ChatGPT now watermarks its text in the EU: what that means for your firm

On 5 October 2026 OpenAI announced that, over the coming weeks, it will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union. Customers of its API, anywhere in the world, can switch the same watermark on for selected models. In the API it stays off unless you turn it on.

OpenAI says it is doing this in response to the EU AI Act. If your firm drafts client letters, reports or proposals with ChatGPT, here is what has actually changed, and whose problem it is.

What OpenAI is doing

• The watermark sits in the wording. OpenAI's system, called textGrain, adds an invisible statistical signal to the model's word choices. Nothing on the page looks any different.
• EU only, for now. It applies to eligible ChatGPT and Codex users on all plans in the EU. OpenAI says it is not making text watermarking a global default at launch.
• Almost nobody can check for it yet. Access to the detector starts with approved researchers and expert organisations. Your clients cannot run your letter through it today.
• Editing weakens it. In OpenAI's own tests on 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%. Replacing 25% cut it to 17%. Short passages, and text with little choice of wording, are harder to detect.
• It proves less than it sounds. OpenAI is clear that a watermark does not identify the user, does not measure how much a person edited the text, and that the absence of a watermark does not prove a human wrote it.

The marking duty belongs to the tool's maker, not to you

The duty OpenAI is meeting is Article 50(2) of the EU AI Act. It is addressed to providers: the businesses that make AI systems generating synthetic text, images, audio or video. They must make the output machine-readable as AI-generated.

A firm that uses ChatGPT to draft its work is not the provider of ChatGPT. You do not have to watermark your own documents, and nobody should be selling you a tool on the basis that you do.

On timing: Article 50 has applied since 2 August 2026, and a generative tool launched since then has had to mark its output from its first day. Under the Digital Omnibus amendments, tools already on the market before 2 August 2026 have until 2 December 2026 to bring their marking into line.

What the Act does ask of a firm that uses AI

Article 50(4) puts a disclosure duty on the business using the tool in one situation that matters to professional firms. If you publish AI-generated text to inform the public on matters of public interest, you must disclose that it was AI-generated, unless:

• the text has been through human review or editorial control, and
• a person or organisation holds editorial responsibility for publishing it.

A letter to one client is not publication to the public. A briefing, article or newsletter on a public-interest topic can be. Where someone at your firm reviews it and the firm stands behind it, the exemption applies. Where AI-drafted text goes out unread, it does not.

The Act reaches beyond the EU here. Article 2(1)(c) covers providers and deployers established outside the Union where the output their AI system produces is used in the Union. A UK, US or Canadian firm publishing for EU readers is in scope.

Telling a client that a letter was drafted with AI help can be good practice, and your engagement terms may require it. It is not an Article 50 duty for a firm that only uses the tool.

Three things to do this month

1. Write down your review step. Who reads AI-drafted work before it leaves the building, and who signs it off. That single control is what keeps published content inside the Article 50(4) exemption, and it is the quality check your clients assume you already have.
2. Check whether you have built anything on the OpenAI API. If your firm offers clients a chatbot, a document generator or another tool built on the API under your own name, you may be the provider of that system. The marking duty could then be yours, and OpenAI's API watermark is off unless you switch it on.
3. Expect your clients to ask. Watermarking makes "did AI write this?" a question people will start putting to professional firms. A written AI policy that answers it is easier to send than an explanation.

Sources

• OpenAI, Our approach to EU text provenance rules, 5 October 2026
• Regulation (EU) 2024/1689 (the EU AI Act), Articles 2 and 50

To see which parts of the EU AI Act reach your firm and when, start with our EU AI Act guide.