What the Article 50 deadline actually changed
2 August 2026 passed quietly. Here is what actually changed that day, why the silence does not mean low risk, and what to check now.
Guides to what the law requires, kept separate from how Clause50 does it. If you are starting from scratch, read the whole-Act guide first — it places every other page here in context.
The four risk classes, what each one constitutes, provider versus deployer, the timeline, and where each class leads next.
The four transparency duties, who each one binds, the exemptions, and the evidence an auditor asks for — the obligation most systems in production today actually have.
Which systems Annex III makes high-risk, what Annex IV's nine-point technical file must contain, and why that evidence can only be produced by someone recording it at the time.
Articles 53 and 55 bind general-purpose AI model providers. If you build on someone else's model, those are not your obligations — Article 50's transparency duties are.
Every obligation our rule pack models, rendered live from the pack itself and split into what is in force now and what arrives later — plus a plain statement of where we stop.
The ten practices the Act bans outright, in plain English — two of them from 2 December 2026 — and why a system recorded as one is refused an evidence pack rather than offered a lesser one.
2 August 2026 passed quietly. Here is what actually changed that day, why the silence does not mean low risk, and what to check now.
A completed checklist proves someone believed a statement was true. Article 50 asks whether the information was actually given — a different, and harder, question.
Two different products get sold under the same name. A category comparison of what each actually produces, and the one question that tells them apart.
If an AI system touches personal data, the GDPR already applies — today, not in 2027. The duties that matter most for AI, where they meet the AI Act, and why Clause50 answers both from one system record.