27 days out: Article 50 is the part of the EU AI Act that did not move to 2027

If you skimmed the AI-policy headlines in May and June, you probably now believe the EU AI Act moved to 2027. That is half true — and the half that is not true applies in 27 days.

Two clocks, not one

The Digital Omnibus — politically agreed 7 May 2026, adopted by the European Parliament on 16 June and by the Council on 29 June 2026 — rewrote the AI Act's timeline. But it rewrote only part of it:

Obligation	Old date	Current date
High-risk systems (Annex III) — biometrics, critical infrastructure, education, employment, migration; incl. the Article 26 deployer duties	2 Aug 2026	2 December 2027
High-risk systems embedded in products (Annex I) — machinery, medical devices, lifts, toys	2 Aug 2027	2 August 2028
Article 50 transparency — chatbot disclosure, machine-readable marking of AI-generated content, deepfake labelling	2 Aug 2026	2 August 2026 — unchanged
Article 50(2) machine-readable marking only, for generative systems already placed on the market before 2 Aug 2026 — a narrow Omnibus grace, limited to the marking duty alone	2 Aug 2026	2 December 2026
Penalties for Article 50 breaches — up to €15M or 3% of worldwide turnover	2 Aug 2026	2 August 2026 — unchanged

The European Commission's own AI Act page states it plainly: the transparency rules come into effect in August 2026, while the high-risk rules follow from 2 December 2027. So strictly there are three clocks, and the middle one is narrow: systems already placed on the market before 2 August 2026 get until 2 December 2026 for the machine-readable marking duty under Article 50(2) — and only for that duty. Chatbot disclosure and deepfake labelling apply from 2 August 2026 for everyone, with no runway; and a generative system placed on the market after 2 August 2026 gets no marking grace at all. Nobody can lean on the December date for anything except legacy marking.

Why this catches people

Article 50 is not a "high-risk" rule, so the postponement headlines did not apply to it — but almost nobody reads past the headline. And unlike the high-risk regime, Article 50 is broad: it covers any AI system that interacts with people or generates content. The Commission's AI Act Service Desk FAQ confirms it applies to AI agents. If you ship — or deploy — a chatbot, an assistant, an agent that talks to users or produces text, images, audio or video, this is your lane.

Since 10 June 2026 there is also an official yardstick: the AI Office's Code of Practice on Transparency of AI-Generated Content — concrete commitments on marking, watermarking and disclosure that make "are we ready?" a checkable question rather than a legal vibe.

What the data says: most shipped agents show no readiness signal

Hlido independently tests shipped AI-agent products and publishes the evidence. Our Article-50 Readiness Register scores each tested agent's public transparency-readiness signal — is the AI nature of the product disclosed, is there a clear stated purpose and access path, are the transparency-relevant surfaces present — from what the product actually shows, not what the vendor self-attests.

As of 6 July 2026, across 821 tested agent products:

• READY: 144 (17.5%) — clear transparency signal on the public surface
• PARTIAL: 123 (15.0%) — some signal, with gaps
• NOT-READY: 554 (67.5%) — no meaningful transparency-readiness signal we could verify

Every score links to per-agent evidence — what we checked, what we found, when. The register is machine-readable, free, and re-scored as agents change.

Two honest caveats. This is an independent signal built from public-surface testing — it is not legal advice and not a compliance certification, and a NOT-READY band does not mean a vendor is non-compliant; it means we could not verify a transparency signal from the outside — which is also exactly what a user, a buyer, or a market-surveillance authority sees first.

What to do with 27 days

1. Know your lane — and your clock. High-risk deployer duties (Article 26) bite in December 2027. Disclosure and deepfake-labelling duties bite on 2 August 2026 for everyone. Machine-readable marking bites 2 August 2026 for new systems, 2 December 2026 for systems already on the market.
2. Check the agents you ship or deploy — look them up on the register, or request a check if they are not yet tested.
3. Use the Code of Practice as the rubric. It is voluntary, but it is the closest thing to an official checklist for the marking and disclosure duties.
4. Build the evidence trail now. The December-2027 high-risk clock did not delete the diligence question; it extended the runway for it.

Hlido is an independent review desk in Prague. No vendor pays for a score; methodology is public; evidence is published per finding.