Building centaur responders: is emergency management ready for artificial intelligence?

Building centaur responders: is emergency management ready for artificial intelligence? - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Disasters . 2026 Apr 21;50:e70054. doi: 10.1111/disa.70054 Search in PMC Search in PubMed View in NLM Catalog Add to search Building centaur responders: is emergency management ready for artificial intelligence? Christopher Whyte Christopher Whyte , PhD 1 Virginia Commonwealth University, United States Find articles by Christopher Whyte 1, ✉ , Brittany ‘Brie’ Haupt Brittany ‘Brie’ Haupt , PhD 1 Virginia Commonwealth University, United States Find articles by Brittany ‘Brie’ Haupt 1, ✉ Author information Article notes Copyright and License information 1 Virginia Commonwealth University, United States * Correspondence , Christopher Whyte, PhD, L. Douglas Wilder School of Government and Public Affairs, Virginia Commonwealth University, 923 West Franklin Street, Box 842028, Richmond, VA 23284‐2041, United States, Email: [email protected] ✉ Corresponding author. Received 2024 Apr 10; Accepted 2026 Mar 10; Issue date 2026 Jul. © 2026 The Author(s). Disasters published by John Wiley & Sons Ltd on behalf of ODI Global. This is an open access article under the terms of the http://creativecommons.org/licenses/by-nc-nd/4.0/ License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non‐commercial and no modifications or adaptations are made. PMC Copyright notice PMCID: PMC13096858  PMID: 42011101 Abstract This article examines the preparedness of emergency management (EM) for addressing questions pertaining to artificial intelligence (AI), encompassing its benefits to EM missions, the potential biases, the societal impacts, and more. We pinpoint two key shortcomings in early EM research on AI: (i) insufficient discussion of both AI's history and evolution; and (ii) a lack of frameworks for organising AI‐related knowledge‐building efforts. We offer a comprehensive survey of AI that is tailored for EM scholars and practitioners. Then, utilising the case of cyberspace and domain concept development, we explore the challenges of applying legacy ideas to disruptive technologies. We argue that current EM frameworks struggle to align with emerging technologies; as a result, we propose a new framework that bridges EM‐specific concerns and broader social science inquiries and AI. This framework aims to facilitate research and practical efforts to navigate the complexities of AI technology within the context of EM. Keywords: artificial intelligence, cyberspace, disruptive technology, large language models, machine learning 1. INTRODUCTION Few developments in the twenty‐first century have impacted societal expectations about future prosperity, security, and equity as greatly as has the artificial intelligence (AI) boom of the early 2020s. Statements like this seem almost remarkable given the unprecedented experiences of extremism, climate change, digital insecurity, and public health that have characterised the first two‐and‐a‐half decades of the new millennium. And yet, machine intelligence that can mimic human behaviour and biological intuition has rapidly come to underwrite transformation in most strata—social, political, and economic—of the developed world. Luminary voices like Steve Wozniak and Bill Gates have espoused the prospective value of AI for humanity if we can get the technology right (Evgeny, 2023 ), forecasting futures of immense prosperity and the resolution of diverse inequalities (Baum et al., 2023 ). At the same time, a large and vocal community of experts continues to voice concern about AI, going as far as advocating for the cessation of its development until its safety can be assured (Struckman and Kupiec, 2023 ). After all, they contend, bugs in the technology today could spell existential challenges for humanity in decades to come. The contemporary AI boom is driven particularly by recent experiences with new generative large language model (LLM) tools like ChatGPT, Claude, and Gemini. Within this context, professional fields of research and practice across the social sciences are attempting to grapple with questions of innovation and adaptation. What does such a complex exogenous shock 1 mean for the traditional paradigms of knowledge building in a field like emergency management (EM)? How should such a dynamic technological transformation change the toolkit of the responder? How might AI alter the society being made safe by EM activities? Indeed, these issues are particularly difficult to address in practitioner‐oriented fields like EM. 2 Here, clearer intersection between foundational research and on‐the‐ground practice elevates the significance of established frameworks and increases the likelihood of iterative approaches to novel conditions (that is, putting new wine in old bottles). In this article we take up the question of EM's readiness for AI and the prospect of a much‐changed public safety paradigm in years to come. While we admit (and even argue) that contemporary discourse involves rampant AI sensationalism, there is little doubt that recent advances in computing power, algorithms, and the data priors required for truly intuitive machine intelligence