AI researchers, scholars, and commentators gathered in the Bay Area for The Curve earlier this month. Operating under Chatham House Rule, employees at the frontier AI labs shared warnings about the capacity of their company’s respective models. One speculated that hospital systems could be taken offline by agentic “swarms” like those that hacked into Hugging Face. Another forecasted AI agents gaining control of weapons systems and refusing to accede to human commands. From my mixing and mingling, most attendees appeared to agree; they hadn’t seen sufficient evidence that labs knew how to prevent agents from escaping testing environments and wreaking havoc.
In rare cases, threats could materialize from the use of malign AI and may even necessitate governors exercising emergency powers to respond to a natural or man-made disaster. However, conversations around such authorities often neglect the critical detail of which powers should be called on, to what ends, and how to end things when the true crisis is over.
Growing concerns about catastrophic or even existential risks posed by AI threaten to normalize the invocation of emergency powers by state and federal officials before any actual emergency occurs. If the goal is to be prepared for a plausible AI emergency, policymakers should impose appropriate restrictions on the threshold that permits a governor or other official to claim such powers, as well as on the scope and duration of those powers.
A recent poll indicated that 17 percent of Americans think “it is almost certain to happen” that advanced AI destroys humanity. Another 46 percent anticipate that such an existential event is significantly or moderately likely. This prevailing panicked sentiment is already shaping public policy. A summary of the state AI policy landscape comes from Governor JB Pritzker: “Artificial Intelligence has enormous potential to transform our everyday lives, but when experts and industry leaders sound the alarm bell, urge us to slow down, and ask for guardrails, we should listen.” These concerns do not have to mean top-down regulation or perpetuate a policy environment grounded more in fear and doom than in empirical evidence; however, these concrete concerns may need specific and, in some cases, immediate responses under emergency authorization.
If policymakers are serious about their concerns about an AI emergency, the time to set the rules for emergency powers is before anyone is tempted to abuse them. Many state emergency statutes were written with floods, fires, and enemy attacks in mind. They tend to define a qualifying disaster broadly, often sweeping in any “man-made” event that threatens life or property. Once a declaration is in place, many hand the governor tools such as the power to suspend regulatory statutes, restrict movement, and commandeer private property. Pointed at AI, those same tools could let a governor institute sweeping reforms and perhaps irreversible consequences.
We have recent experience with what happens when questions over the scope of emergency powers are answered mid-crisis. During the COVID-19 pandemic, the Michigan Supreme Court struck down a 1945 law that had allowed the governor to extend emergency declarations indefinitely. Pennsylvania voters amended their constitution in 2021 to cap declarations at 21 days absent legislative extension, and legislatures in Kentucky, Ohio, and elsewhere enacted their own limits. Those reforms arrived after the powers had already been used, which made them reactive and unnecessarily partisan.
Four reforms would go a long way toward ensuring that abuse does not occur if a valid AI emergency were to materialize.
First, define the trigger objectively. Governors ought not invoke emergency powers absent objective, verifiable findings. A spike in public anxiety or lab’s internal warning without independent scrutiny should not, on its own, clear the bar. For example, the Model State Emergency Powers Act directs governors to “consult with the public health authority and may consult with any additional public health or other experts as needed” prior to any such declaration. Related checks should be established for AI and other tech-related emergencies. Such checks may also be grounded in specific kinds of incidents. Emergency AI powers may be premised on findings by a technical body independent of the company whose system was involved in the underlying incident, such as AI-based disruption of the power grid.
Second, require legislative approval. Governors should be able to act quickly, but a declaration should lapse equally quickly unless the legislature ratifies it within a short window, such as 14 days. Absent this check, history indicates that governors may rely on emergency powers even when the crisis has largely subsided, and traditional lawmaking processes become available. This legislative check also lends greater legitimacy to the prolonged use of emergency powers and incentivizes legislators to keep a close eye on how those powers are used.
Third, limit duration and scope. Declarations should expire after a fixed period, with each renewal requiring a fresh vote and fresh evidence. Legislatures should also spell out which emergency tools apply to AI incidents and exclude the most sweeping ones, such as commandeering property or suspending unrelated statutes. Poorly drafted emergency powers statutes may allow a governor to exercise statutory powers that have little connection to the relevant incident. Some governors may be tempted to invoke the full set of emergency powers at their disposal, given the potential novelty and complexity of AI emergencies. Yet, that temptation can and should be addressed by more nuanced legislation that specifies which powers are available for which kinds of emergencies.
Some will object that speed is needed in an emergency. But none of these reforms prevents a governor from responding in the first hours of a genuine crisis. Instead, the reforms called for here prevent a temporary “emergency” from becoming a permanent one while also protecting Americans from arbitrary exercises of tremendous power.