OpenAI adds invisible watermarks to ChatGPT and Codex text in the EU
OpenAI will add an invisible watermark to text generated by ChatGPT and Codex for eligible users in the European Union, according to its announcement on 5 October 2026. The rollout covers all plans and will take place over the coming weeks. The mark is called textGrain. It changes the statistical pattern of word choices rather than adding visible symbols, so a reader will not see it. A detector can look for the pattern, but a missing mark does not prove a human wrote the text, and a found mark does not prove the text was entirely AI-generated. For businesses, the practical point is that the watermark does not move responsibility. A company that publishes AI-drafted text in the EU still answers for what it signs.
In short
- OpenAI will add an invisible textGrain watermark to eligible ChatGPT and Codex text for EU users over the coming weeks.
- OpenAI's own tests show that synonym swaps weaken detection, with a 10% swap cutting it from about 92% to 66%.
- A missing watermark does not prove human authorship, and a found mark does not prove the text is entirely AI-generated.
- The watermark does not transfer legal or editorial responsibility away from the business that signs the text.
- Marking obligations for systems already on the market are deferred to 2 December 2026, according to Paul Weiss.
OpenAI said on Monday, 5 October 2026, that it will add an invisible watermark to text produced by ChatGPT and Codex for users in the European Union. According to the company, the change is a response to the EU AI Act, the law that requires AI providers to make generated content identifiable by machines. The rollout will reach eligible users on all plans, but only in the EU. OpenAI has not made the mark a global default for ChatGPT.
The technology is called textGrain. According to TechCrunch, the mark does not add hidden characters, unusual punctuation or invisible spaces to the output. Instead, the model is nudged toward certain word choices when it picks among possible next words, which leaves a statistical pattern that a detector can later find. Reports from Neowin and Cybernews describe the same mechanism, and the 5 October announcement is the only primary source for the details.
The numbers that matter most are OpenAI's own test results. According to Cybernews, the detector identified watermarks in about 80% of 200-token psychology passages and about 95% of 400-token passages. Detection was weaker for mathematics and recipes, where writers have fewer ways to word a sentence. A separate summary, from the MLQ news site, reports that swapping 10% of words with synonyms cut detection from about 92% to 66% in a 400-token test. Neowin reports that replacing 25% of words dropped detection to 17%. These figures come from OpenAI's testing, not from independent labs, and they describe test conditions rather than real-world documents.
Why it matters for businesses that automate work
Many firms now run AI drafting inside routine work. Follow-up emails, product descriptions, billing notes, policy summaries and client reports can all start as model output. Until now, a company rarely had a way to tell from the text alone where it came from. The watermark gives detection tools a signal to look for, which may matter to regulators, platforms and clients who ask how a document was made.
The limits matter just as much. According to OpenAI, detecting a watermark does not prove a text was entirely AI-generated, and its absence does not prove a human wrote it. A legal commentary from Shibolet & Co. adds a point for deployers: a business using a provider's tools must not remove or alter the watermarks the provider embeds. Automation that rewrites AI output through another tool, translation or paraphrase can weaken the signal, which is why the mark should not be treated as a reliable audit trail.
What changes in practice for a business that automates work
For most staff using ChatGPT in the EU, the daily work does not change. Nothing visible appears in the text. Copying and pasting into an email or a report works as before. The change is in what a business should be able to answer if asked about a document.
| Situation | What the watermark does | What the business still does |
|---|---|---|
| Internal drafts, such as meeting notes | Nothing visible to staff | Keeps a record of which tools were used where it matters |
| Client documents and contracts | May be detectable, but not proof of authorship | Assigns a named person to review and sign off |
| Public-interest text, such as news or public notices | Does not replace labelling | Checks the labelling duty under Article 50 of the AI Act |
| Text passed through translation or rewriting tools | Detection can drop sharply | Treats the mark as unreliable once the text has been edited |
According to Paul Weiss, the Commission published finalised guidelines on Article 50 on 20 July 2026. The same memo says that the Article 50 obligations apply from 2 August 2026 for most AI systems. Marking obligations for generative systems already on the market are deferred to 2 December 2026, and watermark detection interoperability to 2 February 2027. Paul Weiss also notes that deployers must label AI-generated text published on matters of public interest, unless it has been reviewed or taken on by a legal or natural person with editorial responsibility.
A five-person firm and a group with thousands of staff face the same question: who is responsible for the output? A small firm may only need a short written rule that says which tools may draft client text and who signs it off. A large group will need that rule in its systems, with logs, approvals and retention periods. Neither case is solved by a watermark alone.
What to watch next
Several items are still open. OpenAI has opened applications for its detector, but access will initially go only to approved researchers and expert organisations. According to Technobezz, OpenAI intends to release textGrain as open source and is working with cloud partners to offer watermarking through their services in the coming weeks. OpenAI has also said that the API will offer watermarking to customers worldwide, with the option switched off by default.
Two dates are worth noting. The deferred marking deadline of 2 December 2026 falls within weeks, and the watermark detection interoperability deadline of 2 February 2027 follows. Businesses should also check OpenAI's own announcement and the EU Commission's guidelines before relying on any detail in this article.
Automation cannot settle what a watermark cannot show. It cannot tell a client who wrote a paragraph, how much a person edited it, or whether the facts are correct. Those checks still need a person, and a business that automates drafting should assign that person in writing before the watermark rolls out to its own accounts.
How AiStaffo would automate this
AiStaffo would start by connecting the tools that draft text, such as ChatGPT or Codex, to the systems where final documents are stored and sent. Every AI-assisted output would be logged with its tool, date and purpose, so the business can answer questions without relying on a watermark. Routine items, such as follow-up emails or billing notes, would run automatically. Client-facing or public text would be routed to a named person for review before it leaves the business. The owner still decides who is accountable for each document and approves the rules. Book a free automation audit.
Questions people ask
Will ChatGPT text look different after the watermark rolls out in the EU?
Can a detector prove that a document was written by AI?
Does the watermark apply to ChatGPT outside the EU?
What should a business do if it publishes AI-drafted text in the EU?
Does the watermark survive translation or rewriting?
Book a free automation audit
Thirty minutes. We look at one process you run every week and tell you exactly what an AI worker would take off your desk, and what it would not.
