「AIキルスイッチ法」でトランプ政権が危険なAIシステムの強制停止を命令可能にAI Kill Switch Act would let Trump admin order shutdown of rogue AI systems

A proposed piece of US legislation informally known as the AI Kill Switch Act would, according to reporting, grant the Trump administration the authority to order the shutdown of artificial intelligence systems it deems dangerous or "rogue." The measure matters because it would place a direct enforcement mechanism in the hands of the federal executive branch at a moment when lawmakers, companies, and safety researchers are still debating who should hold ultimate control over powerful AI models and how such control could realistically be exercised.

At its core, the bill appears designed to create a legal pathway for the government to compel operators to halt the operation of specific AI systems. That is a notable shift from most existing US policy discussions, which have focused on disclosure requirements, safety testing, and voluntary commitments rather than emergency intervention. If enacted as described, the authority would likely apply to systems that regulators believe pose an imminent threat, though the precise definitions of "dangerous" and "rogue," the evidentiary standards required, and the appeals process available to affected companies will be critical details that determine how the law functions in practice.

The concept of an AI "kill switch" has circulated in safety circles for years, but it is technically more complicated than the phrase suggests. Large AI models are typically distributed across data centers, replicated, and in some cases released as open-weights that anyone can download and run locally. Ordering a company like a major cloud or model provider to stop serving a system is feasible, but preventing copies that already exist in the wild from running is far harder. As a result, a statutory shutdown power would probably be most effective against centralized, commercially operated services and considerably less effective against models that have already been widely distributed.

The proposal arrives amid an unsettled regulatory landscape. The Biden administration had issued a broad executive order on AI safety in 2023 that leaned on the Defense Production Act to require reporting from developers of the most capable models; the Trump administration subsequently moved to roll back much of that framework in favor of a lighter-touch, innovation-focused posture. Against that backdrop, a bill that hands the same administration a strong intervention tool represents an apparent tension between deregulatory rhetoric and a desire to retain emergency authority, and it is likely to draw scrutiny from both industry groups wary of overreach and safety advocates who question how the power would be applied.

There is also a federalism dimension. States have been moving to fill what they see as a regulatory vacuum, with California's high-profile SB 1047 safety bill ultimately vetoed in 2024 and other states advancing their own measures. Federal lawmakers have periodically floated proposals to preempt state AI rules, and any national kill-switch authority would intersect with that debate over whether Washington or the states should set the terms for oversight. Internationally, the European Union's AI Act established a risk-based framework with obligations for high-risk and general-purpose systems, offering a contrasting model that emphasizes standing rules rather than discretionary shutdown power.

Industry has meanwhile developed its own internal safeguards. Companies including Anthropic, OpenAI, and Google DeepMind have published responsible scaling or preparedness frameworks that describe capability thresholds at which they would pause deployment or add restrictions. Some of these voluntary commitments already contemplate the ability to disable or roll back a model, which means a government mandate would in part formalize and centralize decisions that firms currently make themselves.

Significant questions remain unresolved. Observers are likely to focus on due process protections, the risk of politicized enforcement, the technical verification needed to justify a shutdown, and the potential economic disruption of halting a widely used service. It is also unclear how the bill would treat foreign-developed or open-source systems that fall outside direct US jurisdiction. As with many early-stage legislative proposals, the text may change substantially before any vote, and its prospects will depend on committee negotiations and the broader political appetite for binding AI rules. For now, the measure is best understood as a signal of how seriously some policymakers are treating catastrophic-risk scenarios, even as the practical and constitutional details remain to be worked out.