have started to transform society. Within the EM enterprise, AI has begun to be integrated into preparation and response planning along with risk and crisis communication efforts (Photis and Grekousis, 2012 ; Sun, Bocchini, and Davison, 2020 ; Gupta et al., 2022 ). Fortunately, AI is far from the only exogenous shock to impact society in recent history, even if it promises to be the most dynamic one. In the sections below we discuss AI and those recent developments that make now the right time to ask questions about corresponding implications for the discipline's core paradigms. Indeed, since few treatments of AI exist in the EM field, we attempt to do so in some detail. Next, after briefly surveying the limited amount of EM work that considers or leverages AI, we consider adaptation to a recent exogenous technological shock—the rise and socialisation of the internet since the 1990s—in public affairs and the broader field of political science. Our analysis suggests that readiness to grapple with the implications of complex exogenous shocks like cyberspace or AI—meaning an ability to organise clear research programmes of inquiry and translate scholarly know‐how into practice—most clearly corresponds with the presence of organising frameworks that help stakeholders overcome multifaceted uncertainty about new circumstances. That said, such frameworks also come with risk, particularly that of oversimplifying emergent conditions. The dominant cyberspace framework conceived and exported by American national security scholar‐practitioners in years past exemplifies this risk—capable of anchoring multidimensional development of new thought but overly rigid in the face of hybrid phenomena (such as digital influence operations). Clearly, organising frameworks must thus not only serve to remedy uncertainty for the most stakeholders possible, but they must also be geared to complement, but not supplant, existing ones. The final sections of this article assert that traditional EM frameworks are hampered by a general incompatibility with the underlying technology under consideration, a challenge that more technical or informationally‐oriented fields like communication studies may lack. To resolve this, we offer a framework for organising research and practical efforts around new AI technologies. This is intended to serve as an intermediate lens between the EM foci of preparation, mitigation, response, and recovery, and the broader social science areas of inquiry that underwrite AI research. Specifically, we propose that EM scholars, professional organisations, and practitioner communities adopt a four‐part typology of AI as it pertains to the EM mission: AI as a tool ; AI as an intermediary (or agent); AI as a talent pool ; and AI as an attack surface . We call this the TITAS (Tool/Intermediary/Talent pool/Attack Surface) Framework. Under such an organising scheme, we believe that EM would become more agile and capable in its unavoidable need to grapple with the implications of AI and effectively build centaurs—responders, researchers, and planners that effectively blend the intrinsic strengths of both machine and human being. To make that organising payoff immediately concrete, we preview the TITAS Framework against the EM lifecycle in Table 1 . Rather than treating ‘AI’ as a single object or tool category, TITAS forces a structured accounting of how AI manifests (as a tool, intermediary, talent pool, and attack surface), which can be mapped across mitigation, preparedness, response, and recovery to clarify effectively how the same AI capability can create readiness in one phase while generating new dependencies or vulnerabilities in another. TABLE 1. TITAS Framework for operational opportunity planning. Operational planning by EM phase Mitigation Preparedness Response Recovery TITAS element AI as a tool (applications that augment EM processes) Risk modelling and hazard forecasting; scenario planning support; infrastructure inspection/prioritisation assistance. Training and exercise support; pre‐positioning/logistics planning; resource and capability planning. Decision‐support for ICS/EOC; routing/resource allocation; situational awareness summarisation. Damage assessment support; case management triage support; rebuilding prioritisation and programme monitoring. AI as an intermediary (AI ‘agency’ absent direct human inputs) Automated anomaly detection in critical infrastructure; autonomous monitoring/triggering for protective actions. Automated alerts/escalation logic; agent‐mediated coordination workflows; automated intake/triage ‘front doors’. Semi‐autonomous prioritisation (dispatch/triage queues); bot‐mediated public information and rumour triage; autonomous sensor fusion alerts. Automated eligibility/benefit routing; agent‐driven fraud detection; autonomous backlog prioritisation in recovery pipelines. AI as a talent pool (priors: data, algorithms, supply chains, inputs) Baseline risk datasets; standards for geospatial/asset data; procurement/vendor choices that shape future capacity. Data governance and sharing MOUs; model validation/testing pipelines; workforce skill building and QA processes. Real‐time data feeds and interoperability; ‘human in the loop’ QA staffing; compute/platform continuity