On Sunday 2 August the transparency obligations in Article 50 of the EU AI Act began to apply. Nothing visibly happened, which is how most regulatory dates go. A week of writing about it comes down to three things worth keeping.
What actually happened on 2 August
A rule that had been public since 2024 became live. Providers of generative systems now have to ensure output is marked in a machine-readable format as artificially generated or manipulated. Deployers, meaning whoever publishes the result, have to disclose it. Deep fakes and AI-generated text published to inform the public on matters of public interest are named specifically.
No enforcement wave arrived on the Monday, and anyone who told you one would was selling something. What changed is that the obligation is now a fact about your publishing rather than a date in a calendar.
The three things worth remembering
One. Almost everyone reading this is a deployer, not a provider. Teams assume the duty belongs to the vendor whose tool they use. If you publish the output, the disclosure duty is yours. That was the first post of the week, a plain reading of what Article 50 asks you to do, and it is still the one to send to a colleague who has not looked at any of this.
Two. A label is a claim, and claims get tested. This was the thread running through the whole week. Disclosure is a statement you make on a surface you control, which is exactly what the rule asks for and exactly what stops being sufficient the moment somebody has a reason to check it. A checkbox is not proof set out why the records teams already hold, file dates, email threads, project history, cloud version history, all weaken at the same point: they are held by the party making the claim. The compliance-proof post covered what a record needs instead.
Three. Picking the wrong label is worse than picking none. The obligation is about accuracy. The four AI declarations went through AI generated, AI modified, AI assisted and human authored, and the honest finding is that most working teams are AI assisted more often than they would like to write down.
The part underneath all of it
Halfway through the week the subject changed and did not really change. There is no world copyright register: under the Berne Convention copyright arises automatically on creation in 181 countries, with no formality required, and almost none of those countries keep any record of it. The United States runs a register, and that registration does not travel, because copyright is territorial and national treatment carries the right across a border without carrying the American procedural record with it.
That is the same problem as the Article 50 one wearing different clothes. In both cases the law gives you something real and gives you no evidence of it, and the evidence is what the argument turns on when there is an argument. An AI disclosure and a claim of authorship both come down to: what was made, by whom, and when.
What to do about it
Sealing and declaring a file takes about five minutes, and the step that matters is the last one, where somebody outside your organisation verifies the record without an account and without contacting you. Every internal record fails at precisely that step.
The practical advice from the whole week is unglamorous. Put the record at the point of publication rather than at the point of audit. Do not try to seal four years of back catalogue. Start with client deliverables, anything published into the EU, and anything on a public-interest topic.
What is coming next week
Week two goes deeper into the mechanics: what a qualified timestamp actually is, how verification works from the recipient's side, and what separates a provenance record from a compliance checkbox in practice.
The full week
- 1 August: what Article 50 actually asks you to do, in plain terms. Who it binds, what is in scope, what labelling means.
- 2 August: Article 50 is live. Here is how to prove your content is compliant. The four parts of a record that survives being questioned.
- 3 August: a checkbox is not proof. Why file dates, email threads and version history all give way at the same point.
- 4 August: the four AI labels every EU publisher needs. How to decide in under a minute, and why understating is the expensive mistake.
- 5 August: the world has no copyright register, except the United States. What automatic protection does not give you.
- 6 August: why a US registration does not protect you outside the United States. Territoriality, national treatment, and what travels instead.
- 7 August: how to seal and declare a file in under five minutes. The whole flow, including the step people skip.
If you read one thing, make it the Article 50 page, which holds the whole argument in one place.
