Commentary. Washington’s latest argument over artificial intelligence is arriving at the wrong destination. After the U.S. midterm elections, calls for a pause on advanced AI and even a ban on “artificial superintelligence” have collided with industry warnings, voluntary safety pledges and a White House preference for keeping regulation light. The result is a familiar political spectacle: maximalist language on one side, confidence in corporate self-policing on the other.
Our view is simpler. The United States should regulate AI, but it should regulate concrete dangers rather than hypothetical headlines. That means enforceable duties for developers of the most powerful systems, independent testing, disclosure of serious failures and clear accountability when companies deploy tools they know are unsafe.
That position is not a rejection of innovation. It is a demand that innovation earn public trust.
The case for restraint
The strongest argument for tougher rules comes from the companies building these systems themselves. Anthropic chief executive Dario Amodei has called for “more serious and binding regulation,” while other leading developers have supported stronger safety measures. Their warnings deserve attention—not because executives should write the law, but because firms closest to the technology understand how quickly capabilities can change.
Congressional concern is also broadening. Representative Yassamin Ansari has called the industry “out of control” and urged a pause on advanced AI development, while Senator Bernie Sanders has proposed far-reaching measures addressing data centers, corporate ownership and the distribution of AI’s economic gains. These proposals reflect legitimate anxieties about labor displacement, concentration of power, energy use, misinformation and systems that could be deployed before their risks are understood.
A voluntary agreement announced by President Donald Trump with six technology companies offers a useful starting point but not a finished policy. Internal safeguards, monitoring and independent assessments may improve practice; they do not substitute for a legal baseline. Companies can revise commitments, interpret them narrowly or abandon them when commercial pressure rises.
The case against a blanket pause
The opposing argument is not merely a defense of Silicon Valley. A broad pause could freeze beneficial research while leaving less transparent actors outside the agreement. It could also drive development overseas, weaken American scientific leadership and delay applications in medicine, accessibility, education and public services.
There is a practical problem, too: “advanced AI” is an imprecise category. A rule aimed at frontier models might inadvertently capture smaller systems used for ordinary business, research or safety testing. Legislators who promise to halt “superintelligence” may be responding to a genuine fear, but fear is not a workable regulatory definition.
Nor should every alarming prediction become law. Policymakers must distinguish demonstrable harms—fraud, discriminatory decisions, privacy violations and unsafe products—from speculative scenarios. Regulation that is vague, technologically rigid or dominated by incumbent firms could protect the largest companies while burdening researchers and smaller competitors.
A workable middle path
The answer is a risk-based framework with teeth. The most capable developers should register systems above clearly defined thresholds, conduct pre-deployment evaluations and report major incidents to an independent public body. Companies should retain records that allow auditors and courts to determine what a model did, what safeguards were used and who approved deployment.
Rules should also follow the use of AI, not just the size of the model. A system assisting a doctor, hiring a worker or determining access to credit deserves stricter obligations than a tool generating a restaurant menu. People affected by consequential automated decisions need notice, an explanation and a meaningful way to appeal.
Finally, Congress should fund public-interest testing and technical expertise. Government cannot oversee systems it cannot evaluate, and it should not rely exclusively on companies’ own benchmarks. Independent researchers need access, privacy protections and protection from retaliation when they uncover failures.
There will be costs. Compliance may slow releases, raise prices and make some investment less attractive. Those are real trade-offs. But the alternative—allowing commercial urgency to determine the safety standard—merely shifts the cost to workers, consumers and the public.
The AI debate should therefore reject both complacency and theatrical prohibition. America does not need rules that stop every experiment. It needs rules that make powerful companies prove their systems are safe enough to use, disclose when they are not and pay for preventable harm. That is not anti-technology. It is the minimum condition for technology to remain worthy of public confidence.