ChatGPT watermark in the EU: what it means for your firm

What OpenAI has announced

A watermark is a signal that lets you tell later where a piece of content came from. On 5 October OpenAI announced it will add its own, called textGrain, to ChatGPT and Codex text in the European Union "over the coming weeks", according to BleepingComputer. It is not a hidden symbol: the model slightly adjusts its choice of words, leaving a statistical pattern that a detector can look for. Nothing shows when you read or copy the text. Outside the EU it will not be the default for now, and anyone using the API, the connection software uses to talk to the model, can switch it on for some models, though it comes turned off.

The detector will not be public. Hipertextual explains that, at first, OpenAI will only accept applications from researchers and expert organisations, and links the change to the EU's AI rules. Article 50 of the AI Act, in force since 2 August 2026, requires those offering systems that generate text to mark the output in a machine-readable way.

The product pages of a bike shop

Imagine you run a bike shop with a repair workshop. For months ChatGPT has been helping with the product pages on your website, the monthly newsletter and replies to reviews. You paste in a few details from the supplier, it hands back something that reads well, and you publish it between one repair and the next.

One Tuesday a customer walks in, phone in hand. The page for one bike says it has hydraulic brakes, and it does not. Nobody meant any harm: the text came out that way and nobody checked it against the catalogue. That same afternoon you read about the watermark and wonder whether people will now be able to tell that a machine writes your copy.

Why do so many businesses write with AI without any rules?

The scene is made up, but it is a familiar one. AI slipped into small businesses through the back door: someone tried it for a tricky email, it worked, and it stayed. Nobody decided which texts it is used for, what gets pasted into it or who looks at the result before it goes out.

That is understandable. When there are three of you, nobody writes a usage manual. The trouble starts as volume grows: twenty product pages a month, a newsletter, replies to customers. Every text that goes out without a second read is a small gamble.

What an unchecked text can cost you

The watermark is not the risk. What costs you is a wrong spec on a product page, a promise you cannot keep in a reply, or a newsletter that sounds just like a competitor's. You pay for that in returns, in annoyed customers and in trust.

Nor is the mark a reliable way to know what a machine wrote. OpenAI's own evaluation, as reported by BleepingComputer with a 1% false-positive target, gives these figures (a token is a fragment of a word that the model works with):

• 200-token answers on psychology: around 80% are detected.
• 400-token answers on psychology: around 95%.
• 400-token texts with 10% of the words swapped for synonyms: detection falls from about 92% to 66%; with 25% swapped, to 17%.

OpenAI also warns that finding no mark does not prove a person wrote the text, and that finding one does not reveal who generated it or how much was edited afterwards. Hipertextual adds that it does not confirm the content is accurate either.

What should a small business that writes with ChatGPT do?

Nothing urgent, and certainly not hide it. What is worth an afternoon is writing down how your business uses AI. One page is enough:

• List the texts that already go out with AI help: product pages, emails, social posts, review replies, quotes.
• Decide where it is used and where it is not. Anything that commits the business, such as prices, delivery times or terms, deserves extra care.
• Name the person who checks each type of text before it is published, and what they check: facts, prices, tone.
• Agree on what never gets pasted into an AI tool, such as customers' personal details or contracts.
• Review the page every few months: the tools change quickly, and so do the rules.

As for the law, Article 50 says that anyone who publishes AI-generated text to inform the public on matters of public interest must disclose that it was AI-generated, unless the text has gone through human review or editorial control and someone takes responsibility for publishing it. If you are unsure about your own case, ask an adviser. Either way, checking your texts is good practice.

How we help at digizone

At digizone we offer technology consulting: we sit down with you, look at where AI fits in your business and where it does not, and help you put tools and rules in order so your team uses it sensibly. If you need a tool of your own, we build it.

One example of AI doing the writing inside a business is Docufy. Law firms specialising in road traffic accidents were spending much of their time on manual work, such as gathering paperwork or drafting documents. We built a platform with automations and with AI that generates the documents, where each firm can use its own templates. What the professionals value most is the time they save gathering paperwork and producing documents.

A first rule, not a manual

You do not need a forty-page policy. The idea behind an MVP, a minimum viable product, works here too: start with the smallest thing that already solves a real problem, and improve it as you use it.

In the bike shop, that first version would be a rule for product pages: AI drafts them, and before anything goes live someone checks every spec against the supplier's catalogue. Later, a small tool could do that check on its own and flag anything that does not match. The newsletter and the review replies would follow.

If you would like to see how it would work for you, our free workshop, in Spanish, walks through it step by step.