Nearly four weeks ago, parts of the EU AI Act still became applicable. The high-risk deadlines moved. Others did not.
The EU's Digital Omnibus entered into force on 27 July, and most coverage led with the deferral: the main high-risk requirements moved to December 2027 for Annex III systems, and to August 2028 for product-related Annex I systems.
Here is the part that travelled less well. Not everything moved.
- Most Article 50 transparency duties applied from 2 August 2026, covering certain AI interactions, synthetic content and deepfakes. Systems already on the market have until 2 December 2026 for the machine-readable marking requirement.
- Obligations for providers of new general-purpose AI models have applied since 2 August 2025. Models on the market before then generally have until 2 August 2027.
- Most prohibited-practices rules have applied since 2 February 2025, and a small number added by the Omnibus apply from 2 December 2026.
The practical risk is simple. An organisation that heard "the high-risk rules are delayed" and paused its whole programme may now be out of step with the provisions that continued on their original timetable.
Relief creates drift too. The belief about what applies and the reality of what applies can separate quietly, and the headline is not going to correct it.
If you provide or deploy AI that interacts with people, generates or alters content, or may fall within the Act's EU scope, the useful exercise this week is not a new programme. It is one check per system: what role are we playing, which provisions might cover it, and which date applies?
I am not a lawyer and this is not legal advice. It is a prompt to go and look.