Commentary. Europe’s artificial-intelligence rulebook has reached the moment when political promises meet operational reality. Since 2 August 2026, the European Commission and national authorities have begun enforcing major provisions of the EU AI Act, including transparency duties for chatbots and labels for deepfakes. The law’s central ambition is sound: permit useful AI while placing guardrails around systems that can affect safety, rights and democratic life.

Yet the implementation has already been softened. The EU’s Digital Omnibus, which entered into force in July, extends deadlines, simplifies documentation for smaller companies and expands regulatory sandboxes. The European Parliament’s legislative summary says the changes are intended to reduce duplication and compliance burdens while preserving fundamental-rights protections. That is a reasonable objective. Our concern is whether the political pressure for competitiveness will gradually turn simplification into deregulation.

A necessary correction

The case for adjustment is not difficult to make. Small firms cannot absorb the same legal and technical workload as global technology companies. Unclear standards and overlapping obligations can delay beneficial products, discourage investment and push entrepreneurs toward jurisdictions with more predictable rules. The Omnibus responds by extending some simplified requirements to small mid-cap companies and allowing more proportionate technical documentation. It also expands testing environments where companies can develop systems under regulatory supervision.

Those measures deserve support. Regulation that only lawyers at the largest firms can navigate is not genuinely pro-innovation or democratic. A start-up building a medical-support tool should spend its resources improving reliability, not deciphering duplicated paperwork. The EU is right to seek clearer deadlines and a more workable path to compliance.

The risk of enforcement by postponement

But the counterargument is stronger than a slogan about Europe falling behind. AI systems are already being used to rank applicants, assess risk, generate public-facing information and shape what people see online. Delayed obligations can mean delayed accountability, especially where errors fall disproportionately on people with little power to challenge automated decisions.

The European Parliament’s implementation timeline sets key high-risk obligations for later dates, with the AI Act not fully effective until 2028. That staggered approach may be administratively necessary, but it also creates a long period in which companies and public bodies can deploy systems before the strongest safeguards apply. The answer cannot be to pretend that every delay is harmless.

Transparency rules are a useful test. From August, certain interactive systems must disclose that users are dealing with AI, while synthetic or altered material must be labelled. These requirements will not solve manipulation or misinformation by themselves. Labels can be ignored, stripped away or rendered meaningless by confusing design. Still, a visible duty to disclose is better than treating deception as an unavoidable cost of innovation.

What Europe should defend

Our position is not that every AI application needs maximal regulation. It is that the burden should track the consequence. A spellchecker and an employment-screening system should not face identical scrutiny. But where software can influence access to work, education, credit, healthcare or public services, enforceable rights must come before speed.

That requires three priorities. First, national authorities need money, staff and technical expertise, not merely statutory powers. Second, regulators should publish plain-language guidance so companies can comply before violations occur. Third, enforcement must be visible enough to deter misconduct, including by large firms whose resources allow them to treat fines as a business expense.

Industry groups are right to warn that fragmented national enforcement could undermine the single market. Civil-society groups are right to warn that vague exemptions and postponed deadlines could weaken protections in practice. Both concerns can be true. The solution is a common enforcement baseline, transparent decisions and a presumption that affected people deserve an explanation and a meaningful route to appeal.

Europe should not apologize for regulating powerful technology. Nor should it confuse a thick rulebook with effective protection. The AI Act’s credibility will be decided less by the elegance of its text than by whether ordinary people can tell when an automated system affects them—and do something when it gets things wrong.

Sources