Trapped in the whiplash of increasing dystopian headlines describing the latest disaster and utopian headlines about the promise of artificial intelligence (AI), many have asked the natural question: “How can the second save us from the first?” Several entities, including the Naval Postgraduate School and Department of Homeland Security’s Science and Technology Directorate, have commissioned studies to explore this exact question, which have uncovered a treasure trove of research, applications, and existing implementations of AI in emergency management (EM). An in-depth landscape assessment conducted in 2024 dug deep into research articles, preprints, code repositories, and surveys at the intersection of applied AI and EM.
In the wake of recent or local emergencies, many articles focus on exceedingly specific threats (e.g., floods), and they apply technologies with the same specificity. It is hard to understand whether that technology could apply to other threats – or even the same threat in a different locality or at a different time. Instead, understanding the landscape of AI in EM requires zooming out to understand each threat and technology in context. This landscape assessment sought to do the following:
- Categorize where the technology applies along the emergency timeline: from the mitigation stage – well before the event – to detection around the onset of the event, on to response to the event, and recovery after the event.
- Identify the applicable domain, for example, a physical system such as a public building, telecommunications infrastructure, the natural environment, or others.
- Evaluate the technology based on scientific data, assessing the data type (such as imagery or geospatial information) and other technical details about the data.
Implementing this framework on hundreds of sources illustrated trends, gaps, and a broader vision for AI in EM. Researchers are finding some of the biggest challenges in emergency management and tackling them head-on, making great progress with many. Researchers, entrepreneurs, private and public industry, and others have roles to play, as well, but gaps, challenges, and unanswered questions remain. One of the biggest questions is that of generalization. In AI, a technology “generalizes well” if it performs its task as well in a future emergency as it did during the emergency in which it was trained. Unfortunately, many AI technologies are so focused they do not generalize beyond the data on which they were trained. Generalization is a gate through which researched technologies must pass before being implemented in a real system. There is also a potential risk if not properly assessed through rigorous testing and evaluation before and during implementation. Another important filter is that of pure feasibility, which is challenging to assess from a technological standpoint.
Emergency management experts must validate each proposed technology against the chaos of real emergencies. This was never more apparent than in the recent discourse between national laboratory data scientists, state and local emergency managers and first responders in New York, and the University of Albany faculty in computer science and emergency management. Applying AI to EM must be done at the confluence of the two fields – understanding the potential and pitfalls of AI technologies (best analyzed by data scientists) but also with an appreciation for the real world that only those who have managed emergencies have.
The landscape assessment garnered insights focused on four technologies foundational to future AI technologies in EM and nine broad technologies with great potential to benefit EM.
A Good Backbone
Like many other industries are discovering, using AI requires infrastructure that had little reason to exist before AI. This applies to hardware and physical networks but also to the software surrounding AI, and even to the processes and public perception of AI technologies.
Operational AI depends on highly sophisticated infrastructure to enable data delivery to high-powered servers and back to its users in a timely fashion. The increase in demand for graphic processing units (GPUs) – the hardware that runs much of the development of AI models – has been prolific and will be critical. In fact, since emergencies can happen quickly and with little warning, unused computer resources must be kept as “inventory” to be deployed during the event. Network infrastructure will likely be as important to enable AI broadly, but the challenge for network connectivity is amplified in EM, whereas it is normal to require better-than-average network resources in rural, remote, or even destroyed areas. Additionally, such sophisticated infrastructure requires ongoing maintenance and the power to support the infrastructure.
Trusting the output of AI is also crucial to future success in EM. That trust depends on the governance of and communication about AI – from setting reasonable policies of how public entities adopt AI technologies to allowing for human overrides at the time of the event. Feedback garnered from EM practitioners during stakeholder outreach for the landscape assessment heard fundamentally opposite arguments regarding these points: Some believe that “the most cutting-edge AI should be used because people’s lives are at stake,” yet others believe that “we can’t use AI technologies until it is completely mature and trusted because people’s lives are at stake.” The outcome of that argument – and it may not be homogeneous across all countries or states – will have an outsized effect on how AI contributes to EM. However, trusting AI goes beyond simply understanding the output of the specific technology and expands into trusting that the whole system – data, network, software, and all – has not been compromised by malicious actors. Cybersecurity will, therefore, grow in importance as AI is further incorporated into EM.
