Why there is so much confusion about this date

The EU AI Act does not have a single "effective date" — it has a staggered application schedule by risk category, published since 2024. August 2, 2026 corresponds to the full application of a specific block of obligations, not the entire regulation at once. This is why press articles mention the date with varying levels of precision.

To complicate the picture further, in May 2026 the European Commission agreed to specifically postpone the obligations for high-risk AI systems until December 2027 — a delay of more than a year from the original schedule, in response to industry pressure over the lack of harmonized technical standards ready in time.

What DOES apply from August 2, 2026

  • General-purpose AI model (GPAI) governance: transparency and technical documentation obligations for providers of general-purpose foundation models, a category that covers the large commercial LLMs.
  • Penalties regime: national supervisory authorities (such as the one already active in Finland since January 2026) have full capacity to impose administrative penalties within this block.
  • Institutional governance: the European and national supervisory structure (AI Office, market authorities in each country) must be fully operational.

What does NOT apply yet (postponed to December 2027)

  • High-risk system obligations: conformity assessment, registration in the EU database, and specific risk management requirements for systems used in hiring, credit scoring, critical infrastructure, education and healthcare.
  • Third-party conformity certification: the notified bodies that certify high-risk systems do not yet have the harmonized standards infrastructure needed to operate at scale.
The risk of this postponement: some companies are reading the high-risk postponement as "the AI Act was relaxed" in general, when in fact the general-purpose model obligations continue on course unchanged. Confusing the two blocks can lead to underestimating compliance obligations already in force.

What a company operating in or selling to the EU should do

If your product uses a general-purpose model (any third-party commercial LLM) as a component, your transparency obligation toward end users about the use of generative AI is already in force and did not move with May's postponement.

If your system falls into any high-risk category (HR, credit, healthcare, education, critical infrastructure), you have until December 2027 as the formal deadline — but building the risk management and documentation processes now, instead of waiting for the last quarter of 2027, remains the safest decision given the EU's track record of not postponing again once a date has been set a second time.