AI & strategy
AI Act Delayed: What Europe's Retreat Teaches Canada
· 8 min read

August 2, 2026 was supposed to be the day the European AI Act's obligations for high-risk artificial intelligence systems took effect: hiring, education, credit, critical infrastructure. On that day, they did not apply. On July 24, nine days before the deadline, the European Union published Regulation (EU) 2026/1744, known as the "Digital Omnibus on AI," in its Official Journal; it entered into force three days later, on July 27 (Regulation (EU) 2026/1744; Cooley, 2026). It pushes those obligations back to December 2027 and August 2028.
The European Parliament's co-rapporteur, Arba Kokalari, summed up the move in one phrase: Parliament is "pressing the pause button on the AI Act" (European Parliament, 2026b).
Is Europe, which had established itself as the global reference for AI regulation, giving up on its model? And what does this retreat mean for Canada, which still has no AI law?
What changed, and what did not
The core of the regulation is a change of calendar. Stand-alone high-risk systems listed in Annex III must now comply from December 2, 2027, and those embedded as safety components in regulated products from August 2, 2028 (European Parliament, 2026b). That is a 16-month delay for the former and a 12-month delay for the latter (Cooley, 2026).
Other relief comes with the delay. Systems already on the market before August 2, 2026 have until December 2, 2026 to mark AI-generated content in a machine-readable way (European Parliament, 2026b). The AI literacy obligation, which fell on providers and deployers, becomes a duty for the Commission and member states to support and facilitate AI literacy (Cooley, 2026). Exemptions for SMEs are extended to small mid-cap enterprises, and the deadline for national regulatory sandboxes moves to August 2, 2027 (Cooley, 2026). Finally, products covered by the Machinery Regulation are no longer directly subject to the AI Act and will follow that regulation's sectoral approach (Council of the European Union, 2026).
It would nonetheless be inaccurate to speak of a repeal. The regulation adds two prohibitions, applicable from December 2, 2026: systems that generate child sexual abuse material, and systems that create non-consensual intimate depictions of identifiable people, the so-called "nudifier" apps (European Parliament, 2026b). The obligation to register systems that their providers have themselves assessed as not high-risk was maintained, despite the proposal to remove it (Cooley, 2026). Above all, the AI Act's architecture remains intact: a risk-based approach, prohibitions in force since February 2025, and rules on general-purpose AI models in force since August 2025 (Heuking Kühn Lüer Wojtek, 2026).
The outcome is therefore a recalibration: a postponed calendar, targeted relief, and two new prohibitions that allow the co-legislators to present the whole as a balance.
Why Brussels backed down: competitiveness and missing standards
Two forces explain the delay. The first is political. Mario Draghi's report on European competitiveness, published in September 2024, established the idea that regulatory burden holds back innovation in Europe, and the Commission's wave of legislative simplification follows from it (European Parliamentary Research Service, 2025). A year later, Draghi himself described the AI Act as "a source of uncertainty" and called for a pause on the rules for high-risk systems while their drawbacks were assessed (Euronews, 2025).
The second force is technical, and it is the official argument. The harmonised standards and guidelines meant to help companies comply were not ready. Parliament justified the delay by the need to give those tools time to be finalised (European Parliament, 2026a). Requiring compliance with rules whose implementing standards did not yet exist would have created real legal uncertainty.
The speed of the procedure also says something. Proposed by the Commission on November 19, 2025, the text reached political agreement on May 7, 2026, a final vote in Parliament on June 16, adoption by the Council on June 29, and entry into force on July 27 (European Commission, 2026; Council of the European Union, 2026; Cooley, 2026). A regulation meant to postpone a deadline is useless if it arrives after it: the calendar was dictated by the date that had to be avoided.
A consensus that is wearing thin
The European Parliament's numbers tell a story the press releases leave out. In March 2026, Parliament's position was adopted by 569 votes to 45, with 23 abstentions (European Parliament, 2026a). In June, the final text received only 423 votes in favour, 57 against and 174 abstentions (European Parliament, 2026b). In three months, abstentions rose more than sevenfold. For a political scientist, that shift is a signal: many members did not want to block the compromise, but did not want to endorse it either.
Criticism came from both sides. A coalition of rights organisations, including EDRi, Access Now, AlgorithmWatch and Amnesty International, argues that the text weakens the AI Act and undermines the EU's credibility as a serious digital regulator (EDRi et al., 2026). The European consumer organisation BEUC recalls that simplification should make rules easier to apply, not serve as cover for deregulation (BEUC, 2026). Conversely, the tech industry association CCIA said the deal "falls short," notably because it keeps the registration requirement for systems deemed non-high-risk (The Record, 2026).
When rights advocates denounce a rollback and industry complains of timidity, the result is a classic compromise. The question is what it costs the credibility of the European model.
