On October 5, 2026, OpenAI announced it would add an invisible watermark to text generated by ChatGPT and Codex, but only for users located in the European Union. Called textGrain, the mechanism answers a transparency obligation under the European AI Act. Two months after Anthropic made a similar announcement for Claude, here is what this ChatGPT watermark really changes for an SMB that uses AI daily to write or code.
In brief
- Since October 5, 2026, OpenAI has been rolling out textGrain, an invisible watermark embedded in ChatGPT and Codex text responses, but only for accounts located in the European Union.
- The rollout is gradual, over "the coming weeks," across all ChatGPT and Codex plans in the EU. API developers worldwide can turn it on as an opt-in, off by default.
- The watermark survives copy-pasting, but its reliability drops fast after editing: around 92% detection on an unedited 400-token passage, versus 66% after 10% of words are replaced with synonyms, and 17% after 25%.
- Unlike Anthropic, which has applied its Claude watermark worldwide since August 2026, OpenAI is for now limiting textGrain to Europe.
- For an SMB, this change requires no technical action: it is automatic on OpenAI's side. But it changes how AI-generated content or code can be identified after publication.
What OpenAI announced on October 5, 2026
textGrain is OpenAI's invisible watermarking method. It works by subtly influencing word choice during generation: an algorithm uses a secret key to steer, at each step, the selection among statistically likely next words. Taken alone, that choice is invisible to a reader. Repeated across hundreds of words, it forms a detectable statistical signal that OpenAI's own tool can analyze, without changing the meaning, tone, or quality of the generated text.
The watermark covers text responses from ChatGPT and from Codex, OpenAI's coding assistant. It applies to every plan, from free to business, but only for accounts tied to the European Union. Companies using the OpenAI API elsewhere in the world can voluntarily enable it on select models; it stays off by default outside the EU.
One notable point: OpenAI itself states that the watermark cannot identify a specific user, cannot reveal the content of a conversation, and cannot prove that a text is inaccurate or copyrighted. It only indicates a probability that the text came from an OpenAI model.
Why now: Article 50 of the AI Act
This announcement follows directly from Article 50 of the EU AI Act, which took effect on August 2, 2026, and requires providers of generative AI systems to mark their content in a way that is detectable by automated means. OpenAI is among roughly 190 organizations that signed the European Commission's Code of Practice on the transparency of AI-generated content.
Anthropic moved first, deploying a similar watermark for Claude on August 11, 2026, and applying it globally, not just in Europe. OpenAI is taking a more cautious approach, for now staying within the strict legal boundary.
Anthropic (Claude)
Text watermark active since August 11, 2026, applied worldwide. Paired with signed C2PA metadata for generated image files.
OpenAI (ChatGPT, Codex)
textGrain watermark rolled out since October 5, 2026, limited to European Union accounts. API users worldwide can opt in on select models.
Reliability that drops fast with editing
OpenAI published its own detection figures, based on internal testing. They show the watermark stays reliable on lightly edited text, but loses accuracy quickly once a human reworks the content.
Key takeaway
A detected watermark is a probability signal, not proof. OpenAI notes that short texts, math answers and translations are particularly hard to detect, even without any deliberate edit.
What this actually changes for an SMB
Most SMBs use ChatGPT and Codex as deployers: to write emails and marketing content, or to generate and complete code. Three practical consequences follow from this announcement.
First, for text content: text written with ChatGPT and published from an EU account can now be identified as such after publication, even when copied into a CMS or an email tool. A company that publishes raw, unreviewed AI content, especially public-facing content (website, press releases, social media), needs to factor this in.
Second, for code generated with Codex: the practical reach is still limited at this stage. A watermark hidden in a code comment or a variable is far less exploitable than one in editorial text, and OpenAI has not communicated concrete use cases for detecting AI-generated code. For now, this is mostly a principled compliance move.
Third, this announcement does not change obligations already in force since August 2, 2026: an SMB that publishes AI-generated content resembling reality (a photo, a voice, a news article) must still disclose it to its users, regardless of whether a technical watermark exists.
A simple checklist for an SMB
No special technical skills are needed. Three habits are enough: review and substantially enrich any ChatGPT content meant for the public rather than publishing it raw, remind marketing teams that this watermark exists and can be detected by third parties (clients, journalists, competitors), and check that the transparency notices already required under AI Act Article 50 are properly in place on the company's website and conversational tools.
With measured optimism, this broader trend, driven by European regulation but followed voluntarily by major labs, points toward more traceable AI. For an SMB that is transparent about its AI use, that is an extra point of trust with clients, not a burden to endure quietly.
FAQ
Is the ChatGPT watermark active in France?
Yes, as soon as the ChatGPT or Codex account is tied to the European Union. The rollout is gradual, over several weeks starting October 5, 2026, across all plans, including free ones.
Does the watermark prove a text came from ChatGPT?
No. OpenAI states that detection remains probabilistic: it does not identify the user or the conversation content, and the absence of detection does not prove a text was written by a human. Even a light edit can erase the signal.
Does an SMB need to change its practices because of this watermark?
Not technically: the marking is automatic on OpenAI's side. It is, however, a good moment to check that published AI content is actually reviewed and enriched by a human, and that the transparency notices required under AI Act Article 50 are already in place.
Why don't Anthropic and OpenAI take the same approach?
Anthropic chose to apply its Claude watermark worldwide starting in August 2026, going beyond the strict legal requirement. OpenAI, for now, limits textGrain to European accounts and leaves the option open, but off by default, for developers using its API elsewhere in the world.
To learn more about your AI transparency obligations, read our guide AI Act Article 50: mandatory AI transparency for SMBs and browse all our AI resources for business.
Sources: TechCrunch, "OpenAI will start watermarking ChatGPT's text in the EU" (October 5, 2026); Tech Startups, "ChatGPT text is getting an invisible watermark in Europe to comply with the EU AI Act" (October 5, 2026); European Commission, Code of Practice on the transparency of AI-generated content (AI Act Article 50).
