How to avoid needing to stamp text as AI Generated
Nobody wants to stamp AI GENERATED on their own work. Not journalists, not marketers, not you. It reads as low effort.
If you don’t stamp it and you don’t claim the exemption, you can be fined, even where the AI only corrected grammar, rewrote sentences, and gave structure to text you wrote yourself.
In this post, we go over the easiest way to claim it, so you never have to mark reviewed text as AI Generated.
Since the 2nd of August, 2026, if you publish AI generated text to inform the public, you must disclose that it is AI generated. But there is an exemption: human review and editorial responsibility.
They are big sounding words, but they are simple.
But first of all, what type of content is covered and should be disclosed as AI Generated?
You are to disclose if you post about:
- Thought leadership, data studies written to be quoted by journalists, or any content built to inform the public (such as this post).
- Elections, health, safety, courts, and local or international government.
If you use AI to write product pages, support copy, or ads (without the topics above, of course), you are fine.
What to record
- What was published (link or file)
- How AI was involved, and what it did
- Which AI did it
- Who reviewed it
- When it was reviewed
- What was changed, or why nothing was changed
- Who holds editorial responsibility
- When it went live
- Store it with an immutable third party to avoid tampering claims
Record it when it happens, not when you are asked
A record made after someone complains proves nothing. It only proves you can type.
If you cannot show the record existed before the article went live, it is a statement, not evidence. That is the reason for the third party. The timestamp does the work, not the text.
What counts as human review
There is one question behind all of this: could the review have caught a problem?
It counts if a named person read it before publication, had the authority to stop it, and either changed something or can say why nothing needed changing.
It does not count if a second AI checked the first one, if the approval was logged after the piece went live, or if one person cleared forty articles in nine minutes.
Who holds editorial responsibility
This is the second condition and it is the one people forget. Someone has to be accountable for the piece going out. A named person, or the company itself.
If you cannot say who it was for a given article, the exemption does not apply, no matter how good the review was and trust us, they will know its AI made as Anthropic has been hiding signatures in the text for a while now.
Where the exemption does not reach
Deep fakes. Images, audio and video of real looking people or events get no human review carve out. Editing them changes nothing. You disclose.
Chatbots. A separate rule. If a person is talking to your AI, tell them at the start.
Everything else. Advertising rules and consumer protection law still apply on top of this.
Who reports you
Anyone can. Not just a regulator, not just a customer. A competitor who reads your blog, a reader who did not like your angle, a contractor you stopped paying.
They file, the authority asks about one article, and the answer is either on file or it is not. Penalties run to 15 million euro or 3% of worldwide turnover, but the fine is rarely the expensive part. The finding is.
You probably do not need to label your text. But you almost certainly need to be able to prove you did not have to.
Those are two different jobs, and only one of them needs doing today.
Art50 does the second one.
You log the review as it happens. We seal it and publish the proof to a public transparency log that you do not control, and neither do we. If anyone ever asks who reviewed a piece and when, you send them a file and the conversation ends there.
Free for journalists, newsrooms, charities, NGOs and public bodies.
