AI-generated content: 4 pieces of evidence to keep to ensure you remain defensible

Picture the scene. Your team publishes a position paper. A journalist runs it through a detector, and the detector says: "This text has been through an AI." This is no longer science fiction. It has been possible since this summer. And there is one detail that changes everything: the detector cannot tell whether the AI wrote your text or just tidied it up. Let me explain how it works, what the law says about AI-generated content, and the four simple records to keep so you can rest easy.

How AI-generated content gets recognised

A mark nobody can see

Let's start at the beginning. In August, Anthropic, the company behind Claude, announced that its new language models would leave a mark in every text they generate. It is called a watermark, like the one on a banknote. Except that you will never see this one.

Anthropic is clear that nothing is added to the text. No hidden characters, no invisible code.

So where is this mark? It is in the choice of words.

Here is the idea. When an AI writes, it constantly hesitates between several words that are equally good. "The meeting was long" or "the meeting was endless"? To you, it makes no difference. Normally, chance decides. With the watermark, it is no longer chance but a secret key. Over hundreds of words, those choices form a pattern. And whoever holds the key can find it.

This is not Anthropic's invention, by the way. The method comes from Google researchers, who published it in the journal Nature in 2024. Google already uses it in its own AI, Gemini. And according to their tests, users see no difference in quality.

The detector sees a visit, not an author

The watermark says "the AI was here". It does not say what the AI did. So there is no way to tell a text written by Claude from a text that Claude heavily reworked.

In practice, that gives you 3 situations.

You ask the AI to fix your spelling. Almost every word is still yours, so there is next to nothing to detect.

You ask it to restructure your text in depth. Now it is choosing a lot of words for you. Your text can come out marked, exactly like a text generated from start to finish.

You ask it to translate. Here, every single word of the translation is chosen by the AI. The translated text is therefore fully marked, even if the original is entirely yours. If your organisation publishes in several languages, remember this one.

The watermark does have limits. On a very short text, it is hard to detect. And if you rewrite everything, word by word, it disappears.

Who is allowed to check?

Not everyone, for now. The detection tool is reserved for certain organisations: regulators, law enforcement, the media, fact-checkers (journalists who specialise in verifying facts), researchers, schools and universities, and European civil society organisations. Anthropic says the list will grow.

Two reassuring points along the way. The watermark contains no information about you, your organisation or your conversations with the AI. And it says nothing about who the author of the text is in the legal sense.

What the law says: Article 50

Why is Anthropic doing this now? Not for fun. European law requires it.

Europe has adopted a major regulation on artificial intelligence, usually called the AI Act. Its Article 50 is about transparency, and it has applied since 2 August 2026. The thing to understand is that it contains two different obligations, for two different players. If you want to go further, the European Commission sums up these rules on one page.

First obligation: AI makers must mark

That is the watermark. Every AI maker operating in Europe has to mark what its tools produce. New models since 2 August, older ones by 2 December 2026 at the latest. Anthropic signed up in July, along with around 190 other signatories.

In other words, by the end of the year this will not be about Claude alone. It will cover more or less every tool your teams use.

Second obligation: whoever publishes must disclose

This one is for you. And many people have never heard of it.

If your organisation publishes a text generated or modified by an AI, and that text informs the public on a matter of public interest, you have to disclose it. Whatever tool you used.

What does "public interest" mean? The European Commission, in its guidelines, takes a broad view: health, environment, fundamental rights, consumer protection, the economy, politics, science. If you work for a European association or an institution, that is pretty much your list of topics.

Fortunately, there is an exception, and it is common sense. You do not have to disclose if two conditions are met: a human has genuinely reviewed the text, and someone takes responsibility for it. Mind the word "genuinely". A thirty-second glance before clicking "publish" does not count. The Commission is strict about that.

And if you break the rule? Fines can reach 15 million euros or 3% of worldwide annual turnover.

To be clear, I am not a lawyer: for your specific case, take advice.

The real risk is your reputation

Let's be honest. The rule is brand new, and the fine is not what should worry you first.

Go back to the opening scene. Your team writes a position paper. Someone asks the AI to restructure it a little. A journalist runs it through the detector. Result: "probable involvement of Claude". You can see the headline already: "Their position paper was written by an AI." When all the AI did was tidy things up.

At that point, answering "but we did review it" carries little weight. What matters is being able to prove it.

There is something else you should know: the detector gives a probability, not a proof. Researchers at ETH Zurich even managed to plant a fake watermark on texts written by humans, in more than 80% of cases and for less than 50 dollars. So a fully human text can, in theory, come out marked. One more reason to keep your own evidence.

AI-generated content: the four records to keep

The good news is that the answer is the same for the law and for your reputation: being able to prove who reviewed, corrected and approved. And honestly, it does not take much. No software, no complicated procedure. Just 4 records, filed with every text you publish.

1. The version written by your team, before the AI. This is what shows that the ideas are yours.
2. The name of the person who reviewed the final version. Properly reviewed, line by line.
3. The date they approved it.
4. The name of the person or organisation that signs it off, meaning who takes responsibility for publishing it.

Where do you keep all this? In your website's revision history, in a shared folder, in a simple sheet per document. It does not matter. What matters is doing it every time. A system you apply one time in three protects nobody.

One last tip: put in writing what your teams are allowed to do with AI. Correcting, restructuring, translating and writing a first draft are not the same thing, and they do not have the same consequences. And if one day you publish a text that nobody has genuinely reviewed, the answer is simple: disclose it.

Key takeaways

The watermark does not change how you work with AI. It changes one thing: you can no longer claim what you cannot show. AI-generated content, or even content the AI merely reworked, can now be detected by people other than you. And the law says human review is what exempts you from disclosing it.

Four records per text, and you are covered. If you use AI without keeping any trace, you will have nothing to show the day someone asks.

At Simpl., we help associations and institutions put this framework in place: an editorial guide, human validation, a written trail. Let's talk.

Useful resources

• European Commission: the Article 50 transparency rules on one page. The best starting point: who has to do what, the penalties, the December 2026 deadline.
• European Commission: the guidelines on Article 50. The text that explains what counts as a matter of public interest and what counts as human review.
• The text of Article 50, on the official AI Act Service Desk.
• The Code of Practice on Transparency of AI-generated Content. Section 2 is the one for organisations that publish.
• The EU icons for labelling AI-generated content. Useful the day you need to disclose a text.
• Anthropic: how Claude's text watermark works. The maker's own answers, including on proofreading, translation and the limits of detection.
• Google DeepMind in Nature: the original study on text watermarking. For those who want the scientific detail.