planning. After‐action data curation; model updating/retuning governance; documentation for lessons learned and auditability. AI as an attack surface (AI shapes EM issues and recursively reshapes EM) Model deception affecting long‐horizon planning; disinformation shaping mitigation preferences and policy coalitions. Training data poisoning; supply chain compromise; adversarial testing/red teaming gaps. Adversarial inputs under time pressure; spoofed sensors/LLM ‘hallucinated’ guidance; communication manipulation during a crisis. Fraud/scams amplified by AI; misinformation about aid eligibility; long‐tail trust erosion in institutions/processes. Open in a new tab Note: ICS/EOC = incident command system/emergency operations centre; MOUs = Memorandums of Understanding; QA = quality assurance. Source: authors. Table 1 is not meant to be an exhaustive typology, but rather a deliberately compact way to signal what TITAS analytically procures. First, it shows that ‘being prepared for AI’ is not a single capability or procurement matter, but instead, a phase‐specific configuration problem in which the balance among tools, intermediaries, talent pools, and attack surfaces shifts as EM work moves from risk reduction to time‐pressured action and then to reconstruction. Second, the mapping clarifies why preparedness requires organising structures (that is, shared definitions, governance choices, training, and coordination routines), namely because AI's operational value and its failure modes are produced by the same socio‐technical arrangements that enable them. In the remaining sections, this paper builds the need for better organising frameworks sequentially. It does so first by illustrating the need therefor today as distinct from in years past. Then, it uses a historical exploration of a previous case of orientation towards a novel disruptive technology—that of cyberspace—to contextualise contemporary EM engagement with AI. Next, after fleshing out the framework above, we show that the TITAS Framework's deeper utility for emergency management lies not only in mapping AI to mission phases (as in the preliminary illustration shown in Table 1 ), but in using its analytic assumptions to structure knowledge organisation and strategic planning to clarify what must be learned, governed, and coordinated as ‘centaur’ capacity is built. 2. WHY SHOULD WE CARE ABOUT MACHINES THAT MOVE, SENSE, LEARN, AND INTUIT? As we pass the midpoint of the twenty‐first century's third decade, it feels like we are at a watershed moment of sorts for machine intelligence. That said, it is important to recall that the conversations we are having now are far from new or novel. Today, ‘artificial intelligence’ is an umbrella term that represents a broad array of techno‐scientific and techno‐social innovations spread across an almost bewildering array of scientific disciplines, from information theory to sociology. 3 But there is a linearity found in the history of humankind's consideration of artificial thought and form that quite clearly paints a contemporary picture of AI as something balanced on a precipice. On the one side, AI remains a constellation of techniques enabling mimicry of human agency, most commonly in quite narrow settings. On the other side, a select set of AI productions in just the past decade have finally exhibited the characteristics of intuition, a quality of intelligence correlated with animal sentience and a phenomenon that—quite ironically—presents as quite alien when found in a machine. What AI looks like tomorrow depends on how this sophistication of machine intelligence is guided, a process that will inevitably reflect human society, values, and interests. 2.1. Constant idea, recent reality The development of the modern AI field is closely linked to the development of computers. True, prior to the first digital computers of the 1940s—the Colossus machines of Second World War vintage—ideas about non‐human intelligence permeated philosophical and scientific thought going back millennia. Aristotle thought intelligent tools impossible and used the concept, in The Politics , as a justification for Athenian views on slavery (Qerimi, 2020 ). Leonardo da Vinci and others designed and even built mechanical automata, the forerunners of modern robots (Iavazzo et al., 2014 ). And mathematicians like Gottfried Leibniz, the father of calculus, designed logical systems of rules that they thought might be used by a future machine to view and react to the world around it (that is, ‘if X is true, then Y …’) (Kramer, 1996 ). Until the computer, however, no effort to realise machine intelligence had access to anything that could truly mimic the raw processing potential of the human brain. Early computers, of course, were only powerful in a limited sense and so the first modern efforts to organise thinking on AI were narrow in the extreme. Building on work undertaken by scientific luminaries in the years after the Second World War, a conference in 1956 at Dartmouth College in the United States brought together the best minds of the Anglo‐American scientific community to explore a simple prompt: to investigate if ‘every aspect of learning or any other feature of intelligence [could] in principle be so precisely described that