Commentary. The United States is approaching an avoidable policy failure on artificial intelligence. Congress is leaving Washington for the midterm campaign without agreeing on basic rules for systems that can affect consumers, workplaces, public services and national security. The result is not neutrality. It is a decision to let the fastest-moving companies define the safeguards.
Recent developments make that choice harder to defend. A bipartisan-looking effort to establish a Commerce Department safety board, require incident reporting and create pre-deployment standards was blocked in the Senate.[Politico] Separately, Congress is reportedly leaving without adopting limits on AI deployment or even measures addressing the rising electricity costs associated with data-center expansion.[The New York Times]
Our view is straightforward: AI governance should begin with enforceable, narrow safety duties, not with either complacency or maximalist prohibition. Companies should have to document serious risks, report major failures, permit independent testing and demonstrate that high-risk systems meet defined standards before release. Those requirements would not settle every philosophical question about AI. They would, however, establish accountability where the consequences are tangible.
The case for action
The strongest argument for regulation is practical rather than speculative. AI systems already influence decisions and generate content at scale, while their developers possess far more technical information than users, journalists, regulators or elected officials. When a system fails, the public often learns about the failure only after harm has occurred. A reporting regime and external audits would not eliminate mistakes, but they could make recurring failures visible and create incentives to correct them.
Even lawmakers who oppose mandatory rules have acknowledged the need for oversight and transparency. House Speaker Mike Johnson has expressed support for external auditing, while preferring that leading firms adopt such measures voluntarily.[Politico] That preference deserves consideration: voluntary standards can be faster, more flexible and less likely to freeze useful innovation in a rapidly changing field.
But voluntary commitments have a structural weakness. They ask companies to bear costs that competitors can avoid. A firm that spends heavily on testing, security and red-teaming may be placed at a disadvantage against one that makes broader claims with fewer checks. Public rules can create a floor beneath which no serious company is allowed to fall.
The counterargument matters
Critics are right to warn against vague mandates and bureaucratic overreach. A government board that must approve every model could delay medical, scientific and educational applications. Poorly designed rules might favor the largest firms, which can afford compliance teams, while burdening smaller laboratories and open-source developers. They could also turn contested predictions about future dangers into legal standards that quickly become obsolete.
Those risks argue for proportionate regulation, not legislative paralysis. Rules should focus on measurable outcomes and clearly defined high-risk uses, with exemptions or lighter obligations for low-risk experimentation. Regulators should publish technical criteria, permit appeals and review standards regularly. Congress should also address energy and infrastructure impacts transparently, rather than allowing communities to discover that data-center growth has shifted costs onto household power bills.
Why blanket bans are the wrong answer
A proposal backed by several Democratic lawmakers would ban “superintelligence” and pause advanced AI development until a vetting regime exists.[Politico] The motivation—concern about systems that could evade safeguards or create catastrophic risks—is serious. Yet a broad pause would raise difficult questions about definitions, enforcement and international competition. It could push development into less transparent jurisdictions without guaranteeing that dangerous research stops.
The better course is democratic control through specific obligations: disclose capabilities and limitations, test for predictable harms, secure sensitive systems, report serious incidents and impose penalties when firms mislead the public or regulators. The United States should welcome innovation, but it should no longer confuse speed with leadership.
Congress still has time to choose responsibility over spectacle. The first law need not be perfect. It must be real, enforceable and strong enough to ensure that public safety is not merely another voluntary feature.