OpenAI to Introduce Invisible textGrain Watermark for Selected EU Outputs
OpenAI plans to mark eligible ChatGPT and Codex text outputs in the European Union with the invisible textGrain signal in the coming weeks.
OpenAI textGrain is expected to be added in the coming weeks to eligible text outputs from ChatGPT and Codex for users in the European Union. It is an invisible watermark intended to enable the technical identification of AI-generated text without the signal being visible to ordinary readers.
The announcement comes in connection with the European AI Act rules. Article 50 requires generated or manipulated content to be marked in a machine-readable and detectable way. These obligations apply from August 2, 2026.
How OpenAI textGrain Works
The textGrain technology affects the statistical selection of words or parts of words when generating a response. It does not add hidden characters to the text or use unusual punctuation. The signal therefore is not intended to change the text’s normal visual appearance or be recognizable during ordinary reading.
The watermark is intended to serve as a signal for tools designed to detect it, not as a label that a person would see directly in the text. OpenAI said access to the detector will initially be limited to approved researchers and professional organizations.
For the API, watermarking will be available globally for selected models, but it will be optional. It is not expected to be enabled by default at launch. The company has not yet published a complete list of supported models, output types or exceptions.
What the Watermark Can and Cannot Do
A text watermark is not standalone proof of authorship, user identity or the truthfulness of content. If the detector finds the signal, it means only that the relevant marking was technically detected. Likewise, failing to find it does not prove that a human created the text.
OpenAI also warns that text edits and paraphrasing can significantly weaken detection. These limitations are why the text detector will not initially be publicly available. When evaluating text, the watermark therefore cannot be used as the sole basis, for example, when assessing plagiarism, content origin or disinformation.
EU Deployment Still Has No Exact Date
OpenAI says the rollout will take place over the coming weeks, but it has not announced an exact completion date. It also remains unclear which specific ChatGPT and Codex outputs the marking will cover and in which cases it will not apply.
For users in the EU, this means that some text created by supported services may carry an invisible, machine-readable signal. The ability to independently verify this signal, however, will at least initially be limited to a selected group of organizations.
It will be important to monitor the conditions for accessing the detector, any possible expansion to the wider public and independent error-rate tests. These should show how the marking behaves with short texts, translations and subsequent wording changes.
Sources
- OpenAI — Our approach to EU text provenance rules – Confirms the planned introduction of textGrain in ChatGPT and Codex in the EU, the globally optional API setting, limited detector access and measured detection limits.
- OpenAI Help Center — Provenance signals in OpenAI-generated content – Describes how textGrain works, its invisible nature, its regional use in the EU and its limitations after extensive text edits.
- European Commission — Transparency obligations under Article 50 of the AI Act – Confirms the requirement for machine-readable and detectable marking of generated text and the date Article 50 applies.
Verified and updated: 10/06/2026 06:22