a machine [could] be made to simulate it’ (McCarthy et al., 2006 , p. 12). Knowing that digital computers were capable of undertaking tasks that a human could not—or, at least, at a speed and scale that a human could not match—the attendees attempted to answer the question of whether a machine could simulate ‘the higher functions of the human brain’, both practically and philosophically (Moor, 2006 ). The result was the first intentional organisation of the modern AI field, focused on exploring (i) the feasibility of advanced computers that looked more like the human brain than a piece of machinery, (ii) concepts of language processing and understanding, (iii) creative and abstract reasoning, and (iv) machine self‐improvement (Jensen, Whyte, and Cuomo, 2020 ). These efforts reflected different takes on a simple and historically ubiquitous question that had received new answers in the years following the Second World War: can machines think? Many people, like Norbert Wiener ( 2019 ) in his seminal 1948 book titled Cybernetics or Control and Communication in the Animal and the Machine , followed Leibniz and other mathematicians in asserting that a computer could be designed to play and even outplay a human at games like chess (and thus perform other dedicated tasks) if given sufficient rational rules from which to work. John von Neumann, a genius even by the standard of geniuses, suggested that complex systems within computer architecture could be made to mimic biology, an idea that dominates work on deep learning and environmental recognition today (Al‐Hashimi, 2023 ). And Alan Turing, today a household name, proposed an even simpler answer to the question with one of his own: if a machine can fool a human into believing it is actually another human, then is it not intelligent? (French, 2000 ). What followed was the first ‘summer’ of AI development in which researchers created ‘worlds’ in the form of narrow application machine systems (Toosi et al., 2021 ). These were excellent at performing simple tasks using just a small amount of data that might otherwise require a human being. Yet these early successes had extreme limitations, most notably a lack of progress on environmental recognition and programmes that could only handle small amounts of data before veering towards bizarre behaviour. This produced the first ‘winter’ of AI in 1974 when funders in and the interest of both industry and government declined dramatically (Floridi, 2020 ). A decade later, in the 1980s, a new paradigm of building machine intelligence for more generalised tasks, which emphasised broad but shallow (rather than deep but singular) knowledge of a subject, brought about the second ‘summer’ (Floridi, 2020 ). Despite limited progress on building better computers upon which AI could operate, this approach found success, even producing a machine—Deep Thought—that problem‐solved its way to beating a chess grandmaster for the first time (Hsu et al., 1990 ). Governments jumped into AI development again, with Japan in particular providing the investment model that others would follow (Gonsalves, 2019 ). Almost as quickly, alas, interest dried up and a second ‘winter’ arrived. The hardware simply was not sufficient to sustain a boom in more generalised AI applications. The difference this time was that scientists and technologists saw more clearly the problems to be solved and so did not drift away from AI. Another decade later, computers continued to become more powerful and the second ‘winter’ ended, ushering in a third ‘summer’ that continues to this day. Early in this third boomtime, new mathematical advances in decision trees, neural networks, and naïve Bayes classifiers made it easier to provide large amounts of data to the ‘expert systems’ developed during the preceding ‘summer’, culminating in perhaps the most famous AI event of the twentieth century (Buchanan and Imbrie, 2022 ): in 1997, a programme called Deep Blue decisively beat the (arguably all‐time greatest) chess grandmaster and world champion, Garry Kasparov (Newborn, 2012 ). 2.2. Why now? Those interested in AI often differentiate weak AI from strong AI. Weak (or narrow ) AI is the only AI that the world currently knows and implies any system that can mimic the cognitive functions of a human being in a limited setting. Strong AI (often artificial general intelligence ) is machine intelligence that can solve problems without prior context, like science fiction's Skynet, HAL 9000, or Data (of Star Trek fame). Other authors break down this distinction further into categories that describe types of intelligence: reactive machines; limited memory; theory of mind; and self‐awareness (see, for example, El Samad, Nasserddine, and Kheir, 2023 ). Respectively, these are machines that (i) can only respond to stimuli in a single instance, (ii) can use iterative experience to learn, (iii) can perceive intangibles in the environment like emotion or psychological conditions, and (iv) can understand their own existence in terms distinct from those programmed in by a creator. In this scheme, the limitations of AI at the start of the third ‘summer’ become quite clear: Deep Blue, as impressive as it was, was still only a reactive automaton. The reason why conversations about AI at the quarter point of the twenty‐first century are far more pressing than they were in the time of Deep Blue is that several recent advances have begun to move it beyond reactive formats. The quest for general machine intelligence has always been about the interaction of three things (other than the input of humans): data; algorithms; and computer power. In just the past 15 years, each of these areas has seen revolutionary advances and AI programmes that vastly outclass Deep Thought or Deep Blue in their capacity to perform not just a few narrow tasks, but many undertakings. 