Europe wrote the world’s first comprehensive artificial intelligence law. Passing it, however, may have been the easy part.
The European Union Artificial Intelligence Act entered into force in 2024, and was widely heralded as a global benchmark. Implementation was always meant to be phased. Now, before the law is even fully applied, Europe has postponed some of its hardest obligations.
In May 2026, EU lawmakers agreed to delay major compliance deadlines. Standalone high-risk AI systems, originally scheduled for regulation from August 2026, will now have until December 2027 to comply, while high-risk AI embedded in regulated products will have until August 2028. The reason is pragmatic. The technical standards and enforcement tools needed to make those rules workable were not ready.
This is not a retreat from regulation as other key provisions remain on schedule. Transparency obligations, such as labelling AI-generated content, still apply from August 2026, subject only to a short grace period for existing systems. Obligations for general-purpose, or “foundation”, AI models began in 2025 as planned.
Europe has not abandoned its rulebook, and it has acknowledged that building the institutional and technical capacity to enforce such a sweeping AI law takes time. In effect, the world’s largest AI rulebook is already being reshaped by implementation reality.
The lesson is that writing rules was only the first phase of AI governance. The second phase is making them work on the ground.
The EU AI Act was widely expected to play a role similar to Europe’s General Data Protection Regulation, which became a durable global reference point for data protection. AI governance is proving harder. Technology evolves faster than laws can be updated, and the methods needed to audit, test and supervise AI systems are still being built.
The EU AI Act will shape global debates, but whether it becomes a stable default model is now less certain. In AI, unlike data protection, constant recalibration may be unavoidable. Europe’s own revisions suggest that flexibility is not a design flaw but a necessity, especially as foundation models blur boundaries between sectors and use cases.
Policy Drift
For those familiar with machine learning, the need for constant recalibration is intuitive. A deployed model cannot be “fire and forget”. Its accuracy degrades as the real world changes, a phenomenon known as data drift. The standard practice is to build continuous monitoring and retraining into the system’s design from the start.
Similarly, writing an AI law is not a one-time exercise. It needs built-in feedback loops, the policy equivalent of model monitoring, to stay effective as technology evolves. Europe’s experience, essentially a late retraining of its AI Act, shows why smart governance should be designed for adaptation from the outset.
As Indian policymakers consider the next steps in AI governance, Europe’s experience underscores a simple point: implementation capacity matters as much as legislative ambition.
India’s AI Governance Guidelines, released under the National AI Mission in early 2026, take a principle-based and institution-led approach rather than rushing into a comprehensive statute. The guidelines explicitly argue that effective governance includes not just regulation but also capacity building and institution building, and that many AI-related risks can, at least initially, be addressed through existing laws and targeted guidance.
The guidelines even call for new bodies, such as an AI Governance Group and an AI Safety Institute, to start building oversight muscle before any binding AI law is in place.
This sequencing is worth contrasting with Europe’s experience.
The EU passed its AI Act first and only then had to stand up and empower new institutions, including an EU AI Office, while coordinating dozens of national regulators. That institutional lag is one reason implementation timelines slipped. India’s approach suggests an attempt to have the guardrails ready when the rules eventually arrive.
None of this guarantees success as institutions on paper do not automatically translate into effective enforcement, and principle-based frameworks can drift if not followed by operational detail. Sceptics will argue that India’s stance reflects caution or delay rather than strategy. That is a fair concern.
But the alternative is visible in Europe’s experience: legislate early and discover later that the system cannot yet support compliance. In a country as large, diverse and administratively complex as India, the cost of rewriting a premature law could be higher than the cost of waiting to legislate once the ground is better prepared.
Execution Test
Singapore offers a brief contrast, as instead of adopting a single, comprehensive AI statute, the country has relied on detailed guidance and an evolving governance playbook, updated through engagement with industry and researchers. The emphasis is on operational clarity and iterative adjustment.
In fast-moving technological domains, that approach can sometimes travel better than dense legislation. Singapore’s experience reinforces the broader lesson from Europe: effective AI governance depends less on the elegance of the statute and more on the machinery behind it.
No jurisdiction has fully cracked AI governance yet. Europe at least deserves credit for forcing the issue onto the global agenda and setting a high bar for accountability. Its willingness to revise timelines now reflects realism, not retreat. The harder work lies in turning legal ambition into administrative practice.
For India, the takeaway is not that it should legislate less, though a measured approach may well make sense, but that it should never undervalue execution. India has not rushed into an AI law, a stance that may reflect strategic patience or may simply reflect caution.
If this period is used to strengthen oversight tools and coordination, through the proposed AI Governance Group, AI Safety Institute and other capacity-building measures, India could eventually craft rules that endure with minimal rewriting.
In the next phase of global AI governance, prestige will not come from writing the thickest rulebook. It will come from writing one that does not need to be rewritten every time reality catches up.
.png)
