# ChatGPT watermark in the EU: what your company must mark by 2 December > For owners and managers in Europe whose staff write in ChatGPT or who run a chatbot on the OpenAI API. Six situations, who marks the text, and the dates. - **Canonical URL:** https://appwave.dev/en/blog/chatgpt-watermark-eu-what-your-company-must-mark - **Language:** en - **Updated:** 2026-10-16 --- - **Published:** 2026-10-16 - **Author:** Mateusz Cieśliński - Co-Founder, DevOps Engineer - **LinkedIn:** https://www.linkedin.com/in/mateusz-cie%C5%9Bli%C5%84ski/ - **Reading time (min):** 7 - **Topics:** EU AI Act, AI security, AI agent, ChatGPT ![A paragraph under a magnifying glass beside a desk calendar with one day marked. Under the glass part of the text lines up into a glowing sine wave: an invisible text watermark and a deadline to mark AI text.](https://cms.appwave.dev/uploads/chatgpt_watermark_eu_what_your_company_must_mark_5db98a0806.webp) **On 5 October 2026 OpenAI announced an invisible watermark for ChatGPT text in the EU. In the API it is off by default. Run an API chatbot under your name, and the Article 50(2) AI Act marking duty may be yours, and systems placed on the market before 2 August have until 2 December 2026. As of 16 October 2026.** Most coverage asks whether a teacher can catch an essay. A company has a different question. Who has to mark the text it sends to customers, and has anyone already done it? ## What changed: ChatGPT in the EU, the API only on request OpenAI is adding a watermark called textGrain to “eligible” ChatGPT and Codex text on all plans, but only in the European Union. The announcement does not mention the UK, Switzerland, Norway, Iceland or Liechtenstein. There is no firm date, only “over the coming weeks”. The mark is not a label and not a hidden character. The model shifts its word choices according to a secret key, and a detector holding the same key spots the pattern. That is why tools that strip invisible spaces from text do not touch it. Five points from OpenAI’s own documents change what a company should decide: - in the API, watermarking is off, and you switch it on under Organization settings, Data controls, Text provenance, or per project, choosing the models; - according to OpenAI, the EU Code of Practice does not require a watermark on output under 200 tokens (a token is roughly part of a word) or on code snippets; - in OpenAI’s test on 400-token English passages, swapping 25% of words for synonyms cut detection from about 92% to 17%; - the detector goes only to approved research and academic organisations at first, not to businesses or schools checking work; - the mark does not identify who used ChatGPT, and its absence does not prove a human wrote the text. The test figures are for English. OpenAI publishes per-language results in its Help Center, and they vary: at a 1% false-positive rate, Spanish scored highest at 69.0% and Romanian lowest at 42.2%. ## Does the ChatGPT watermark affect your company? It depends on who generates the text and where it goes. Article 50(2) asks the provider of a generative AI system to mark its output in a machine-readable way. Article 50(4) asks the deployer, meaning the business using the tool, to label text on matters of public interest visibly, unless a person with editorial responsibility reviewed it. - **Staff write emails and proposals in ChatGPT.** Who marks: OpenAI as provider. What you do: add nothing. Decide who reads text before it goes out. From when: ChatGPT rollout over the coming weeks. - **You subscribe to a ready-made chatbot.** Who marks: the tool’s vendor. What you do: ask the vendor in writing how it marks output. From when: 2 August 2026, older systems 2 December 2026. - **Your own chatbot or agent on an API under your brand, placed on the market before 2 August 2026.** Who marks: probably you, as provider of the system. What you do: turn on Text provenance and document it. From when: 2 December 2026. - **The same, launched after 2 August 2026.** Who marks: probably you. What you do: as above, with no transition period. From when: now. - **Internal tool, strictly technical text that never leaves the company.** Who marks: possible exception. What you do: record who has access. Ask a lawyer. From when: to be confirmed. - **You publish AI text on public matters, e.g. health, politics, the environment.