2.2.1. Data The data revolution of the twenty‐first century is simple enough to understand. After the success of building expert systems in the 1990s based on new design principles, numerous researchers began to experiment with how such systems might perform if given large amounts of data alongside more limited instructions. For instance, a 2001 experiment on language identification (that is, can computers interpret confusing wordage correctly?) found that existing programmes were about 82 per cent accurate when fed millions of examples ahead of time, but that scaling up this firehose of data by a factor of 1,000 consistently produced 97 per cent accurate estimates (Buchanan and Imbrie, 2022 ). This realisation galvanised young scientists and prompted a renaissance in focus on a potential competitor to the ‘expert system’ model of machine intelligence that had generally been ignored for decades: machine learning (Patterson, 1990 ). Expert systems learn from data in a supervised fashion, meaning that the user tells the algorithm what rules to follow or what labels to adopt. With machine learning, knowledge that the algorithm develops about the world is attained via an unsupervised survey of immense amounts of data. This process of pattern recognition is aided by a neural network. Neural networks, the tool at the core of what is called deep learning , are a series of layers of stores of information that are organised based on the patterns initially found in training data (Buchanan and Imbrie, 2022 ). Imagine that a response unit receives a blurry satellite image during a natural hazard situation. To make this useful for incident response (such as to identify stranded persons), teams of professionals might apply their talents one after the other (for instance, adjusting the sharpness of the image or using personal knowledge of terrain to identify landmarks). A neural network does exactly this layered task, first by training itself what to look for by studying huge amounts of data and thereafter by filtering new inputs through this lens. And, in the late 2000s, neural networks demonstrated their power, with programmes like GoogLeNet (Al‐Qizwini et al., 2017 ) and AlexNet (Alom et al., 2018 ) able to recognise imagery accurately at scale in a way that Deep Blue and its ilk never could. 2.2.2. Algorithms The algorithmic revolution of the 2010s is intrinsically tied to the unprecedented expansion of focus on machine learning. Following successes with image recognition, language classification, and more, renewed focus on machine learning principles produced breakthroughs in the unsupervised generation of content. Generative adversarial networks, for example, took the principle of unsupervised machine learning forward by forcing two algorithms to interact with one another (Creswell et al., 2018 ). While one algorithm used its knowledge foundations to generate requested content based on simple parameters, a second discriminating algorithm gauged its performance. Outputs based on this interaction, which continued until the discriminator was satisfied, rapidly overcame conventional challenges in asking computer programmes to generate convincing multimedia and led directly to one of today's most notorious AI applications: the deepfake (Whyte, 2020 ). Tied to such breakthroughs, perhaps the most significant advance in algorithmic capacity in AI development lies with something called reinforcement learning (Buchanan and Imbrie, 2022 ). This is the process by which a programme learns not from enormous amounts of training data, but from direct interaction with the environment. The most famous examples of this approach are probably the AI models of the company DeepMind, such as AlphaGo (Chen, 2016 ). These were specifically designed to beat the world's best Go players, beginning with the assumption that the game was so much more complicated than chess that something like Deep Blue would never stand a chance against Go grandmasters. To prepare its AI, DeepMind simply forced its neural networks to play games against themselves millions of times. After some initial success, the approach was refined to include a pair of neural networks that allowed the AI not only to simulate all possible moves, but also to judge which one was most likely to achieve victory. The result was something incredible: an AI that, in competition against the world's best Go grandmaster, Lee Sedol, won by making moves that the global Go community could not fathom. AlphaGo, in beating its opponent, had demonstrated intuition (Buchanan and Imbrie, 2022 ). 