Quick answer: OpenAI says it will begin adding an invisible, machine-readable watermark to eligible text generated by ChatGPT and Codex for users in the European Union. The rollout is expected over the coming weeks and is tied to transparency requirements under the EU AI Act. The watermark is not visible to readers, is not a cryptographic proof of authorship, and is not expected to survive every possible rewrite or transformation.
What is the new ChatGPT text watermark?
The watermark is a hidden signal embedded in generated text so approved detection systems can identify whether content is likely to have been produced by an eligible OpenAI model. Unlike a visible label, it does not add a badge, footer or “AI-generated” notice to the page.
OpenAI’s system has been described in reporting as textGrain. It is designed to create a machine-readable pattern while keeping the text natural enough that ordinary users should not notice a difference in wording or formatting.
Who gets the watermark?
OpenAI’s initial rollout is focused on eligible ChatGPT and Codex text generated for users in the European Union. Reporting on the announcement says the change will apply across ChatGPT plan levels rather than being limited to a single paid tier.
For API developers, the situation is different: OpenAI is making watermarking available as an opt-in capability for selected models globally. That means an API application outside the EU may be able to enable the same type of signal even though the consumer rollout begins in Europe.
When does the rollout start?
OpenAI announced the change on October 5, 2026. The company says the rollout will happen over the following weeks rather than switching on for every account at once. Users should therefore expect availability to vary by account and product surface during the transition.
Can you see the watermark?
No. The watermark is intended to be invisible to human readers. It is not the same as visible provenance labels used on some AI-generated images. You should not expect a special symbol, color, metadata panel or sentence appended to a ChatGPT answer.
Can ordinary users detect it?
Not initially. Detection access is expected to be restricted to approved researchers and expert organizations rather than exposed as a universal public checker. That limitation reduces the risk of a simple public detector becoming an optimization target for people trying to defeat the signal.
What the watermark can and cannot prove
| Question | Answer |
|---|---|
| Can it indicate OpenAI-generated text? | Yes, when the signal survives and the detector has access. |
| Is it visible to a reader? | No. |
| Does it prove who wrote or submitted the text? | No. |
| Is it guaranteed to survive all rewriting? | No. |
| Is it a plagiarism detector? | No. |
| Is it a cryptographic signature of authorship? | No. |
Why OpenAI is introducing it in Europe
The European Union’s AI Act includes transparency obligations intended to help people and institutions identify synthetic content in certain circumstances. Watermarking is one technical approach to meeting those requirements without forcing visible labels into every generated paragraph.
The move also reflects a broader industry problem: text is much harder to watermark robustly than images. A picture can carry metadata or a pixel-level signal; text can be copied, shortened, translated, paraphrased, reformatted or mixed with human writing.
Will the watermark change answer quality?
OpenAI has not presented the feature as a new writing style or content mode. In normal use, users should expect the same product interface. However, any text watermarking system has to encode a statistical signal through token or phrase choices, so researchers will watch closely for measurable differences in fluency, vocabulary, creativity or model behavior.
For organizations, the important test is not whether a single example “sounds different” but whether large-scale evaluations show quality changes across languages and domains.
Can the watermark be removed?
No text watermark should be treated as impossible to remove. Heavy paraphrasing, translation, summarization, manual editing or mixing model output with human writing can weaken statistical signals. OpenAI itself does not position the mechanism as perfectly robust or as definitive proof of origin.
That is why detection should be interpreted as one piece of evidence rather than a standalone accusation. A school, employer or publisher should not treat a detector result as proof that a specific person cheated or misrepresented authorship.
What this means for students and schools
Schools may be tempted to use watermark detection as a disciplinary tool, but the limitations matter. A watermark can indicate that text resembles eligible AI output; it cannot show who generated it, whether AI use was permitted, how much a human edited it, or whether the content passed through another system afterward.
A better policy is to combine process evidence, drafts, citations, oral explanation and clear AI-use rules rather than relying on one detection score.
What this means for publishers and SEO teams
The presence of a watermark does not automatically make text low quality, spam or ineligible for search. Search engines primarily evaluate usefulness, originality, reliability and policy compliance. Publishers should focus on editorial review and factual accuracy rather than trying to remove a hidden signal simply because it exists.
For high-stakes publishing, keep source notes and human review records. Provenance documentation is far more useful than guessing whether a detector can see the final text.
What this means for developers using the API
API developers should watch OpenAI’s model documentation for the exact models and request parameters that support opt-in watermarking. Before enabling it globally:
- Test whether output quality changes on your production prompts.
- Check supported languages.
- Document the behavior in your privacy and transparency materials if relevant.
- Measure how downstream transformations such as translation affect detection.
- Do not market the watermark as guaranteed authorship proof.
How this differs from AI detectors
Traditional AI detectors try to infer whether text “looks like” model output based on statistical characteristics. A watermark is different because the generator intentionally introduces a detectable pattern at creation time. In theory, that can provide a stronger signal than guessing after the fact.
But the two approaches share an important limitation: neither should be used as unquestionable evidence about a person’s intent or behavior.
Could watermarking expand beyond the EU?
Yes. The API opt-in option already creates a path for applications worldwide to use the technique, and future regulation or product policy could expand consumer coverage. OpenAI has not said that every ChatGPT message worldwide will automatically receive the EU-style watermark, so users should distinguish confirmed rollout plans from speculation.
Related AVARIXO coverage
For more AI platform changes, see our OpenAI Decisions API guide, Gemini Free Tier changes guide, and our coverage of llama.cpp v0.6.0.
Frequently asked questions
Will all ChatGPT text have a watermark?
No. The announced consumer rollout is focused on eligible text in the EU, while API support is opt-in for selected models.
Can I see the watermark in copied text?
No. It is designed to be machine-readable rather than visible.
Can a watermark prove a student used ChatGPT?
It may provide evidence that text was generated by an eligible system, but it cannot prove who generated it or whether use violated a rule.
Does the watermark survive paraphrasing?
Not reliably in every case. Strong transformations can weaken text watermark signals.
Sources
This article was verified against same-day reporting from TechCrunch and The Verge. Because the rollout is ongoing, product scope and detector access can change after publication.
Other Articles
EmbeddingGemma 2: Specs, RAM, Download and How to Run It
