The EU AI Act No Longer Has One Effective Date
Regulation (EU) 2026/1744, the Digital Omnibus on AI, is the subject of this series. Its central change is easy to summarize and easy to misunderstand: it does not postpone the EU AI Act as a whole, but reorganizes when different rules apply to different kinds of systems and actors. This opening explains why the phrase “the AI Act’s effective date” is now too blunt, and gives you the mental model needed to decide which date matters in a real situation.
The change matters because a date is not merely a calendar entry for a compliance department. It can determine whether a provider must redesign a system, whether a public authority can continue using an existing one, whether a product needs a particular conformity assessment, or whether a person receives protection against a harmful practice. The same organization may have to respond to one AI rule in 2026, another in 2027, and another in 2028. Treating those as one event creates two opposite errors: preparing too late for rules that already apply, or duplicating work because rules that could be coordinated are treated as separate regimes.
The Digital Omnibus is the enacted change, not just an announcement
The relevant subject is not simply a Commission statement about simplifying AI regulation. It is Regulation (EU) 2026/1744, adopted on 8 July 2026 and published in the Official Journal on 24 July. It amends the Artificial Intelligence Act, Regulation (EU) 2024/1689, and coordinates related provisions in the EU rules for civil aviation and machinery. The Commission had announced a political agreement on 7 May. That announcement was useful as an early explanation of the proposed direction, but it still described formal adoption and publication as next steps. The enacted regulation is the authority for what now applies.
That distinction is important in fast-moving regulation. A press release can tell readers what policymakers say they have agreed to do. It cannot, by itself, settle the wording of an exception, define which actor is responsible, or establish the date on which a provision applies. The final text does those things. Here, it says that the amending regulation enters into force on the third day after publication. Since it was published on 24 July, the regulation itself took effect on 27 July 2026.
That fact does not mean every amendment became operative on 27 July. “Entry into force” answers when the amending legal instrument becomes part of the law. “Entry into application” answers when a particular rule begins governing the conduct or system within its scope. Those events can coincide, but they do not have to. The Digital Omnibus makes the difference especially visible because it changes the AI Act’s application schedule while adding rules that have their own dates.
The reason for the change is also more specific than “the EU changed its mind.” The regulation’s recitals point to delayed harmonised standards, guidance, national authorities, governance arrangements, and conformity-assessment frameworks. Without those pieces, the original high-risk timetable risked producing heavier-than-expected compliance costs, inconsistent interpretations, or rules that formally started before operators had reliable ways to demonstrate compliance. The stated response is targeted simplification and more time for implementation, while maintaining the AI Act’s objectives around safety, health, fundamental rights, the internal market, and innovation.
That is a policy trade-off, not proof that the protection level has fallen. More preparation time can reduce rushed or duplicated work. It also makes the eventual result depend more heavily on whether standards, authorities, assessment bodies, guidance, and testing arrangements are ready when needed. The law establishes some of those workstreams and deadlines, but the enacted text does not prove that every practical support instrument is already available.
Entry into force is not the same as application
The easiest way to see the new structure is to stop looking for one replacement date. The amended AI Act contains a set of legal clocks. Some are already running; some concern a new prohibition; some concern institutional powers; and the two headline high-risk dates belong to different classification routes.
The regulation keeps Chapters I and II of the AI Act applicable from 2 February 2025, while setting the newly added prohibited-practice provisions on non-consensual intimate material and child sexual abuse material to apply from 2 December 2026. Articles 102 to 110, which concern parts of the institutional framework, apply from 27 July 2026. The main high-risk obligations in Chapter III, Sections 1 to 3, are split: the rules for systems classified under Article 6(2) and Annex III apply from 2 December 2027; the rules for systems classified under Article 6(1) and Annex I apply from 2 August 2028. The text also contains transition rules for certain synthetic-content systems and for high-risk systems intended for public authorities, including a 2030 compliance date for the latter condition.
This is not a list to memorize without context. It is a warning about the unit to which a date attaches. The 2 December 2027 date is not “the date for AI.” It is a date for a specified group of high-risk systems and a specified set of provisions. The 2 August 2028 date is not a general extension for every provider. It belongs to another route into high-risk classification. The 2 December 2026 date is not the start of the entire AI Act; it is the application date for newly inserted prohibitions, alongside a related transition for certain transparency duties.
The difference changes how a reader should interpret a headline. If someone says that the AI Act has been delayed until 2028, the useful response is: which provision, and which system? If someone says that the AI Act is already in force, the next question is: are they talking about the amending regulation, an existing AI Act chapter, a governance provision, a prohibited practice, or a high-risk requirement? Both statements may contain a fragment of truth while still giving the wrong operational answer.
The structure can be represented as a path rather than a single switch:
The path also explains why the calendar cannot be separated from the rest of the regulation. A provider placing a new system on the market is not in the same position as a deployer using an existing system. A system built on a general-purpose model may raise a different supervisory question from an AI component in a lift or toy. A system intended for a public authority may have a different transition from an otherwise similar private-sector deployment. The date is the result of those facts, not a substitute for them.
The date belongs to a rule, a system, and an actor
Three distinctions will organize the rest of this series. The first is the distinction between a legal event and a compliance event: the amending regulation can enter into force before a revised obligation applies. The second is the distinction between types of high-risk classification. The third is the distinction between the people or organizations connected to the system.
Under the revised structure, one high-risk route runs through Article 6(2) and Annex III, which lists use areas such as biometrics, critical infrastructure, education, employment, migration, asylum, and border control. The other runs through Article 6(1) and Annex I, where an AI system can be a safety component of a product governed by Union product legislation. The Commission’s earlier public description used lifts and toys as product examples. These are not simply two industry lists. They are different legal relationships: one turns on a listed use case, while the other turns on a product and a safety function.
The actor matters as well. The Digital Omnibus adds a prohibition that tests a provider’s purpose and foreseeable, reproducible system capability differently from a deployer’s purposeful use. It gives the AI Office a defined Union-level role for some systems built on general-purpose AI models and for systems connected to very large online platforms or search engines, while retaining national and sectoral responsibilities in specified cases. It gives smaller enterprises proportional tools, but with eligibility conditions and without a general reduction in the required level of protection.
Finally, lifecycle matters. The amended rules clarify when an existing type and model can continue under a grace period and when a significant design change brings the relevant high-risk obligations into play. A system’s first lawful placement on the market, later modifications, intended use, and relationship to a public authority can all change the answer. The question is not only “what is this system?” but also “what happened to it, and who is using it?”
This is why the title’s claim is more than a semantic correction. The AI Act now has a layered implementation calendar because the Digital Omnibus links dates to legal provisions, classification routes, institutional roles, protections, and system history. The next task is to make that calendar concrete: to identify which clocks are already running, which begin in 2026, 2027, 2028, or 2030, and what each one actually governs.