** Who marks: you, with a visible label. What you do: a label, or review by a responsible editor. From when: 2 August 2026. ![Diagram: whether your API chatbot must mark its text and by when, depending on who reads it and when it went live.](https://cms.appwave.dev/uploads/chatgpt_watermark_eu_what_your_company_must_mark_diagram_4d63ea3fce.webp) The four middle rows turn on one question: did you put an AI system into service under your own name? We covered the same provider question for chatbot disclosure in [Chatbot and the EU AI Act: what to check before 28 October 2026](https://appwave.dev/en/blog/chatbot-ai-act-checklist-before-28-october-2026). The difference between a simple chat and an agent is in [Corporate chatbot or AI agent](https://appwave.dev/en/blog/corporate-chatbot-or-ai-agent). Companies outside the EU are in the picture too. The AI Act covers providers placing systems on the EU market and those whose output is used in the EU, so a UK firm with a chatbot answering customers in Ireland or the Netherlands is in scope even though ChatGPT’s own watermark skips the UK. ## What we see in the market: AI assistants disagree On 16 October we asked four AI assistants the same question about an API chatbot. ChatGPT said you usually do not need to mark anything yourself, because that is OpenAI’s job. Gemini said to switch on the API setting, but claimed the deadline had already passed. Only Perplexity gave 2 December 2026. **Ask an assistant about your own duty and you may get three different answers.** The Commission’s Article 50 guidelines of 20 July 2026, as summarised by the firm Lawcel, say two relevant things. A system provider may rely on marking done by the model provider if that solution meets Article 50(2), but proving compliance stays with the system provider. They also describe a narrow exception for strictly technical output read only by a limited, pre-defined group of professionals, which is not meant to leave the company and is protected from doing so. Text pasted into a customer email does not qualify. We have not read the full guidelines, because our tools could not open them. The Commission confirmed on 20 July 2026 that systems placed on the market before August must comply with marking from 2 December 2026. Fines for breaching Article 50 reach €15 million or 3% of worldwide turnover, whichever is higher. For small and medium-sized firms, Article 99(6) applies the lower figure. That is a ceiling, not a rate. ## What this means for your company **Probable conclusion.** If your team only writes in ChatGPT, you have no new marking duty. You do have a reason to decide who reads text before it goes out. The watermark does not check whether the price in a proposal is right. **Probable conclusion.** A company with its own API chatbot has a decision to make by 2 December, and nobody will make it for them. OpenAI leaves the API watermark off, so “we use OpenAI” does not mean your answers are marked. **Hypothesis to test.** Chatbot answers are often shorter than 200 tokens, and OpenAI does not have to watermark those. We do not yet know how an authority will view a chatbot whose answers mostly carry no mark. That is a question for a lawyer and for your model provider. Switching vendors does not avoid the topic. Anthropic marks text from supported Claude models worldwide, including the API, and tells businesses to assess their own Article 50 duties. ## What to do this week These steps take hours, not weeks: 1. List every system that generates text sent outside the company: chatbot, email agent, product description generator. Note who launched each one and when. 2. For an API chatbot, check the Text provenance setting at organisation and project level in OpenAI. Write down which models it covers. 3. For a bought tool, send the vendor one question: how and since when do you meet Article 50(2)? 4. Agree one rule for ChatGPT text in your team, for example who reads a proposal before it is sent. We describe the options in [Human in the loop AI](https://appwave.dev/en/blog/human-in-the-loop-ai-human-oversight-in-a-small-business). 