Commentary. Artificial intelligence policy has reached the point where delay is no longer a neutral choice. The United States is debating how to govern increasingly capable systems while also treating technological leadership as a national-security imperative. Our view is that Washington should move quickly—but narrowly—toward enforceable safety rules, independent oversight and clear liability, rather than either freezing innovation or trusting companies to police themselves.
The immediate political argument is familiar. Supporters of lighter regulation say the United States cannot afford to slow development while China competes for technological influence. That concern is legitimate. AI could improve scientific research, productivity and public services, and badly designed rules could entrench the biggest firms by making compliance too expensive for smaller competitors. Industry also has a reasonable objection to one-size-fits-all mandates: a system used to draft marketing copy should not face the same requirements as one deployed in medical diagnosis or critical infrastructure.
But the case against indefinite voluntary oversight is stronger. A recent review of U.S. policy describes a regulatory landscape in which Congress has debated broad legislation without reaching a comprehensive settlement, while the White House has pursued a voluntary program for reviewing powerful frontier models. Voluntary commitments may be useful as a bridge, but they cannot substitute for duties that remain in force when commercial pressure rises.
The danger is not limited to cinematic scenarios about machines escaping human control. More immediate problems include fraud, privacy violations, discriminatory decisions, unsafe advice and the concentration of essential digital infrastructure in a handful of companies. These risks are uneven, but that is precisely why rules should focus on use and consequences rather than imposing identical obligations on every algorithm.
What a workable framework would require
First, developers of high-risk systems should document foreseeable harms, test models before deployment and report serious incidents afterward. Those obligations should be proportionate to a system’s capabilities and the stakes of its use. Second, independent auditors—not marketing departments—should be able to inspect evidence supporting safety claims. Bipartisan proposals in Congress have included independent verification organizations, suggesting that oversight need not be a partisan fantasy.
Third, companies should remain legally accountable for preventable harm. Transparency reports are valuable, but disclosure alone does not help a person denied a loan, impersonated by a synthetic voice or injured by an automated decision. Victims need accessible remedies, and regulators need authority to impose meaningful penalties.
“No one seems to agree exactly how AI should be overseen.”
That uncertainty, reported in recent coverage of the U.S. regulatory debate, is an argument for institutional clarity—not for paralysis. Congress should designate responsibilities among existing agencies, establish baseline rights and set a timetable for updating technical standards. States should not be casually stripped of their ability to respond to local harms, although businesses also deserve a predictable national floor rather than fifty contradictory regimes.
There are serious counterarguments. Critics warn that regulators may misunderstand fast-moving technology, that confidential trade secrets could leak through audits and that excessive caution could push research overseas. Those risks deserve safeguards: protected submissions, technically qualified oversight and periodic review. But uncertainty about the perfect rule cannot justify having no enforceable rule at all. Aviation and pharmaceuticals are not innovation-free sectors because they have safety requirements; they are trusted sectors because safety is part of the bargain.
The United States should therefore reject both extremes now competing for attention. A sweeping moratorium would sacrifice useful research and be difficult to enforce globally. A purely voluntary system would leave the public dependent on the goodwill of firms whose incentives reward speed, market share and investor confidence. The better course is targeted regulation that follows risk, preserves competition and gives independent institutions the power to verify claims.
AI policy will not be settled by one bill or one presidential order. It will be tested through ordinary failures: a misleading output, a security breach, a discriminatory system or a preventable death. The responsible response is to build accountability before those failures become the price of progress.