← Live scorecard
Regulation ledger

EU AI Act: what applies now vs. what got deferred

Last updated: August 21, 2026 · Re-checked when the law moves — deferral dates below are flip conditions, not trivia
By the AGI Scorecard team · methodology & independence · Not legal advice
The deadline didn't vanish — it split. As of August 21, 2026: the transparency duties apply (Article 50 — chatbot disclosure since Aug 2, 2026, no grace period for new systems), while the Digital Omnibus on AI — Regulation (EU) 2026/1744, adopted law, in force since July 27, 2026deferred the high-risk obligations to Dec 2, 2027 (Annex III) and Aug 2, 2028 (Annex I). Most confusion online comes from mixing those two halves up.

The obligation-by-obligation ledger

Law-firm memos explain this in prose; here is the same content as a dated ledger you can audit row by row. "Original date" is the AI Act as adopted in 2024; "current date" reflects Regulation (EU) 2026/1744 (the "Digital Omnibus on AI", Official Journal July 24, 2026, in force July 27, 2026).

ObligationArticleOriginal dateCurrent dateStatus
Prohibited AI practices (social scoring, manipulative AI…)Art. 52025-02-02unchangedIn force
AI literacy duty for providers & deployersArt. 42025-02-02unchangedIn force
General-purpose AI model obligationsCh. V2025-08-02unchangedIn force (enforcement powers from 2026-08-02)
Chatbot / AI-interaction disclosureArt. 50(1)2026-08-022026-08-02In force, no grace period
Machine-readable marking of AI-generated contentArt. 50(2)2026-08-02splitNew systems: now · pre-Aug-2026 systems: 2026-12-02
High-risk AI systems, standalone (Annex III)Ch. III2026-08-022027-12-02Deferred
High-risk AI embedded in regulated products (Annex I)Ch. III2027-08-022028-08-02Deferred

Instrument for every deferral row: Regulation (EU) 2026/1744. Sources for each row: Morgan Lewis (Aug 2026) · White & Case · K&L Gates (Jul 31, 2026) · European Commission · DLA Piper (Aug 2025).

The one everyone gets wrong: Article 50(2) is a split, not a deferral

Widely-repeated summaries say the content-marking duty "moved to December 2026." That is only half true. The grace period to December 2, 2026 covers only AI systems placed on the market before August 2, 2026. A system launched today must ship machine-readable marking of AI-generated content from day one. If you build on a third-party API, this lands on you as deployer — it is not just the model provider's problem.

The practical companion is the Commission's final Code of Practice on marking and labelling AI-generated content, published June 10, 2026 — a voluntary compliance tool for exactly these Article 50 duties (EC announcement, Jones Day summary).

Penalties in force today

Article 50 breaches have been enforceable since August 2, 2026: up to €15 million or 3% of total worldwide annual turnover, whichever is higher — for SMEs, whichever is lower (Article 99(4)). In Germany, the market-surveillance authority is live: the KI-MIG implementing act entered into force July 29, 2026, and the Bundesnetzagentur (competence centre "KoKIVO") has been the central AI supervisor since August 2, 2026 (Bundesnetzagentur).

Why this ledger lives on an AGI-timeline site

Because the two clocks are now officially out of sync. The AGI-2027 thesis says transformative capability arrives around 2027; the EU just moved its high-risk compliance clock to December 2027 and August 2028 — i.e., Europe's rulebook for high-risk AI now becomes fully enforceable after the window in which the strongest capability forecasts expect near-transformative systems. Whether that is prudent flexibility or regulatory lag depends on which side of the capability bet turns out right — which is exactly what the scorecard measures.

The part a chat answer goes stale on
62.5/100 of the AGI-2027 thesis still standing — the capability clock this regulation is racing against, recomputed only when a graded verdict flips
All eight verdicts, each with its pre-registered flip condition →

Dates to watch next

Frequently asked questions

Does the EU AI Act apply from August 2026?

Partially — and that split is the confusion. The transparency duties (Article 50: chatbot disclosure, deepfake labelling) apply since August 2, 2026. But the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force July 27, 2026) deferred the high-risk obligations: standalone high-risk systems (Annex III) to December 2, 2027, and product-embedded AI (Annex I) to August 2, 2028.

Does my chatbot have to tell users it's AI?

Yes — since August 2, 2026, with no grace period, unless it is obvious to a reasonably informed user that they are interacting with AI. This applies to deployers building on third-party APIs, not just model providers. Non-compliance risks fines up to €15 million or 3% of worldwide turnover (Article 99(4)).

Which EU AI Act deadlines were deferred?

Three moves, all made by Regulation (EU) 2026/1744: high-risk Annex III obligations moved from August 2, 2026 to December 2, 2027; high-risk Annex I (product-embedded) moved to August 2, 2028; and the Article 50(2) machine-readable marking duty got a grace period to December 2, 2026 — but only for systems already on the market before August 2, 2026. New systems must comply immediately.

What are the penalties for breaking the AI Act's transparency rules?

Up to €15 million or 3% of total worldwide annual turnover, whichever is higher (for SMEs, whichever is lower), under Article 99(4) — applicable to Article 50 breaches since August 2, 2026.

The live scorecard updates as models ship and verdicts change.

View the live scorecard →

Be first to know when a verdict flips

One email when one of the eight verdicts changes and the AGI-2027 Thesis Tracker moves — the single auditable score no other tracker has. Not a weekly newsletter; nothing arrives until the score actually moves.

Subscribe free →