The Brussels effect put to the test
Legal scholar Anu Bradford coined the term "Brussels effect" for the EU's ability to export its rules through the sheer size of its market: to sell there, companies worldwide align with European standards, then apply them elsewhere for convenience (Bradford, 2020). The AI Act was designed in that spirit. It applies to providers placing systems on the European market wherever they are established, and to providers and deployers in third countries when the output of their systems is used in the Union (Regulation (EU) 2024/1689, art. 2).
The delay does not remove that extraterritorial reach. It does, however, change the message sent to the rest of the world. The EDRi-led coalition sees a precedent: implementation becomes a second legislative arena, where adopted rules can be renegotiated as soon as some actors complain (EDRi et al., 2026). A standard whose calendar can slip under competitive pressure is a less reliable anchor for countries that planned to moor themselves to it.
Canada, with no law and no anchor
Canada is precisely one of those countries. Its proposed Artificial Intelligence and Data Act (AIDA), part of Bill C-27, rested like the AI Act on a risk-based logic, targeting "high-impact" systems. It died on the Order Paper when Parliament was prorogued in January 2025 (Legal 500, 2026). The federal AI for All strategy, published on June 4, 2026, announces no new general AI law. It relies instead on targeted tools: standards, certification, procurement, sector-specific regulation and privacy reform (Fasken, 2026; Innovation, Science and Economic Development Canada, 2026).
Canada is not, however, a legal vacuum. In Quebec, Law 25 has required businesses since September 2023 to inform individuals when a decision about them is based exclusively on automated processing (Act respecting the protection of personal information in the private sector, s. 12.1). Quebec's public administration set rules in December 2025 for the use of generative AI by public bodies, with requirements on governance, risk management and staff training (Gouvernement du Québec, 2025). In Ottawa, the Directive on Automated Decision-Making governs systems used by federal institutions (Treasury Board of Canada Secretariat, 2019). In Ontario, a law passed in 2024 governs the use of AI systems by public sector entities (Strengthening Cyber Security and Building Trust in the Public Sector Act, 2024).
The European retreat lends itself to two readings. The first sees it as a vindication of Canadian caution: legislating before standards exist is costly, and Europe is now paying the price. The second sees it as a warning: without a law, Canada has no general lever, and the European anchor whose logic AIDA shared is moving. Both readings are well founded. What they have in common is decisive: the credibility of regulation depends less on its ambition than on the capacity to enforce it.
What Quebec organisations should take away
For a business or public body here, five practical conclusions follow.
- The delay is not an exemption. A Quebec company that sells an AI system on the European market, or whose system outputs are used there, remains subject to the AI Act (Regulation (EU) 2024/1689, art. 2). The prohibitions and the rules on general-purpose models already apply, and the high-risk obligations will arrive in December 2027.
- The delay is a runway, not a holiday. Technical documentation, risk management, human oversight: what will be required in 2027 must be prepared now, especially since the expected harmonised standards will serve as the reference.
- Law 25 already applies. The duty to inform individuals of fully automated decisions was never postponed (Act respecting the protection of personal information in the private sector, s. 12.1).
- The public sector has its own rules. Quebec public bodies must comply with the generative AI framework published in December 2025 (Gouvernement du Québec, 2025).
- Standards are becoming the common language. In the absence of a general Canadian law, standards, certification and procurement clauses will set expectations in practice (Fasken, 2026). Preparing for them means preparing for both the Canadian and the European markets.
Conclusion: regulating also means being able to enforce
Europe has not given up on regulating artificial intelligence. It has bought itself time, at the cost of a visibly weakened political consensus and a dented international credibility. Its mistake was not to legislate, but to set deadlines before having the standards and capacity to enforce them.
For Canada, the lesson is therefore neither "do not regulate" nor "copy Brussels." It is to build first what makes a law enforceable: recognised standards, certification mechanisms, clear rules for the public sector and well-equipped regulators. When Ottawa does legislate, the law will then be able to apply from day one, which the AI Act did not manage to do.
(The author)

Strategy, Digital and AI Consultant
Trained in political science (from an honours bachelor's to a master's at Université de Montréal), I'm now completing a graduate diploma (DESS) in private- and public-sector management at HEC Montréal.
Keep reading
The AI ROI Paradox: Why Adoption Isn't Showing Up in the Bottom Line Yet
Nearly nine in ten organisations use AI, yet its effect on financial results is flat. What the data say, and the decisions that set apart the few organisations that truly benefit.
November 13, 2026 · 7 min
Sovereign AI: What Quebec Really Controls
Agreements with Cohere and Mistral, the CLOUD Act, U.S. trade pressure: what Quebec's digital sovereignty in AI actually covers, and what is left to build.
November 4, 2026 · 10 min
Can Canada Really Break Free from Washington?
From the "rupture" declared at Davos to the Alliance for the Future signed with the European Union, Canada's middle-power doctrine tested against international relations theory and the reality of asymmetric interdependence.
November 14, 2026 · 10 min