5. If you would rather go through the list with someone who does this daily, we do it as part of an [AI audit](https://appwave.dev/en/services/ai-audit). You get a description of each system that you can read without us. The worry we hear most from owners is not the watermark. It is a bot saying something wrong to a customer. A watermark does not prevent that. Rules, logs and a person who reads them do, the same point we made about [background agents such as OpenAI dots](https://appwave.dev/en/blog/openai-dots-in-europe-availability-and-first-settings). We are not a law firm and this is not legal advice. We have not tested the watermark on a ChatGPT account in the EU, and whether your company is the provider of a system is for a lawyer to decide. [Book a 45-minute call](https://appwave.dev/en/appointment). ## Sources All pages accessed on 16 October 2026. - OpenAI, 5 October 2026: [Our approach to EU text provenance rules](https://openai.com/index/eu-text-provenance/). - OpenAI Help Center: [Provenance signals in OpenAI-generated content](https://help.openai.com/en/articles/8912793-provenance-signals-in-openai-generated-content), API settings, the 200-token threshold, per-language detection and detector access. - European Commission, press release IP/26/1653, 20 July 2026: [2 December 2026 for systems placed on the market before August](https://europa.eu/newsroom/ecpc-failover/pdf/ip-26-1653_en.pdf). - European Commission: [Quick facts: transparency rules for AI systems](https://digital-strategy.ec.europa.eu/en/factpages/quick-facts-transparency-rules-ai-systems), fines and the editing exception. - Regulation (EU) 2026/1744, recital 38, four-month transition: [EUR-Lex](https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX:32026R1744). - Anthropic Help Center: [How Claude marks AI-generated content](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content). - Paul, Weiss: [EU Finalises Transparency Rules for AI-Generated Content](https://www.paulweiss.com/insights/client-memos/eu-finalises-transparency-rules-for-ai-generated-content), examples of public-interest matters in the guidelines. - Lawcel: [Article 50 applies on 2 August 2026, and your model vendor cannot carry it for you](https://lawcel.com/blog/ai-act-article-50-transparency-august-2026), summary of the Commission guidelines and Article 99(6). - Answers from ChatGPT (gpt-6-luna), Gemini, Perplexity and Claude, collected on 16 October 2026 via DataForSEO. ## Questions & Answers (FAQ) ### Does ChatGPT have a watermark? In the EU, yes, or soon. On 5 October 2026 OpenAI said that, over the coming weeks, eligible ChatGPT and Codex output on every plan in the European Union will carry an invisible watermark. It lives in word choice, not in hidden characters. Outside the EU, ChatGPT text is not watermarked by default. ### Does the ChatGPT watermark apply in the UK, Switzerland or Norway? OpenAI names only the European Union. Its announcement does not mention the UK, Switzerland or the EEA countries outside the EU. The AI Act itself can still reach a UK or Swiss company whose chatbot serves people in the EU, because it covers systems placed on the EU market or whose output is used there. ### Is my chatbot on the OpenAI API watermarked automatically? No. Text watermarking is off by default in the API. You switch it on under Organization settings, Data controls, Text provenance, or per project, and pick the models. OpenAI also says the EU Code of Practice does not require a watermark on output under 200 tokens or on code, and many chatbot answers are that short. ### By when must a company mark text from its own chatbot? Article 50 of the AI Act has applied since 2 August 2026. Generative systems already on the market before that date have until 2 December 2026 for machine-readable marking, as the European Commission confirmed on 20 July 2026. Whether your company counts as the provider of the system is a question for a lawyer. ### Do I have to label emails written with ChatGPT as AI-generated? Usually not. Marking ChatGPT text is OpenAI’s duty as provider. The AI Act asks a business for a visible label when it publishes AI text on matters of public interest without review by a person who holds editorial responsibility. A sales email to a customer is not that kind of text. --- ## About AppWave AI development company from Łódź, Poland. We build custom AI software, automation agents and web applications. We co-create them with the client and guarantee every system we deliver. - **Legal name:** AppWave sp. z o.o. - **Address:** ul. Kolumny 147E/1, 93-611 Łódź, Poland - **Phone:** +48 538 441 413 - **Email:** office@appwave.dev - **Business hours:** Monday-Friday, 09:00-17:00 - [Book a Free Consultation](https://appwave.dev/en/appointment)