August 19, 2026

What the AI Act changed for anyone who publishes

The transparency obligations of the European AI Act took effect on 2 August 2026, and two of them concern anyone who writes with AI. One applies to the companies that build the models. The other may apply to you.

This is not legal advice. It is how we read the rules that shape our own product, and we would rather you knew than found out from a client.

The models now mark their output

Providers of generative systems have to mark what their systems produce in a machine-readable format, so that the output can be detected as artificially generated. For images this had a shape already: metadata, provenance credentials. For plain text it is newer, and it works by shaping the choices the model makes as it writes, which leaves a statistical signature a matching detector can read.

You cannot see it. You will not see it in a diff. It survives light editing, and how much editing breaks it is genuinely unclear, which is a useful thing to know before you rely on either outcome.

This is not the same object as an AI detector. A detector guesses from style, badly. This is a mark deliberately placed by the provider, and read with the provider's own key. Very different reliability, and very different politics.

Some publishers have to disclose

The second obligation lands on the people using the systems. If you publish AI-generated or AI-manipulated text to inform the public on matters of public interest, you have to disclose that it was artificially generated.

Read the scope carefully before you panic. It is about publication to the public on matters of public interest, not about your internal memos, your support replies or your client proposals. And there is an exemption that matters: it does not apply where the content underwent human review or editorial control and a natural or legal person holds editorial responsibility for it.

Where that leaves a rewriting tool

We could have stayed quiet about this. Most tools in our category have. We would rather say it plainly: text processed through our AI rewriting may carry the provider's provenance mark, because it passed through a model. We cannot promise otherwise, and anyone in this field who promises otherwise is either guessing or lying.

What our AI tier does is make your text read like a person wrote it. That is a real thing and it is what most people need. It is not a claim about provenance.

The human tier

A Youmanize human review is rewritten by hand by a professional reviewer. No model touches it. Every reviewer signs a no-AI pledge when they join, and certifies on each individual delivery that the text they are handing back was written by them.

No machine ever generated it, so there is no provider mark to carry. And when disclosure rules are in play, a human wrote it and a person holds responsibility for it. Delivered within 24 hours, with the reviewer's first name attached.

Our position on the marking

Marking generated text is, on balance, good. It makes provenance a technical fact rather than an argument, and it is a better instrument than the detectors that have been failing students and non-native speakers for three years.

It costs the industry a comfortable belief: that you can run a text through a model and call the result yours. That was never quite true. It is now checkable.

Back to the blog