If you read any tech press this month, you heard that the EU "delayed the AI Act". That is half true, and the half that is missing lands on 2 August.
Here is the actual position, from the primary sources.
What happens on Saturday 2 August
- Article 50 transparency obligations go live. If you deploy an AI system that interacts with people (a chatbot on your website counts), you must tell them they are talking to AI. If you deploy AI that generates or manipulates content, disclosure duties apply. This was not delayed. Commission guidelines, 20 July 2026
- Enforcement powers for general-purpose AI switch on. The Commission can fine GPAI model providers up to EUR 15m or 3% of worldwide turnover from 2 August. The model providers' obligations have applied since 2025; what arrives now is the ability to punish. Regulation (EU) 2024/1689, Chapter V
What was actually delayed: the "high-risk" obligations under Annex III (hiring tools, credit scoring, biometrics and similar) move to 2 December 2027, and embedded high-risk systems to August 2028. That is the Digital Omnibus amendment the headlines were about. Council press release, 29 June 2026
The detail almost everyone is getting wrong: as of this writing, the Omnibus has been signed but has not yet appeared in the Official Journal. Until it is published and in force, the original AI Act text remains the binding law. The delay is coming, but on Saturday morning the delay itself is not yet on the statute book, while the Article 50 and GPAI provisions are. EUR-Lex Official Journal, L series (checked 24 July 2026)
One genuine grace period exists, and it is narrower than reported: generative AI systems already on the market before 2 August get until 2 December 2026 to implement the machine-readable marking mechanism under Article 50(2). That is it. New systems get no grace, and the duty to disclose AI interaction to users is not covered by it. Some coverage has blurred this with a separate new prohibition (on non-consensual intimate imagery) that also transitions on 2 December 2026. They are different provisions.
If you are a UK SME asking "does this even reach me": the AI Act applies where your system's output is used in the EU, regardless of where you sit. A UK company with EU customers using its chatbot is in scope for Article 50.
Meanwhile, Westminster rearranged itself in 48 hours
While everyone watched Brussels, the UK side moved more in two days than in the previous six months.
- DSIT no longer exists. The department that owned UK AI policy was abolished on 21 July in the new Prime Minister's machinery-of-government changes. AI policy and the AI Security Institute move to the Cabinet Office under a new AI Taskforce, and for the first time an AI minister attends Cabinet. gov.uk
- The ICO is becoming the Information Commission. Seven non-executive directors were appointed on 15 July and recruitment for the Commission's first Chair is open until 19 August. The single-Commissioner model is ending, as provided by the Data (Use and Access) Act 2025. gov.uk, 15 July 2026
- The FCA published the Mills Review on 6 July, the first review of its kind by a financial regulator anywhere, and asked government to consider whether large language model providers should be brought inside the FCA's regulatory perimeter. fca.org.uk
- Consultation worth your time: the DRCF call for input on consumer interest and AI is open until 2 September. drcf.org.uk
Three things to do this week
- 01
Inventory your customer-facing AI. Anything that talks to people or generates content, list it. If any of its output reaches EU users, Article 50 applies to you from Saturday.
- 02
Check your disclosures. "You are chatting with an AI assistant" where users see it. The Commission's 20 July guidelines set out what good looks like.
- 03
Diarise 19 August and 2 September. The Information Commission Chair appointment will shape UK enforcement culture for years, and the DRCF consultation is a rare chance to be on the record early.
The Blackletter Brief tracks UK and EU AI governance from official sources only: legislation, regulators and standards bodies, every claim linked to the primary record. No law firm marketing, no speculation. Built by Blackletter, a UK AI governance intelligence platform launching soon. Forward this to the person in your company who quietly worries about this stuff.