OpenAI to add invisible watermarks to ChatGPT text in the EU
OpenAI said on 5 October it will mark ChatGPT and Codex text in the EU in a way readers cannot see. Ireland is covered, the UK is not, and the marker fades with heavy editing.
By Jason Sibley
What happened
On 5 October 2026, OpenAI said it will add an invisible watermark to eligible ChatGPT and Codex text output in the EU. The rollout is due over the coming weeks and covers all plans. It responds to the EU AI Act, which requires providers of generative AI to make generated text identifiable in a machine-readable way.
The technology is called textGrain. It adds a statistical signal to the words the model chooses, so you cannot see it when you read the text. OpenAI's detector is limited at first to approved researchers and expert organisations, so there is no public checker.
This is EU only and not a global default at launch. Ireland is covered. The UK is not in the EU, so UK users are not covered by this rollout. API customers anywhere can opt in to watermarking, but it is off by default there.
OpenAI was open about the limits. Shorter text is harder to detect: about 80% of 200-token passages, against about 95% of 400-token passages, at a 1% target false positive rate. Editing weakens the signal too. Replacing 10% of words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.
OpenAI also said a watermark does not identify the user, measure human contribution, establish ownership or responsibility, or verify accuracy. No watermark does not prove a person wrote the text.
Sources: OpenAI, 5 Oct 2026
What it means for a small business
If you run a business in Ireland and use ChatGPT to draft website copy or customer emails, some of that text may soon carry a hidden marker. You will not see it, and you do not need to switch anything on. UK firms are outside this rollout for now. A UK business with staff based in the EU who use ChatGPT there may still find some of its copy marked.
I would not try to scrub it. Reworking text just to dodge detection is a poor use of an afternoon, and the figures above show the signal fades with heavy editing anyway. The better reason to edit is that it makes the copy properly yours: accurate, and written the way you would say it to a customer.
A watermark tells you nothing about quality or truth. A marked draft can be wrong, and an unmarked one may still have come from a machine. In practice, whoever publishes the words answers to the customer for them. Whether the AI Act creates any duties for your own business, as opposed to the AI providers, is one to check with your adviser.
What you can do this week
- Agree a short, honest AI-use line for your website and marketing. Something plain such as "We use AI tools to help draft some of our content, and a member of our team checks it before it goes out" will do, as long as it is true.
- Name one person who edits and fact-checks anything customer-facing before it goes live. Prices, opening hours and claims about your service deserve a second look.
- Find out who on the team is EU-based and uses ChatGPT there, including anyone working remotely from Ireland or elsewhere in the EU.
- If a developer or agency connects ChatGPT to your systems through the API, ask whether watermarking is switched on. It is off by default, and opting in is your choice.
- Drop any idea of removing the watermark. Put that time into editing the draft so it sounds like your business.
How we can help
If you would like a hand, we offer free website help, an AI sales & marketing desk and SMB AI sessions for owners who want to use these tools sensibly. Get in touch whenever it suits you.