChatGPT and Codex users in the EU. The rollout will cover all plans in the EU, but will apply only to eligible users and outputs. OpenAI says it will not make the feature a global default initially. Worldwide, API customers will be able to optionally enable watermarking for selected models; the feature will remain off by default. The company is also working to make watermarking available in the coming weeks for model outputs offered through cloud partners.

OpenAI says its comparisons using the Astra model found no meaningful performance difference between watermarked and unwatermarked text. The company also plans to release the technology as open source, but has not given a date. It says it will continue to make its image and audio verification tools available to organizations seeking to examine content provenance.

Implications for architecture and visualization workflows

Architecture and interior design teams that use AI to draft project descriptions, presentation copy, or client communications can treat the watermark as an additional signal about a text’s origin. However, because it does not reveal how much a person contributed or whether the content is accurate, it is no substitute for editorial review.

Teams should keep track of the rollout for eligible users in the EU, and, if they build their own tools through the API, note that the feature will be off by default and limited to selected models. Since text editing and translation can weaken the signal, detector results should not be treated as definitive on their own.