The EU AI Act: Claude now marks all AI-generated content — what it means for your business
The EU AI Act: Claude now marks all AI-generated content — what it means for your business
When you write a text with Claude, the AI now hides an invisible mark in the words. You and your customers can't see it, but it can reveal that the text came from an AI. The reason is the EU AI Act, the EU's new law on artificial intelligence, which since 2 August 2026 requires AI-made content to be recognisable.
This isn't news you need to worry about, but you should understand it. Below we go through what Claude does, what the law says and what actually applies to you as a business owner.
How Claude marks text and images
According to Anthropic's help centre, two methods are used:
- Watermark in text. Claude weaves a pattern into how words are chosen. It doesn't change meaning, quality or readability, and it can survive copying and light editing. It applies to all of Claude's products, including when you use Claude through Amazon, Google or Microsoft.
- Digital signature on files. Images in formats like PNG and JPEG get a hidden label following the C2PA standard. Think of it as a receipt baked into the image: it shows the file was processed by Claude.
Anyone can check a file for free in Anthropic's Claude Content Checker at claude.com/check-content. Detecting the watermark in text, however, is open only to selected groups such as authorities, journalists and researchers.
What the marking can't tell you
The mark is a clue, not proof. Anthropic itself says a mark only means the content may have passed through Claude. Perhaps the AI only translated or proofread a text you wrote yourself.
The reverse is also true: if there's no mark, the text isn't necessarily human-written. Heavy rewriting, very short texts, screenshots and re-saved files remove the mark. Don't expect a perfect AI detector.
The EU AI Act: what applies to you as a business owner
The duty to mark sits with the AI companies, such as Anthropic, OpenAI and Google. If you use Claude, you don't have to add a mark yourself. But the Act also has rules for those who use AI in their business:
- Chatbots: If an AI talks to customers on your website, the customer must be told it's an AI.
- Realistic AI images and AI videos of real people or events must be marked as AI-made.
- Texts informing the public about matters of public interest must be disclosed as AI-written, unless a human has reviewed them and takes editorial responsibility.
An ordinary blog about your products, an email or a quote normally falls outside. You don't need to write "made with AI" on every text. Breaking the rules can still cost up to 15 million euros or 3 percent of turnover, so it pays to stay on top of it.
The heaviest requirements (for so-called high-risk systems, such as AI that screens job applications) have, according to several legal reviews, now been pushed back to December 2027. The marking rules, however, apply now. AI systems already on the market before 2 August also get a transition period until 2 December 2026. This text isn't legal advice, so ask a lawyer if your case is unclear.
Three things to do this week
- Go through where you use AI toward customers: website chat, automatic emails, images in ads. Is there an AI the customer might take for a human?
- Mark realistic AI images of people or places, for example with a small text in the corner or beneath the image.
- Keep a human in the loop. Read and approve what the AI writes before it's published. It gives better quality and a safer legal position.
If you already use Claude, you don't need to change anything. If you'd like help reviewing how you use AI in a way that both saves time and holds up under the EU AI Act, we're happy to help as an AI consultant. The basics of the assistant itself are in our guide to Claude AI.
Frequently asked questions
Does Claude mark AI-generated text?
Yes. Claude adds an invisible watermark to all text created by models that support it, across all of Claude's products. The mark doesn't change how the text reads and can remain when the text is copied or lightly edited. Images and files in formats like PNG and JPEG instead get a digital signature following the C2PA standard.
What does the EU AI Act require of businesses that use AI?
Under Article 50 of the EU AI Act, businesses must tell people when they are talking to an AI (such as a chatbot), mark realistic AI images and AI videos as AI-made, and disclose AI-written texts on matters of public interest published without human review. The rules have applied since 2 August 2026. The marking of the AI content itself rests mainly with the AI companies.
Do I have to say that a text was made with AI?
Normally no. Ordinary business texts such as product descriptions, emails and quotes are not covered by any AI-labelling requirement. The exception is texts published to inform the public on matters of public interest that haven't been reviewed by a human who takes editorial responsibility. Realistic AI images of real people must always be marked.
Can you tell if an image was made by Claude?
Yes, if the image is a PNG or JPEG file that still has its hidden C2PA label. You can check it for free in Anthropic's Claude Content Checker at claude.com/check-content. The label disappears if the image is re-saved, screenshotted or converted, so an image without a label can still be AI-made.
What does it cost to break the EU AI Act's marking rules?
Fines for breaching the transparency requirements of the EU AI Act can reach 15 million euros or 3 percent of a company's global annual turnover, whichever is higher. For small businesses the lower of the two amounts applies. This text is general information and not legal advice.