OpenAI is adding an invisible provenance signal to some AI-generated text, but it is not a universal AI detector. On October 5, 2026, OpenAI announced that API customers worldwide can opt in to text watermarking for select models. It also said eligible ChatGPT and Codex text in the European Union will be watermarked over the coming weeks. The API setting remains off by default.
For teams shipping AI features, the immediate action is to check whether watermarking is available for their models and to decide what, if anything, should change in their content-disclosure and verification workflow. A watermark result should be treated as one clue about provenance—not a verdict about authorship or truth.
What changed on October 5?
| Surface | OpenAI’s announced status | Practical implication |
|---|---|---|
| OpenAI API | Global opt-in for select models, off by default | An organization or project owner can choose supported models in Text provenance settings. |
| ChatGPT and Codex in the EU | Planned rollout to eligible text over the coming weeks | Do not assume every existing or future output is already marked. |
| Text watermark detector | Applications open to approved researchers and expert organizations | It is not a public, general-purpose “AI-written” checker at launch. |
| Images and audio | Existing provenance tools continue separately | Do not confuse textGrain with Content Credentials or image/audio SynthID checks. |
OpenAI frames the change as part of its response to European text-provenance requirements. The official announcement gives the rollout scope and cautions; its help-center guidance explains the customer settings and interpretation limits. Availability can vary by model, product, region, and the date content was generated.
How textGrain works, in plain English
OpenAI calls its method textGrain. Rather than adding hidden characters, it subtly changes how a model chooses among plausible next words or word pieces. Across enough text, those choices form a statistical pattern that a detector with the matching configuration can look for.
That distinction matters. Copying text does not expose a secret tag, and a visible watermark is not placed on the page. The signal is in the wording itself. OpenAI says the method is designed to leave response quality broadly intact, but its published evaluation is OpenAI’s own testing—not a guarantee for every language, format, or application.
What a result can—and cannot—tell you
A positive result can support the narrower conclusion that an OpenAI system likely generated or processed some of a passage. It cannot tell you who prompted the system, how much a person edited the result, whether the text is accurate, or who owns it. It is not proof that a document was published responsibly.
A negative result is even easier to overread. OpenAI reports that short passages and highly constrained text are harder to detect. Editing and synonym replacement can weaken the signal; translation, older content, and unsupported models create further gaps. In OpenAI’s examples, detection on one type of 400-token passage dropped sharply after substantial word replacement. Those figures describe specific tests, not a reliability score for every article or language.
If you run a news site, hiring workflow, or user-generated-content platform, do not label someone dishonest because a detector returned a result—or declare text human-written because it found nothing. Keep source records, editorial review, and correction processes alongside any technical signal.
What API teams should do now
OpenAI’s current instructions say an API customer can open Organization settings → Data controls → Text provenance for an organization default, or Project Settings → Text provenance for a project override. Turn on Allow text watermarking, select supported models, and save. The available model list appears in those settings; do not assume every model is covered yet.
Before enabling it in a production application:
- Identify which text outputs you generate and whether users see, edit, translate, or republish them.
- Test representative outputs in the languages and formats your product actually uses, especially short answers and code-heavy responses.
- Document what the signal means and who may interpret it. Avoid describing it as infallible AI detection.
- Keep normal safeguards: source citations, content review, access controls, logs, and a correction path.
- Recheck OpenAI’s model and regional coverage as the rollout changes.
If you are designing an AI feature for a business, AI consulting and application development can help turn provenance and disclosure requirements into a workflow that people can understand and operate. For a customer-facing application, pair that policy work with website and web application development so UI labels, audit records, and human review have clear ownership.
Bottom line
TextGrain is a meaningful addition to OpenAI’s provenance toolkit because it creates a signal inside generated text rather than relying on removable file metadata. It is also deliberately limited: API use is opt-in for supported models, the EU product rollout is phased, and the detector is not being opened to everyone. The safest product decision is to treat a watermark as supporting evidence, not a substitute for judgment, attribution, or a sound publishing process.
Sources
Frequently asked questions
Is OpenAI watermarking every ChatGPT and Codex response worldwide?
No. OpenAI says eligible ChatGPT and Codex text in the EU will receive watermarks over the coming weeks. API customers can opt in globally for supported models; API text watermarking is off by default.
How do API customers enable textGrain?
OpenAI says customers can enable Allow text watermarking in organization or project Text provenance settings, choose supported models, and save. Check current model coverage in your own settings before changing production.
Can a text watermark prove who wrote a passage?
No. A detected watermark is a limited signal that an OpenAI system generated or processed text. It does not identify the user, measure human contribution, prove ownership, or establish accuracy.
Does no detected watermark prove a person wrote the text?
No. Short or constrained text, editing, translation, unsupported models, and text created before rollout can all prevent detection. The text detector is initially limited to approved researchers and expert organizations.