2.2.3. Computing power The final revolution of recent years is perhaps the quietest but is once again intrinsically tied to the data and algorithm revolutions that have characterised the first decades of the twenty‐first century. Here, the success of AlphaGo and its successors was, as Buchanan and Imbrie ( 2022 ) describe, a kind of Sputnik moment for the global community, as governments and industry began to funnel tens of billions of US dollars into AI research. In particular, the success of AlphaFold (an offshoot of AlphaGo) in rapidly solving the mystery of protein folding that had eluded scientific explanation for decades invited a kind of confidence in AI's relevance that arguably did not exist previously (Ruff and Pappu, 2021 ). In this context, technology futurists turned to one of the perennial challenges of AI development: computing power that failed to support scientific advances just as momentum was building. The result has been the development of computer processor hardware specifically designed to support machine learning calculations—graphics and tensor processing units—and the securing of sufficient computing capacity to support visions of artificial general intelligence in decades to come (Ilievski, Zdraveski, and Gusev, 2018 ). Indeed, this new hardware has been at the heart of incredible demonstrations of AI's potential in just the first years of the 2020s, including another DeepMind product—AlphaStar—that beat Starcraft II grandmasters and has attracted the attention of militaries around the world interested in machine learning for national security (Vinyals et al., 2017 ). Furthermore, of course, this new hardware has most recently underwritten the most visible manifestation of AI to the average global citizen in the form of generative AI like ChatGPT, constituting massive neural networks available to billions that are at once both uncanny and amusingly human in their clumsiness (Wu et al., 2023 ). In short, AI is worthy of serious attention today not simply because of the ‘democratisation’ of AI that has come with the release of new generative LLM tools for the use of global publics. Rather, AI must enter conversations about research and practice in fields like EM because something more than reactive machine intelligence now seems feasible. Specifically, data are not only available in incredible quantities but are now being manipulated and leveraged in unprecedented fashion. Algorithms are improving themselves, creating facsimiles of intelligence independent of what human input or data might produce. And the computer power to sustain development for years to come, thanks to recent design and production advances, now unquestionably exists. 3. AI IN EM: ISLANDS OF FOCUS Despite the relatively ubiquity of conversation about AI and the availability of AI tools since at least 2010–12, engagement with either the utility of the underlying technologies or issues of ethical design are relatively young in the EM field and tend to focus on response practices and policies (Chen et al., 2019 ; Sun, Bocchini, and Davison, 2020 ). This is somewhat surprising particularly given EM's self‐assigned identity as a cognitively‐centred enterprise (Comfort, 2007 ; Comfort and Rhodes, 2022 ). EM scholars highlight AI's ability to process disaster‐related data efficiently, making it an indispensable tool for decision making, particularly in the response phase, and to categorise these methods into supervised and unsupervised models, deep learning, reinforcement learning, and optimisation techniques, each with specific applications tailored to various disaster management tasks (Yuan, Zhang, and Liu, 2015 ; Fotovatikhah et al., 2018 ; Zhang, Lv, and Dhakal, 2019 ; Russell and Norvig, 2022 ). In the mitigation phase, supervised and unsupervised models are widely used for forecasting hazards, assessing vulnerabilities, and creating mitigation strategies. For example, AI techniques like logistic regression and neural networks have been employed to develop susceptibility maps for hazards such as landslides and avalanches, allowing decision‐makers to understand risk areas better. During the preparedness phase, AI supports early warning systems and real‐time disaster detection (Tan et al., 2021 ). Notable examples include the use of convolutional neural networks for image recognition to predict flood zones and recurrent neural networks to process weather data for early storm warnings. In the response phase, AI applications concentrate on event mapping, damage assessment, and resource allocation. For instance, support vector machines and decision trees have been applied to map disaster events using social media data, while neural networks assist in prioritising relief efforts. Lastly, in the recovery phase, AI aids in assessing reconstruction needs and tracking recovery progress, such as using deep learning models to evaluate structural damage and prioritise rebuilding efforts. While AI offers transformative benefits, there are several challenges hindering its broader adoption. These include the need for high‐quality, real‐time data, which are often unavailable or incomplete during disasters, financial resources, training opportunities, and EM practitioner knowledge and technological familiarity. Computational resource demands, particularly for training deep learning models, can be prohibitive, limiting access by smaller