The AI governance vacuum and why Congress cannot fill it

The United States finds itself in an unusual position on artificial intelligence: its president dismisses safety concerns as a “hoax,” its leading AI companies are calling for mandatory oversight their own executives helped defeat, and Congress is deadlocked on any safety legislation with no viable path to passage. Anthropic co-founder Jack Clark has told the BBC that an AI “kill switch” may need to be mandatory, even as Nvidia’s Jensen Huang insists the industry “doesn’t need new laws.” AI-linked stocks slid after tech executives publicly called for a slowdown in what they described as “reckless” development. The US House Speaker has meanwhile warned against any pause, arguing that slowing down would cede ground to China. The result is a governance landscape that satisfies no serious position.

The received wisdom

The dominant progressive case for AI regulation is powerful: these are systems of unprecedented capability being deployed at unprecedented speed by companies with financial incentives that do not align with public safety. The history of technology regulation — from pharmaceutical standards to aviation safety to nuclear power — shows that market participants cannot self-regulate adequately when catastrophic downside risks are involved. The Anthropic co-founder’s call for a kill switch is not alarmism; it reflects genuine uncertainty at the frontier of a technology that its own architects do not fully understand. If we wait until something goes catastrophically wrong before building governance structures, we will have waited too long.

The UK government’s national security minister has made essentially this argument, warning that AI “poses huge risks to national security without guardrails.” It is a reasonable position. The precautionary principle is not always wrong. There are cases where the cost of waiting for evidence of harm is too high.

A different read

And yet. The specific proposals currently in circulation, and the political coalition pushing them, deserve considerably more scrutiny than the urgency of the framing tends to invite.

The problem with the AI safety consensus as it has developed in Washington and Brussels is that it conflates three quite distinct categories of concern. The first is near-term, concrete harm: deepfakes used for fraud, autonomous weapons deployed without adequate human oversight, algorithmic bias in consequential decisions. These are real, addressable with targeted regulation, and do not require a sweeping framework that effectively hands regulatory power to a small technocratic class. The second is medium-term economic disruption: job displacement, market concentration, competitive dynamics. These are genuine policy challenges but are not safety issues in any meaningful sense; they are industrial policy questions. The third is existential risk from superintelligent systems: the scenario that motivates calls for a “kill switch.” This is speculative, contested among serious researchers, and being deployed rhetorically to create urgency that benefits the largest incumbents who have already built moats.

There is a well-established pattern in American regulatory history where large incumbents support regulatory frameworks specifically because those frameworks raise compliance costs that crush smaller competitors. The Sherman Antitrust Act’s early history is instructive; so is the development of financial regulation after 2008, which produced a regime that entrenched the major banks’ dominance. When OpenAI and Anthropic call for mandatory safety standards that only they can meet, it is worth asking who benefits from the resulting regulatory architecture.

The China argument cuts both ways. The House Speaker warns that slowing down means China wins. This is directionally true but does not settle the regulatory question: China is also investing heavily in AI governance frameworks of its own, though for very different reasons (control of information and population management rather than genuine safety). The US competitive disadvantage from thoughtful, targeted regulation is far smaller than the competitive disadvantage from a catastrophic failure of an unregulated system that destroys public trust in AI entirely. The question is not regulation versus no regulation; it is whether we write smart rules now or bad rules in a panic later.

The AI regulation deadlock in Congress reflects a failure not of political will but of institutional capacity. Congress does not have the technical staff to evaluate competing claims about frontier AI development. The Federal Trade Commission, the NIST, and the various executive agencies that have jurisdiction lack clear authority and are themselves subject to the same political winds that have produced the current stalemate. What is needed is not a kill switch controlled by a single actor, but a genuine, adequately funded technical regulatory body — something like the Nuclear Regulatory Commission’s model — with statutory independence and genuine expertise. That kind of institution takes years to build and requires bipartisan consensus that does not currently exist.

In the meantime, the honest answer is that the US is governing AI about as well as it governed the early internet — which is to say, poorly — and will likely muddle through to a workable framework only after something goes badly enough wrong to force political action. That is not a satisfying conclusion. It is, however, a realistic one.

What to watch

  • Whether any bipartisan AI safety legislation can attach to a must-pass budget vehicle in the current Congress — budget riders are historically how contentious regulatory frameworks get enacted.
  • The EU AI Act’s enforcement trajectory: European jurisdictions will produce the first real test cases for AI liability, and US courts will be watching.
  • Nvidia’s Jensen Huang and his “no new laws” stance: shareholder pressure and insurance market dynamics may prove more powerful than legislation in the near term.
  • Any concrete incident — a major AI-enabled fraud, a military miscalculation, a healthcare system failure — that shifts the political calculus overnight.

— J