ChatGPT Watermark: 4 Tips and Risks You Need to Know
OpenAI is adding an invisible watermark to ChatGPT text in the EU. Read what it is, how reliable it is and what it means for your work.
The ChatGPT watermark has arrived this week for users in Europe: OpenAI is making its texts carry an invisible marker. You won’t see anything, but other systems may be able to detect it. That could affect how you use ChatGPT at work.
OpenAI announced this on Monday 5 October 2026, as TechCrunch reports.
What exactly is the ChatGPT watermark?
The ChatGPT watermark is a hidden statistical pattern in the words ChatGPT chooses. You read ordinary text, but a detector can spot the pattern afterwards.
OpenAI calls the technique textGrain. It adds an invisible statistical signal to the model’s choice of words. This creates a watermark that isn’t in the letters, but in the choice of words.
It covers text from ChatGPT and from Codex, OpenAI’s programming assistant. According to ITdaily, the rollout is starting with eligible subscriptions in the European Union. OpenAI also wants to make the technique available as open source later, so that others can research it.
Why is the ChatGPT watermark coming now?
The reason is the European AI Act. It contains transparency rules that require providers of generative AI to label content as AI-generated and make that label machine-readable.
According to ITdaily, those rules have applied since 2 August. OpenAI wants to comply with them using this watermark. According to the same source, Anthropic, Google and Microsoft have also committed to the European labelling requirement.
So expect more providers to take similar measures. If you’re curious about the background to the law, also read our article on AI literacy for business owners.
Who will get the ChatGPT watermark and who won’t?
The ChatGPT watermark is being rolled out to eligible ChatGPT and Codex users in the EU, on all subscriptions. According to OpenAI it is coming over the next few weeks, so you won’t notice it straight away.
If you use ChatGPT in the Netherlands or Belgium, you’re included. Developers worldwide can switch the watermark on via the API for selected models. An API has the watermark switched off by default, so companies have to choose it themselves.
| Situation | Watermark? |
|---|---|
| ChatGPT and Codex in the EU, all subscriptions | Yes, over the next few weeks |
| API, worldwide | Only if the developer switches it on |
For exactly how things stand outside the EU for ChatGPT itself, it’s best to check with OpenAI. The announcement focuses on the EU.
How reliable is a watermark in AI text?
The ChatGPT watermark is not foolproof. OpenAI itself says that it depends on the length and type of text, and that editing quickly weakens it.
ITdaily cites figures from OpenAI’s report. With a false positive rate of one per cent, the detector recognised about 80 per cent of texts of 200 tokens. At 400 tokens it was about 95 per cent, in text about psychology. A token is a piece of a word; 400 tokens is roughly 300 words. For texts with little variation in word choice, such as maths, detection is clearly lower.
Editing helps the sender: for passages of 400 tokens, detection fell from about 92 to 66 per cent when ten per cent of the words were replaced with synonyms. When 25 per cent were replaced, only around 17 per cent remained.
- Short texts are harder to detect than long ones.
- Translating or rewriting weakens the watermark.
- Text from another AI model does not carry this watermark.
Can anyone tell that you used ChatGPT?
For now, hardly. OpenAI is initially giving the detector only to approved researchers and expert organisations. So colleagues, clients or your boss can’t simply check a text.
That may change later. That’s why it’s smart to make agreements now. Be honest about your AI use, even if nobody can see it.
Important: OpenAI itself says that the absence of a watermark does not prove that a human wrote the text. The text may be too short, edited, translated or from another model. A watermark also doesn’t show how much human judgement and editing went into a text. So never accuse anyone on the basis of a scan alone.
What does the ChatGPT watermark mean for your work?
For most people at work, little changes, but you do need to think about honesty and agreements. The watermark mainly affects you if you pass on ChatGPT texts almost unchanged.
Think of a quote, newsletter or report. Read everything through, add your own knowledge and check the facts. That produces a better text, and it makes the ChatGPT watermark almost a non-issue.
- Check whether your organisation has a rule on AI use.
- State in your agreement with clients whether you use AI.
- Check and edit every AI text for facts and tone.
- Don’t use AI detectors as the only proof, for example in job applications or assignments.
Want to know whether the ChatGPT watermark affects you? Look at your own situation: if you use ChatGPT for internal notes, it makes little difference. If you publish texts for clients or on your website, a clear way of working matters more. Write down how you use AI and who does the final check.
What are the risks of the ChatGPT watermark?
The biggest risks are false confidence and unjust accusations. A watermark looks like hard technology, but it remains a matter of probability.
- False alarm: with a false positive rate of one per cent, a genuinely human text can still be flagged as AI.
- False sense of security: no watermark does not mean the text was written by a human.
- Privacy: don’t share personal data or confidential matters in ChatGPT, watermark or not. Read our 7 rules for your privacy.
Also note that this is a rollout that is still under way. Details may change. So check with OpenAI itself for the latest situation.
What does this mean for other AI tools?
The ChatGPT watermark doesn’t stand alone. According to ITdaily, Anthropic already began watermarking AI text and images in August, and Microsoft in Copilot and Google in Gemini did something similar.
If you also use Claude, Gemini or Copilot, keep an eye on announcements from those providers. How each company handles it may differ. Always check with the provider itself and don’t go by rumours.
A practical example: you use ChatGPT to write a first draft of an email to a client and change it considerably. Then little of the watermark is left. If you paste the same text unchanged into a report, the pattern probably stays more visible to a detector.
So the real benefit is not in hiding AI, but in using it well. You remain responsible for what you send, with or without a watermark.
How to make agreements about AI text
Start with a short AI agreement for yourself or your team. A few sentences are enough, as long as they are clear.
Want to start gently? Our guide on getting started with AI at work will help you get going. If contracts are involved, have a lawyer look at the agreement first; this article is not legal advice.
Frequently asked questions
What is the ChatGPT watermark?
It is an invisible pattern in the word choice of ChatGPT and Codex. A detector can recognise that pattern later, but you won’t see anything.
Does the ChatGPT watermark also apply in the Netherlands and Belgium?
Yes, both countries are part of the EU. OpenAI is rolling it out over the next few weeks for eligible subscriptions in the EU.
Can I remove the watermark?
Editing and rewriting weakens the watermark, but OpenAI itself describes that as a limitation. Don’t use this to hide AI use if you are expected to be honest.
Does a watermark prove that a text comes from AI?
No. It is a matter of probability and can go wrong. If the watermark is missing, that says nothing about who wrote the text.
Your next step: learn to use ChatGPT wisely and safely with our guide using ChatGPT for beginners.
Was this article helpful?
