What actually changes on 2 August 2026 (and what does not)
2 August 2026 is the date Regulation (EU) 2024/1689 — the EU AI Act — becomes generally applicable under its Article 113. Expect a noisy week of buyer emails and countdowns. The precise picture is calmer: what starts that day is narrower than "the AI Act takes effect", and part of the Regulation has already been in force for eighteen months.
What starts to apply on 2 August 2026
For software vendors, the substance of the date is Article 50 — the transparency obligations. Its transparency duties apply in full from that day; the Digital Omnibus did not postpone them. The table below maps each paragraph and who owes it; the full analysis, including who counts as provider and who as deployer, is in our Article 50 guide.
One rule to hold onto: a company that ships a third-party model under its own brand is the provider of the resulting system for Article 50(1) purposes — the disclosure duty toward your users does not stay with OpenAI, Anthropic or Google.
Also new that day: national supervision of the Article 4 AI-literacy measures begins. The Article 4 duty itself is not new — see below.
What does not change that day
- The prohibitions (Article 5) — nothing new. They have applied since 2 February 2025. Anyone telling you prohibited practices "kick in this August" is eighteen months late.
- Article 4 AI literacy — already in force since 2 February 2025. The Digital Omnibus softened it from guaranteeing a sufficient level to supporting the development of AI literacy — an obligation of means, not of result. What starts on 2 August 2026 is national supervision.
- The high-risk regime — not yet. The Digital Omnibus (adopted by Parliament on 16 June 2026 and by the Council on 29 June 2026) moved Annex III use cases to 2 December 2027 and high-risk AI embedded in regulated products to 2 August 2028. Near Annex III territory? Start with the Annex III guide.
- Article 14 human oversight — only for high-risk systems. A limited-risk system's obligations are the Article 50 duties; oversight measures on it are voluntary good practice, not an Act requirement.
- Governance and penalties chapters — applicable since 2 August 2025. Fines are imposed by Member States; national enforcement regimes are still being put in place (Spain's is a bill in parliamentary passage).
The date map at a glance
| Provision | Status on 2 August 2026 |
|---|---|
| Art. 5 prohibitions | Already applied — since 2 Feb 2025. Nothing new. |
| Art. 4 AI literacy | Already applied — since 2 Feb 2025. New: national supervision begins. |
| Art. 50(1) — AI interaction disclosure | Applies from this day (provider). |
| Art. 50(2) — synthetic-content marking | Applies from this day (provider) — transition to 2 Dec 2026 for systems already on the market. |
| Art. 50(3) — emotion recognition / biometric categorisation disclosure | Applies from this day (deployer). |
| Art. 50(4) — deepfakes and AI-generated public-interest text (attenuated disclosure for creative and satirical works; exemption for editorially reviewed text) | Applies from this day (deployer). |
| Art. 50(5) — information clear, distinguishable, at first interaction or exposure | Applies from this day (both, per paragraph). |
| Annex III high-risk regime | Not yet — 2 Dec 2027. |
| High-risk AI embedded in regulated products | Not yet — 2 Aug 2028. |
The one transition inside Article 50: marking, until December
The single accommodation inside Article 50 concerns the machine-readable marking duty of 50(2): systems already placed on the market benefit from a transition until 2 December 2026. It is a grace period for marking only — it does not touch 50(1) chatbot disclosure or any other paragraph — and it buys existing systems four months, not a waiver.
What a vendor should have ready that week
- Your role, per AI feature. Provider or deployer decides which paragraphs you owe — shipping under your own brand makes you provider. Six questions in the free Article 50 checker map it, no email required.
- The 50(1) disclosure live in the product, unless the AI interaction is genuinely obvious to a "reasonably well-informed, observant and circumspect" person.
- A documented 50(2) position: marking in place, or a dated plan inside the transition window.
- Article 4 records: who completed AI training, and when — national supervision starts this day.
- Questionnaire-ready language. A bare "yes, we comply" is a red flag. The four-part structure — classification, applicable obligations, status, gaps with a plan — is in our answer guide; the panic version of the question has its own playbook.
On penalties, quoted correctly: breaching Article 50 can draw fines of up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher (Article 99(4)(g)) — for SMEs and startups, the lower of the two (Article 99(6)). Fines are imposed by Member States; no enforcement action should be claimed without a documented case.
If a questionnaire citing the date has already landed on a live deal, start with the first-24-hours playbook.
FAQ
Did the Digital Omnibus postpone the 2 August 2026 date?
No. It delayed only the high-risk regime (Annex III to 2 December 2027; embedded in regulated products to 2 August 2028). Article 50 applies from 2 August 2026 as scheduled — the only accommodation is the 50(2) marking transition to 2 December 2026 for systems already on the market.
Do the prohibitions start on 2 August 2026?
No — the Article 5 prohibitions and Article 4 have applied since 2 February 2025, the governance and penalties chapters since 2 August 2025. Nothing about the prohibitions changes on this date.
Does my SaaS have to meet the high-risk requirements by 2 August 2026?
No. The Annex III high-risk obligations apply from 2 December 2027. Most B2B SaaS sits outside Annex III as limited or minimal risk — its duties from 2 August 2026 are the Article 50 transparency obligations, alongside Article 4, already in force.