Capability-Based Planning for AI Crisis Preparedness
This paper proposes a capability-based planning framework for AI crisis preparedness that replaces traditional likelihood-based risk assessment with a scenario-driven approach to evaluate and prioritize government capabilities under deep uncertainty.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The future of advanced artificial intelligence is a landscape of profound uncertainty. Unlike weather forecasts, where scientists can predict rain with reasonable accuracy, experts cannot agree on when, or even if, powerful new AI systems will emerge. Some foresee them arriving within a few years; others believe they are decades away. Because the timeline is so unclear, and because these systems might behave in ways no one can currently predict, governments face a difficult problem: how do you prepare for a crisis you cannot describe? Traditional planning relies on guessing the most likely danger and building defenses for that specific scenario. But if the danger arrives in a completely different form, or at a time no one expected, those specific defenses may be useless. This is why a different approach is needed—one that does not try to predict the future, but instead asks what a government must be able to do, no matter what the future holds.
A team of researchers has proposed a new method to solve this problem, drawing on strategies used by militaries and disaster response agencies for decades. Instead of asking "What is the most likely AI disaster?", they ask "What capabilities must a government have to handle any severe AI crisis?" They call this capability-based planning. To test this idea, the researchers built a practical tool: a massive grid that cross-references a wide variety of possible AI disaster scenarios with a list of specific government actions. They did not try to guess which scenario would happen. Instead, they looked at four broad categories of severe threats—such as AI systems taking control of their own operations, or AI being used to create biological weapons—and generated a diverse set of specific situations within those categories. For each situation, they asked a simple question: if this happened today, could the government actually stop it or manage it?
The researchers found that the current state of government readiness is uneven and often insufficient, though they caution that these findings are preliminary. They identified 65 distinct government functions, ranging from the ability to coordinate between different agencies to the power to shut down computer servers, but their pilot study rated only 13 of these capabilities against their library of disaster scenarios. Even with this limited sample, a pattern emerged suggesting that most of the time, the government lacks the immediate ability to act without significant prior preparation. For example, if a dangerous AI system were to suddenly escape into the internet, the government would likely not have the legal authority or the technical means to stop it in the moment. The study suggests that existing institutions are not ready to handle these crises on the fly. However, the researchers also found that some capabilities are useful across almost every possible disaster. These include the ability to mobilize different government departments quickly and to coordinate with international partners. Because these actions are helpful no matter what the specific threat is, the researchers argue they should be the first priority for investment, regardless of how likely any single disaster might be.
The study also revealed that the type of threat matters less than the specific nature of the scenario. Whether the danger comes from a loss of control over an AI system or from a cyberattack, the gaps in government readiness look very similar. This suggests that governments should not focus their planning on specific labels like "cyber" or "biological," but rather on the underlying mechanics of the crisis. The researchers discovered that for the most severe scenarios, such as those that could threaten human existence, the government needs a very specific set of tools: the ability to contain the threat, to control the computer systems running the AI, and to stop the deployment of new models. Without these specific capabilities, the response to a crisis would fail completely. In contrast, for other types of threats, the government might be able to rely on existing emergency networks, but for the most dangerous possibilities, the current infrastructure is not enough.
One of the most striking findings is that governments can start preparing for these crises without needing to agree on exactly when they will happen or how likely they are. By focusing on "no-regret" actions—steps that are useful in almost any future—the government can build a foundation of readiness that remains valid even as the world changes. The researchers found that coordination and the ability to act decisively are the most critical needs. They suggest that the first step for any government is to assign a specific lead agency for each type of crisis and to build trust networks with the companies that build and use AI. These steps do not require predicting the future; they simply require acknowledging that the current system is not built to handle sudden, severe disruptions.
The paper does not claim to have solved the problem of AI safety, nor does it offer a perfect list of every possible solution. The researchers were clear that their work is a pilot study, a proof of concept that demonstrates the value of this new way of thinking. They tested only a fraction of the possible government functions against a limited set of scenarios, and they acknowledge that more work is needed to fill in the gaps. However, the tool they built offers a way to move beyond the paralysis of uncertainty. Instead of waiting for a specific threat to appear, governments can now map out the exact capabilities they need to survive a wide range of potential futures. The study concludes that the biggest bottleneck in a crisis is not the technology itself, but the state's ability to respond. By identifying these gaps now, governments can ensure that when a crisis arrives, they are not caught off guard, but are ready to act with the authority and resources required to protect their citizens.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.