Commentary. Artificial intelligence has reached the point where governments can no longer treat safety as a voluntary promise. But the answer is not to slam the brakes on every powerful model, or to write rules around science-fiction fears while neglecting the harms already appearing in workplaces, markets and public infrastructure.
That is the position this newsroom takes as Congress reopens the argument over AI regulation. The debate has produced an unusual coalition: safety advocates, some technology executives and lawmakers from both parties increasingly agree that frontier systems need independent scrutiny. At the same time, industry voices and national-security hawks warn that excessive regulation could slow innovation and weaken the United States in competition with China. Both concerns deserve to be taken seriously. Neither justifies inaction.
The case for rules
Congressional proposals reportedly include independent auditors for large AI developers, risk assessments before new models are released and continuing transparency reports after deployment. One bipartisan proposal would require major developers to retain government-licensed verification organizations to monitor model risks.[1] Those are not radical demands. They resemble the basic logic applied in aviation, pharmaceuticals and financial reporting: companies can innovate, but they do not get to mark their own homework indefinitely.
There is also a more immediate question of who pays for AI’s physical footprint. A proposed Ratepayer Protection Act would encourage states to require data centers consuming at least 100 megawatts to cover the generation, transmission and related infrastructure they require, rather than passing those costs to ordinary utility customers.[1] That principle should command broad support. Communities should not subsidize private computing campuses through higher bills while receiving vague promises of future jobs.
Transparency rules could also make enforcement more credible. Audits, incident reporting and clear responsibility for unsafe deployments would not eliminate bad outcomes, but they would give regulators and the public a way to identify recurring failures. A system that can affect elections, employment, health information or critical infrastructure should not operate behind a wall of corporate secrecy.
The case against overreach
The opposing argument is not simply a demand for corporate freedom. AI development is moving quickly, and poorly designed rules could entrench today’s largest companies by making compliance too expensive for smaller competitors. A licensing regime written around hypothetical superintelligence could also divert attention from ordinary fraud, discrimination, privacy violations and labor displacement.
Some technology leaders and politicians argue that the United States cannot afford a slowdown while China competes for technological and strategic advantage.[2] That warning has force. Advanced computing may improve scientific research, productivity and national defense. Policymakers should not assume that every risk can be solved through a sweeping federal statute, especially when the technology changes faster than legislative cycles.
Nor should lawmakers turn every alarming demonstration into proof of an imminent catastrophe. Calls to permanently ban the development of artificial super-intelligence, reportedly being prepared by Sen. Bernie Sanders and Rep. Greg Casar, would settle a speculative question before society has established workable definitions, enforcement mechanisms or international agreement.[1]
A narrower, stronger settlement
The sensible path is targeted regulation with periodic review. Congress should require independent evaluations for the most capable models, disclosure of serious incidents, meaningful protections for personal data and liability when companies deploy systems they knew—or should have known—were unsafe. It should also prevent energy and infrastructure costs from being shifted onto households.
Rules should focus on measurable capabilities and demonstrable harms, not on branding systems as “safe” or “unsafe.” They should preserve room for open research, protect whistleblowers and give smaller firms a route to compliance. States should retain authority to address local impacts unless a federal standard genuinely provides stronger protection.
Above all, regulation should be designed to learn. Independent auditors need access to systems and records; regulators need technical expertise; lawmakers need sunset clauses and public reporting. A static law will fail if it cannot adapt.
AI does not need either blind acceleration or theatrical prohibition. It needs the boring disciplines that made other powerful industries more accountable: testing, disclosure, responsibility and a clear answer to the question of who bears the cost when things go wrong. That is not an attack on innovation. It is the price of earning public trust.