On October 5, 2026, OpenAI announced that an invisible watermark, textGrain, will be added in the coming weeks to the ChatGPT and Codex text of users in the European Union, to comply with Article 50 of the AI Act. The signal sits in the choice of words. The detector is reserved for approved researchers and organizations, and its rate drops from about 92% to 17% when 25% of the words in a 400-token passage are replaced with synonyms. Google asks site owners to check AI-generated content before publishing it, and human review is what answers that request.
Contents
On October 5, 2026, OpenAI published its approach to the European text provenance rules1, and Search Engine Journal covered it on October 6.2 For a small business that has the pages of its site written with ChatGPT, the question is what this marking changes for its SEO and for its contracts. This article covers what will be marked, what the detector can establish, what Google's page on AI-generated content says, and the place human review keeps.
01What will be marked, and by whom
The obligation comes from Article 50 of the AI Act, applicable since August 2, 2026. Providers of generative AI systems must mark their outputs in a machine-readable format, so that they can be detected as AI-generated.3 Systems placed on the market before August 2, 2026 have until December 2, 2026 to apply this marking.3
What OpenAI is rolling out
In the coming weeks, OpenAI will add an invisible watermark to eligible ChatGPT and Codex text, on all plans, for users in the European Union only. The marking is not enabled by default in the rest of the world at launch.1 In the API, customers worldwide can enable it now for some models, with the option off by default.1
| Product | Announced marking |
|---|---|
| ChatGPT and Codex, EU users | watermark added on all plans, in the coming weeks1 |
| API, worldwide | option that can be enabled for some models, off by default1 |
| Text detector | applications open, access limited to approved researchers and expert organizations1 |
| Images and audio | verification tools already public, including openai.com/verify1 |
Provider and deployer: who owes what
The European Commission distinguishes the provider, which develops the system and places it on the market, from the deployer, which uses it under its authority in a professional setting.3 Marking outputs is the provider's job. The deployer only has a visible labeling obligation for text published to inform the public on matters of public interest, and only if it has undergone neither human review nor editorial control.3
To fall under this obligation, a text must meet three cumulative criteria: being published, informing the public, and dealing with a matter of public interest. The FAQ cites, for example, public health, consumer safety or environmental protection.3 Fines can reach 15 million euros or 3% of worldwide turnover, with possible proportionality for SMEs.3
02What the detector can establish
OpenAI describes textGrain as a statistical signal added to the model's word choices, which the detector then looks for. The post stresses the limits: good performance in ideal conditions does not guarantee reliable detection in everyday use.1
OpenAI's detection rates
OpenAI measures detection with a target false positive rate of 1%. It depends on the length of the text and its type.1
| Case measured | Detection |
|---|---|
| 200-token passage, psychology-type content | about 80% |
| 400-token passage, psychology-type content | about 95% |
| Math-type content | clearly lower, as word choice is less free there |
These measurements cover answers to questions from the ELI5 dataset.1 Search Engine Journal points out that the European Code of Practice does not require marking free text under 200 tokens, which matches the shortest length in OpenAI's results.2
The effect of edits
On 400-token English passages, replacing 10% of the words with synonyms brings detection down from about 92% to about 66%. With 25% of the words replaced, it falls to about 17%.1
What a detection does not tell you
OpenAI lists what cannot be concluded from a detection result.1
- The human share: a watermark indicates that an OpenAI system generated or processed part of the passage, without measuring what a person contributed.
- Ownership and responsibility: it does not say who owns the text, or who is accountable for it.
- Identity: it links the text to no person, account or conversation.
- Accuracy: it does not say whether the passage is true, misleading or well contextualized.
- The absence of a signal: it does not prove a human wrote the text, which may be too short, edited, translated, produced by a model that is not covered or by another company's tool.
Anthropic describes the same limits for Claude: the watermark establishes that Claude probably took part in the text, without distinguishing whether Claude wrote it or heavily edited it.4 Anthropic marks its text worldwide, lacking a durable way to limit marking to one region.4 According to the same source, AI detection software on the market does not hold the key to this watermark and analyzes other clues, such as turns of phrase.4
Search Engine Journal gives the example of an agency whose writers, in Berlin and Toronto, share the same ChatGPT subscription: part of its texts could carry OpenAI's watermark and part not. A contract clause that took a detection result as proof of who wrote a text would ask the watermark a question that, according to OpenAI, it does not answer.2
03Google, watermarked text and human review
Google's guidance on AI-generated content focuses on the accuracy, quality and relevance of content.5
What Google already asks
Google writes that generative AI can help research a topic and structure original content. Generating many pages without adding value for users, on the other hand, may violate its policies on spam and scaled content abuse.5
Google also asks site owners to manually check the accuracy and reliability of any AI-generated content before publishing it, including title elements, meta descriptions, structured data and image alt text.5 These checks match passes 3 to 5 of the five-pass review of AI-produced content, the first two of which cover facts and figures.
What the watermark does not check
The watermark tells you about the origin of a passage. Accuracy, which Google asks you to check before publishing, is not part of what it measures: OpenAI writes that a watermark does not say whether a passage is true, misleading or well contextualized.1
The Commission defines human review as the deliberate examination of the substance of content by one or more people with the knowledge and professional judgment of the subject. It explicitly excludes superficial or purely formal checks, such as spelling or grammar correction.3
Saying how a text was produced
Google also invites site owners to share information about how content was created, for example about the use of automation, when this is relevant to the audience.5 On its page about helpful content, the “How” question asks whether the use of AI is evident to visitors and recommends adding disclosures where readers might reasonably expect them.6 Disclosing the use of AI is one of the checks of honestly signing a small business's pages.