Commentary. The United States is approaching a consequential choice on artificial intelligence: whether to build a credible national safety regime or continue treating public protection as an informal bargain between government and technology companies. Our view is straightforward: Washington should choose enforceable, focused rules—before the next crisis makes that choice for it.
The debate has acquired new urgency. Pentagon Chief Technology Officer Emil Michael said the federal government planned to detail the laws and regulations applying to domestic AI laboratories in November, framing the effort as a way to give Americans confidence without creating a patchwork of rules that could damage the industry. Meanwhile, OpenAI has called for mandatory national safety requirements, including testing, independent assessments, cybersecurity safeguards and incident reporting. [Reuters] [Reuters]
Those developments expose the weakness in the current approach. Companies that build the most powerful systems are being asked to police risks created by their own products, while the federal government has yet to enact a comprehensive AI law. In the meantime, states are filling the vacuum with differing requirements on high-risk systems, transparency and chatbot disclosures. [Reuters]
Voluntary commitments are not worthless. They can be adopted faster than legislation, updated as systems change and tailored by engineers who understand the technology. The industry also has legitimate concerns: rigid rules could burden small firms, discourage beneficial research and hand an advantage to competitors in China. The administration’s stated fear that hundreds of legislators could produce regulations capable of “killing the industry” reflects a real policy risk, not a caricature. [Reuters]
But flexibility is not the same as accountability. A promise to self-regulate is only as durable as the commercial incentives behind it. If safety testing delays a product launch, raises costs or reveals a serious defect, the public should not have to assume that every company will voluntarily choose caution over market share. Nor should rules depend on whether an individual executive remains committed after an election, a merger or a change in investor pressure.
A national floor, not a regulatory maze
The strongest case for federal legislation is not that every AI application requires the same treatment. It is that the most consequential systems need a common baseline. Congress should require documented pre-deployment testing for frontier models, independent evaluation where conflicts of interest are substantial, prompt reporting of serious incidents and meaningful cybersecurity controls. The requirements should scale with capability and risk, rather than placing identical obligations on a classroom tool and a system used in critical infrastructure.
That approach would also answer a reasonable objection from states and civil-rights advocates: national uniformity must not become national weakness. Federal rules should set a floor, while preserving state authority over consumer protection, employment, housing and other areas where local enforcement can identify harms more quickly. A national framework that pre-empts stronger safeguards would substitute one problem for another.
Recent politics suggests the public is not demanding a choice between innovation and safety. Reuters reported that more than 30 data-center measures appeared on ballots across 13 states in November and that almost all passed, with proposals limiting development in various ways. The result signals that communities want a voice over AI’s physical costs—energy use, water demand, land and infrastructure—even when they recognize the technology’s economic promise. [Reuters]
Congress should read that signal carefully. Public consent cannot be manufactured through optimistic speeches or voluntary industry accords. It is earned when people know who is responsible, what protections apply and what happens after a system fails.
America does not need a rulebook for every algorithm. It does need enforceable boundaries for the systems capable of causing national-scale harm. Sensible AI regulation would not reject the technology race; it would make that race politically and socially sustainable. The question is no longer whether government should act. It is whether lawmakers will establish clear safeguards before public trust runs out.