What happened
OpenAI published its approach to EU text provenance rules on October 5, 2026. It has three parts. First, over the coming weeks, eligible ChatGPT and Codex text in the EU will carry an invisible watermark, on all plans. OpenAI says it is not making this a global default at launch. Second, API customers anywhere can opt in to watermarked text for select models starting that day. It stays off by default. OpenAI says it is also working with cloud partners to offer it on its models accessed through their services. Third, OpenAI opened applications for its watermark detector, but access starts with approved researchers and expert organizations. The detector is not public at launch, because of the risk of missed watermarks and false positives. The method, which OpenAI calls textGrain, nudges the model's word choices so a detector holding a secret key can spot the pattern later. OpenAI says it saw no meaningful change in model performance with it on. The Verge and TechCrunch both report that Anthropic announced a similar step for Claude in August, applied worldwide.
Why it matters for business owners
If your team uses ChatGPT to draft emails, proposals, product pages, or reports, the text it produces will start carrying a hidden signal, at least for EU users. The signal travels with the text when it is copied and pasted. Most owners will never see it. But it changes a few quiet assumptions. Text you publish may be identifiable as machine-assisted by someone with the right tool. And as AI providers add these marks, the question of what you disclose to customers gets harder to ignore. This also shows how regulation reaches you in practice. The EU AI Act puts the marking duty on the AI provider. You feel it as a product change you did not ask for.
What owners should not misunderstand
Do not treat the watermark as proof. By OpenAI's own account, editing weakens it. In its test on 400-token English passages, replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%. Short passages, math answers, and translated text are harder to detect. Do not assume no watermark means a human wrote it. OpenAI says plainly that a missing watermark does not prove human authorship. The text may be short, edited, translated, from another vendor's model, or from before watermarking began. Do not assume a watermark says who is responsible. OpenAI says it does not measure human contribution, does not establish ownership, does not identify the user, and does not verify accuracy. And do not assume this settles your own obligations. The watermark is the provider's job. Whether your business has disclosure duties for content you publish is a separate question, and one for your legal adviser. Atlacis does not give legal advice.
The operational lesson
A vendor can mark its output. It cannot decide your content standard for you. The watermark answers a narrow question: did a particular vendor's system help generate this passage? The questions a business actually faces are different. Which documents may AI draft? Who checks the facts? What do we tell customers about how something was made? Those answers live in your process, not in the vendor's setting. It is also a reminder that AI tools change under you. Your team's tool can start behaving differently because of a rule in another country. Knowing which tools are in use, and watching their release notes, is basic operational hygiene.
What a serious business should do next
Write down where AI drafts text in your company, including tools staff picked on their own. If you do not know, that is the first gap. Then decide three things in plain language. Which kinds of content AI may draft without review. Which always need a named person to check and approve. And what you would say to a customer who asked whether AI was involved. If you use the API through a developer or agency, ask whether watermarking is on or off, and whether anyone has decided that on purpose. Since it is off by default, someone has to make that choice. If you sell into the EU, ask your adviser what the transparency rules mean for your own content. Search Engine Journal reports that the European Commission gives systems already on the market until December 2 to meet the marking duty, so vendor settings may keep changing. Do not buy a detection tool because of this announcement. OpenAI itself says the technology has significant limitations, and its own detector is not public.
The Atlacis view
Atlacis helps owners slow down before an AI decision, map the workflow and the risk, and choose the simplest setup that fits the business. A watermark is a good example of why. It sounds like control, but it covers one vendor, one kind of output, and only some of the time. The controls that hold up are yours: a short list of AI tools in use, clear review rules, and a plain answer about disclosure. If you want help drawing that up, we can do it with you.
The short version
- On October 5, 2026, OpenAI said it will add an invisible watermark to ChatGPT and Codex text for EU users over the coming weeks, with an off-by-default API opt-in worldwide.
- It responds to the EU AI Act's rule that AI-generated text be machine-readable as such. The duty sits with the provider, not with you.
- By OpenAI's own tests, editing weakens the mark, and a missing watermark does not prove a human wrote the text.
- The detector is limited to approved researchers and expert organizations, so you cannot easily check text yourself.
- Your job stays the same: know where AI drafts text, set review rules, and decide what you tell customers. Ask a legal adviser about disclosure duties.
Where ATLACIS can help
Sources
- OpenAI: Our approach to EU text provenance rules (October 5, 2026)
- The Verge: OpenAI is adding text watermarking in ChatGPT and Codex (October 5, 2026)
- TechCrunch: OpenAI will start watermarking ChatGPT's text in the EU (October 5, 2026)
- Search Engine Journal: OpenAI to watermark ChatGPT text in the EU, opens API opt-in (October 6, 2026)