The EU AI Act Clock: What Applied on 2 August 2026, and What Moved to 2027

The AI Act entered into force on 1 August 2024 and became generally applicable on 2 August 2026, with a staged schedule around it. Prohibited practices and AI literacy obligations applied from 2 February 2025. Governance rules and obligations for general-purpose AI models applied from 2 August 2025.[1]

Article 50 is the piece that became live on 2 August 2026. Providers must design systems that interact directly with people so those people know they are dealing with AI, unless it is obvious. Providers of systems generating synthetic audio, image, video, or text must mark the output in a machine-readable format that allows it to be detected as generated or manipulated. Deployers must disclose deepfakes and must disclose AI-generated text published to inform the public on matters of public interest, subject to the exceptions in the Article. The European Commission's guidance is explicit that Article 50 applies to all in-scope systems from that date regardless of when they were placed on the market.[2][3][4]

There is exactly one transitional carve-out, and it is narrow: for generative systems already on the market before 2 August 2026, the marking-and-detection duty in Article 50(2) applies from 2 December 2026. Content generated before 2 August 2026 does not need retroactive labeling, though the Commission encourages it; text generated before that date but published after it does need to be labeled.[2][4]

The deferral that generated headlines was elsewhere. Under the Digital Omnibus amendments, the Chapter III obligations for high-risk systems moved out: Annex III high-risk use cases to 2 December 2027, high-risk systems embedded in regulated products to 2 August 2028, with a longer horizon for certain legacy public-authority systems. The transparency date did not move with them.[1][5]

The reporting compressed two different things into one word. "Delayed" described the high-risk conformity work — risk management systems, technical documentation, human-oversight design, conformity assessment — which is genuinely later. It did not describe disclosure and content marking, which are now.

Disclosure also lands on a different team than compliance usually expects. Telling a user they are talking to an AI is a product and design decision in the interface, not a clause in the terms of service. Machine-readable marking is an engineering change in the generation path. Both are cheaper to build now than to retrofit into a shipped product under an enforcement question.

The obligations also cut across the agent trend directly. A chat assistant that acts on a customer's behalf is a system interacting with a natural person. A workflow that drafts public-facing text is generating content that may need marking. The more the interface disappears into an agent, the harder it is to argue that the AI involvement is obvious.

1) Inventory where your product speaks to a person or generates content, including internal tools that produce customer-facing output. Scope precedes compliance.

2) Put the disclosure in the flow. A visible, plain-language notice at the point of interaction is what the obligation is about; burying it in a policy page is not the same control.

3) Wire marking into the generation path, not the publishing step. Machine-readable provenance that survives export and re-upload is the version that holds up.

4) Keep an evidence pack: which systems are in scope, what disclosure each shows, how outputs are marked, when it shipped, and who owns it. Readiness is a record, not an intention.

5) Do not spend the 2027 deferral. If a use case is likely Annex III high-risk, the documentation, data governance, and human-oversight work is the same work — earlier is cheaper, and the deferral can be revisited.

6) Track your other regimes separately. Sector rules, US state law, and contractual obligations do not follow the AI Act calendar.

This is an operator's reading of published deadlines, not legal advice, and ARC is not your counsel. The Commission's Article 50 guidelines are interpretive rather than binding, national implementation and enforcement practice vary, and the Digital Omnibus amendments changed dates that had previously been reported differently — which is precisely why every date here carries a source. Confirm scope and obligations with qualified counsel before making a compliance decision.