Enabling Technologies
Assuming that a trusted AI infrastructure, as described above, will be in place to support AI applications in emergencies, the landscape assessment identified nine technologies predicted to have a large impact on EM in the next decade. Among the list was also the most possible with current technology: Using AI-enabled productivity applications can assist emergency managers in the repeated and sometimes mundane tasks of their daily jobs – from turning emergency plan outlines into full reports to converting their notes into a full after-action assessment. Chatbots and other tools are being released to the private sector to do this in a general context. With some modifications for the EM sector, it could ease emergency managers’ considerable workload. Additionally, using AI as a means of public-facing communication could help emergency managers ensure that everybody gets the right information in the right way during emergencies. The ability of emerging AI-driven capabilities to produce better translations and modify them to avoid fearmongering is exciting. Finally, the “broadband” ability of AI to tend to large amounts of data creates opportunities for more effective and efficient planning. This would allow emergency managers to use AI as a planning assistant, with the benefit of fact-checking previous events to avoid historical missteps or checking others’ plans to avoid allocating the same asset twice during the same emergency.
That same “broadband” ability is beneficial on the ground as well. AI-filtered domain awareness promises to provide only essential and consequential information to the emergency manager, as opposed to the overwhelming data environment they currently grapple with. That capability will also improve its ability to predict and detect emergencies before and during onset, predict a storm’s course, and measure effects during and after an event. While many technologies already perform these tasks, there is room for improvement in detection and how well the ability to generalize to the future. Additionally, adding new data streams, such as chemical sensors or radiation detectors, will benefit prediction capabilities, but this technology is currently nascent.
In the future, the technologies underlying modern AI can be pushed further. Modern optimization methods, including reinforcement learning, can improve the routing of traffic, assets, and resources during an emergency, sometimes with automation. Using risk models in conjunction with these improvements can also help minimize overall consequences, even when the worst consequence is rare and hard to foresee.
The Next Generation
The landscape assessment was cognizant of the current and evolving state of technology – some of the new and exciting concepts in the literature (and researchers’ surprise at them) highlighted the calcification of old ways of thinking. To combat this, researchers solicited ideas from one of the best sources for off-the-wall ideas: a room full of 100 college students, bribed with free food. In this “sandpit” exercise, performed at the University of Albany in conjunction with its College of Emergency Preparedness, Homeland Security and Cybersecurity, students were asked to perform in one day what many researchers do for their career: Identify an important EM challenge, create an AI-enabled solution to the problem, and validate and report on that solution.
The students conceptualized everything from an AI-enabled clearinghouse for privately owned resources to chatbot systems that could reroute needed supplies during an emergency event. These ideas fortified belief in the above technologies – with a majority of ideas pertaining to efficient asset deployment through modern optimization.
Looking Forward
Performing this landscape assessment, which started pessimistically, was an exercise in optimism. So often when discussing a possible EM challenge, there is already an AI technology that could alleviate or even solve that problem. It should be encouraging, not frightening, to practitioners to see how much research is connecting challenges with AI solutions, and also how students – the next generation – are thinking about the field. While there are still infrastructure hurdles that AI will need to overcome in EM, AI technologies will be the subject of deserved utopian press as they begin to help practitioners prevent, mitigate, and recover from emergencies.
Alex Hagen
Alex Hagen is a data scientist who works broadly across detection and material interdiction spaces to improve analysis using modern machine-learning methods. After a half decade designing neutron detectors and active interrogation techniques and subsequently analyzing data from such experiments, he knows how to combine field implementation with advanced analytical techniques. His research has been published across many nuclear engineering and physics journals, including Journal of Physics, Nuclear Instruments and Methods, and the Transactions of Nuclear Science. His conference presentations have taken him across the world, including to the International Conference of Nuclear Engineering in Prague and the Advanced Computing and Analysis Techniques Conference in Saas-Fe, Switzerland. He has contributed to several high-energy physics collaborations, including Belle2, and PICO. He holds a PhD, MS, and BS in nuclear engineering from Purdue University.
Jonathan (Jon) Barr
Jonathan Barr is a senior systems engineer with Pacific Northwest National Laboratory’s Threat Prevention and Resilience Group. Barr works extensively with stakeholders across the national security and first response communities to understand their operational needs and develop research roadmaps to develop and implement the advanced technologies to meet those needs. He has applied his experience in developing human-centric artificial intelligence concepts and technologies to support the Institute of Electrical and Electronics Engineers in their work to develop ethical AI standards and certifications. Barr is an INCOSE-certified systems engineering professional with an MS in mechanical engineering and materials science from Washington University in St. Louis, graduate certification in medical sciences from the University of Washington, and a BS in mechanical engineering from Kansas State University.
