Commentary. The argument over artificial-intelligence regulation is often presented as a choice between innovation and safety. That is the wrong choice. The real question is whether governments can impose focused, enforceable duties before unreliable systems become embedded in decisions about work, education, housing and public services.
Recent developments offer a warning against both complacency and overreach. The European Union has amended its AI framework through a “Digital Omnibus,” delaying some high-risk obligations to December 2027 and August 2028 while retaining rules on prohibited practices and transparency. The United States, by contrast, still lacks a comprehensive federal AI law, leaving states to develop a patchwork of measures. The result is neither certainty for business nor dependable protection for the public.
Our view is straightforward: regulators should start with demonstrable harms, require accountability from institutions that deploy AI, and give companies workable technical standards. They should not attempt to regulate every speculative future capability or prescribe a single model of innovation.
The case for restraint
Industry has legitimate concerns. Artificial intelligence develops quickly, and poorly drafted rules can become obsolete before companies understand how to comply. Small firms may lack the legal and engineering teams available to large technology companies. Overlapping state laws can raise costs, discourage investment and make it harder for organizations to offer the same service nationwide.
Those concerns help explain why the EU’s revised timetable has drawn support from businesses seeking more time to interpret requirements. A risk-based system also makes more sense than treating a spellchecker, a medical device and a frontier model as equivalent technologies. Regulation should follow the consequences of a system, not the novelty of its branding.
But delay is not neutral. It shifts risk to people who may have no practical way to challenge an automated decision. A worker rejected by a screening system, a student flagged by software or a tenant evaluated by an algorithm may never know what data shaped the result. “Move fast and fix it later” is a poor philosophy when the people harmed cannot easily reverse the outcome.
What sensible rules would require
First, organizations using AI in consequential decisions should provide notice, preserve records and offer meaningful human review. Human involvement must mean more than a nominal sign-off by someone who routinely accepts the machine’s recommendation. A person should have the authority, time and information needed to question it.
Second, developers and deployers should test systems for accuracy, bias, security and foreseeable misuse. Independent auditing can help, but audits should not become decorative compliance exercises. Regulators need access to documentation and the power to investigate incidents, while companies need clear thresholds that distinguish serious failures from ordinary model errors.
Third, transparency rules should focus on the user’s actual need. People should know when they are interacting with an AI system or when synthetic media materially affects what they see. Yet disclosure alone cannot cure a discriminatory hiring tool or a dangerous medical recommendation. Transparency is a baseline, not a substitute for accountability.
The United States’ fragmented approach makes these principles more urgent. Analysts tracking 2026 legislation report activity on automated decisions, frontier-model risks, third-party auditing and public-sector use, but no comprehensive federal measure had passed in the current session. State experimentation can reveal what works, but a patchwork should be a bridge to national standards, not the permanent arrangement.
The balance lawmakers should strike
Critics will say that regulation could weaken the West’s position in competition with China or push development to less regulated jurisdictions. That risk deserves consideration. But safety rules do not automatically suppress innovation; predictable rules can reward companies that invest in reliability and prevent reckless operators from gaining an artificial advantage.
The answer is targeted regulation: strict duties for high-impact uses, lighter obligations for low-risk tools, safe harbors for good-faith testing, and regular review as evidence changes. Governments should also fund technical capacity so regulators are not dependent on the companies they oversee.
AI will not be governed well through panic, nor through faith that market competition will repair every failure. The EU’s adjustments and America’s state-level experimentation should prompt a more practical course: preserve room to build, but make the builders and deployers answerable when their systems affect people’s rights and livelihoods.