OpenAI announced three textGrain access paths: a worldwide API opt-in, a phased EU rollout for ChatGPT and Codex, and a separately approved text detector. Their availability and access differ. Announcement.
This distinction matters if you configure a production API, review AI-assisted copy, or need to know whether a passage can be checked. The matrix is a practical map of the setting, rollout status and detector gate.
OpenAI text watermarking by product surface
| Surface | What OpenAI documents | What to do or avoid assuming |
|---|---|---|
| OpenAI API | Worldwide opt-in for selected models. Watermarking is off by default. Announcement. | Use the project or organization setting and the current supported-model selector. Settings steps. |
| ChatGPT and Codex | Eligible users across all plans in the EU will receive text watermarking over the coming weeks. Announcement. | No precise account-eligibility test or end-user switch is specified. Do not promise a date for a particular account. Announcement. |
| Text detector | Qualifying researchers and expert organizations can apply; applications are reviewed case by case. Announcement. | Keep detector access separate from generation setup. Text application route. |
| OpenAI Verify and Content Provenance API | The public tools are documented for supported image and audio provenance signals. API guide. | Not a textGrain checker; text detection follows a separate application route. Text verification guidance. |
How API customers turn it on
OpenAI’s help instructions place the control in platform settings. To set an organization-wide default, go to Organization settings > Data controls > Text provenance. For one project, open Project Settings > Text provenance. Then turn on Allow text watermarking, select the model or models shown as available, and choose Save. OpenAI’s API watermarking instructions.
OpenAI applies the watermark during generation for supported models; customers do not add a separate mark to each response. Check the live settings page for the current model list. OpenAI Help Center.
For a release record, note the project or organization scope, the model names selected, and the date you checked them. Confirm that production requests use that project and one of its supported models. If the setting or model is absent, check current OpenAI documentation or support before assuming the account is eligible.
What textGrain changes
OpenAI’s technical report describes textGrain as a keyed sampling method that uses word and word-piece choices to make a statistical watermark detectable in generated text. OpenAI’s textGrain technical report.
A general AI-writing score does not test textGrain; text verification uses a separate application route. OpenAI’s text-detection explanation; text verification guidance.
For a separate provider’s implementation and editorial context, see our guide to Claude text watermarking for SEO and content teams. OpenAI’s API setting, EU rollout and detector follow separate access rules.
What a text-detector result can tell you
- A signal is detected
- A positive result indicates that an OpenAI system likely generated or processed some of the passage; it does not identify a user or contribution share. OpenAI interpretation.
- No signal is detected
- No signal does not prove a person wrote the text; short, edited, translated, unsupported or pre-watermark text may not be detected. OpenAI limitations.
- You do not have text-detector access
- The text was not checked for textGrain. A general AI detector cannot substitute for the restricted watermark detector.
OpenAI warns detection can produce false positives or false negatives. Short or constrained text is harder to detect, and editing can weaken the signal. Treat a result as limited evidence, not an authorship verdict. OpenAI’s detector limitations.
For content workflow planning beyond provenance checks, see our AI content marketing tools guide.
The EU rule and the watermark are different questions
OpenAI says its EU rollout responds to the AI Act. The European Commission says Article 50(2) requires providers of in-scope systems generating synthetic text to mark outputs in a machine-readable form and make them detectable as AI-generated or manipulated, as far as technically feasible, with exceptions including assistive standard editing or cases where the system does not substantially alter its input or meaning. Its Article 50(4) guidance describes a separate duty for deployers publishing some AI-generated text to inform the public on matters of public interest; that duty has an exception for human review or editorial control with a responsible person or organization holding editorial responsibility. European Commission Article 50 guidelines.
The Commission says Article 50 applies from August 2, 2026. The limited grace period to December 2 applies only to the Article 50(2) marking-and-detection obligation for AI systems placed on the market before August 2. The Code of Practice is voluntary; the underlying transparency requirements are legal obligations. A text watermark is one technical signal, not confirmation that a provider or publisher meets every applicable duty. Commission Article 50 FAQ; Code of Practice guidance.
Prepared with AI assistance from public sources; checked October 6, 2026. No API project or text detector was accessed or tested. Product rollout and supported-model availability can change.
