OpenAI to Watermark ChatGPT Text in the EU. On Oct 5, OpenAI said it will watermark eligible ChatGPT text in the EU over the coming weeks, and API customers can opt in from now using the textGrain statistical watermark.

OpenAI to watermark eligible ChatGPT text in the EU and opens API opt‑in worldwide

On Oct 5, OpenAI said that over the coming weeks it will watermark eligible ChatGPT text in the EU, and API customers can opt in from now. In a blog post titled “Our approach to EU text provenance rules,” OpenAI outlined how it will comply with the EU AI Act’s requirement that generative AI providers make AI‑generated text identifiable in a machine‑readable way.

The company’s solution is textGrain, an invisible statistical watermark that subtly shapes word choices so that a detector can identify AI‑generated text without readers noticing any visible mark. The rollout is region‑specific and opt‑in by design, rather than a global default.

What OpenAI announced on October 5

OpenAI’s October 5 update covers two parallel tracks:

1. EU ChatGPT and Codex users

Where: European Union only.
Who: Eligible ChatGPT and Codex users across all plans.
When: “Over the coming weeks” (no exact date given).
What: An invisible watermark will be added to eligible text outputs.
Default: Watermarking will be on for these EU users once rolled out; OpenAI is not making it a global default at launch.

2. API customers worldwide

Where: Globally.
Who: API customers and app builders using OpenAI models.
When: Starting October 5, 2026.
What: Ability to opt in to watermarked text outputs for select models.
Default: Watermarking remains off by default in the API.

OpenAI says the API opt‑in lets customers decide how watermarking fits their transparency obligations and the experiences they provide to end users.

How textGrain works

textGrain is a statistical watermark embedded in the model’s token selection process:

It does not add visible tags, metadata, or symbols to the text.
Instead, it biases word choices in a way that creates a detectable pattern.
Humans reading the text should not notice any difference, but a dedicated detector can identify the watermark with high confidence when conditions are favorable.

OpenAI has published a technical report describing the method but has not released the detector publicly. Access to the detection tool is currently limited to approved researchers and expert organizations on a case‑by‑case basis, in line with the EU AI Act’s code of practice for text.

Robustness and limitations

OpenAI is explicit that textGrain is not a silver bullet:

Editing weakens detection: In OpenAI’s own tests, replacing about 25% of words with synonyms can drop detection rates to around 17%.
No proof of authorship: The watermark indicates that text likely came from a watermarked model, but it does not prove who prompted it, whether it was edited, or whether the content is accurate.
False positives/negatives: As with any statistical method, there is a risk of misclassification, which is one reason OpenAI is not enabling it globally by default.

These limitations align with broader research showing that text watermarks can be degraded by paraphrasing, translation, or heavy editing.

How API customers can enable watermarking

For developers and organizations using the OpenAI API:

Watermarking is configured in organization or project settings, under Data controls → Text provenance.
It is not a per‑request parameter in the API (as of October 6).
Customers can choose which models in their setup should emit watermarked text.
The setting is off by default and must be explicitly enabled.

OpenAI says it is also working with cloud partners (such as Microsoft Azure) to make watermarking available for OpenAI model outputs accessed through those services in the coming weeks.

Why this matters for EU compliance

The EU AI Act’s transparency rules for generative AI (Article 50) began applying on August 2, 2026. Providers must:

Make AI‑generated text identifiable in a machine‑readable way.
Implement this in a manner that is effective and proportionate, while considering risks like false positives.

OpenAI’s approach—mandatory watermarking for EU ChatGPT/Codex users, optional for API customers globally—is designed to meet the letter of the rule while limiting unintended consequences outside the EU.

Summary: On Oct 5, OpenAI said that over the coming weeks it will watermark eligible ChatGPT text in the EU, and API customers can opt in from now. The invisible textGrain watermark will roll out to EU ChatGPT and Codex users across all plans, while API users worldwide can enable it for select models in their settings, with detection access limited to approved researchers.

OpenAI to Watermark ChatGPT Text in the EU and Opens API Opt‑In Worldwide