organisations. Additionally, integrating AI systems into existing disaster management frameworks incorporates ethical concerns, such as data privacy and algorithmic bias, which further complicates implementation. A survey of journals linked with the field conducted by the authors found no more than eight peer‐reviewed research articles that engaged with AI in some sense. An expanded search added another 10 articles in peer‐reviewed settings that were relevant to EM, including applications of AI connected to the COVID‐19 (coronavirus disease 2019) pandemic (see, for example, Janis and Mann, 1977 ; Gupta et al., 2022 ; Zhang et al., 2022 ). The vast majority of articles published by EM‐linked outlets (only three publications) are to be found in Natural Hazards (Hashemi et al., 2016 ; Sun, Bocchini, and Davison, 2020 ; Tan et al., 2021 ). By and large, EM work on AI falls into one of three categories. First, there are a handful of systematic literature reviews that catalogue specific technical developments relevant to sub‐areas of EM work (see, for example, Huang, Wang, and Liu, 2021 ). Second, there are somewhat more studies that ruminate on the applicability of such developments to the practical work of EM (in the oil industry or for wildfire response, for instance) (see, for example, Zarghami, Ashkan, and Dumrak, 2021 ). Third, a small number of works use AI as an aid to empirical analysis (see, for example, Sharma et al., 2019 ). In many ways, EM's treatment of AI tracks with that seen in other areas of the social sciences and even within the broader political science and public policy disciplines. As work on disruptive societal developments and scientific inquiry notes, the early engagement of academic fields of study with exogenous shocks is often characterised by a divergent focus on the most extreme levels of analysis. On the one hand, scholars are drawn to macro debates about emergent conditions and the implications for foundational knowledge of a given field of study. On the other hand, researchers rapidly orient on technical analyses and implications for practice. Both are desirable and even necessary for the development of scientific inquiry that parses the divide to produce core assumptions, mid‐level theory, and testable hypotheses. Where EM parallels other areas of the social sciences is in the focus on technical implications. What is missing thus far are macro treatments of AI governance, design, and interaction with the core concepts of the field. This absence is not unexpected given the relative newness of the phenomenon, but it does already contrast with what might be seen in other spheres of the social sciences. The international relations field of political science, for instance, has seen a rapid proliferation of treatments of AI's meaning for core concepts (like deterrence, power, and sovereignty) in the discipline's top outlets in the past decade (see, for example, Johnson, 2019 ; Jensen, Whyte, and Cuomo, 2020 ; Goldfarb and Lindsay, 2022 ). Likewise, since 2015, the field of communication studies has seen no fewer than 11 special editions of high‐impact journals centred on AI or specific dimensions thereof. The integration of AI into disaster management policies necessitates a comprehensive approach to harness its transformative potential while addressing its associated challenges. AI technologies provide unprecedented capabilities for predictive modelling, resource optimisation, and decision making, significantly enhancing all phases of disaster management—mitigation, preparedness, response, and recovery. For instance, advanced neural networks and reinforcement learning models enable precise hazard mapping, dynamic resource allocation, and real‐time risk assessment, making disaster response more effective and recovery efforts more efficient. Yet, realising these benefits requires policies that invest in the necessary infrastructure, such as reliable data collection systems, secure communication networks, and accessible computational resources. Ethical and regulatory considerations are critical in shaping AI's role in disaster management. Policies must address concerns around data privacy, security, and the potential for algorithmic bias, ensuring that AI‐driven decisions are transparent, equitable, and accountable. Establishing standards for the auditing and validation of AI systems is essential, particularly in high‐stakes scenarios where biases could exacerbate vulnerabilities or exclude marginalised populations. Furthermore, AI systems rely heavily on high‐quality, unbiased data, making data governance policies a priority. These should include guidelines for data collection, sharing, and usage that protect individual rights while promoting collaborative disaster management efforts across agencies and jurisdictions. Interdisciplinary collaboration is another vital policy consideration. AI disaster management systems are most effective when they integrate the expertise of technologists, emergency managers, and local community leaders. Policies should facilitate this collaboration by creating frameworks that align AI applications with the specific needs and priorities of affected communities. For example, hazard mapping tools and evacuation planning systems must consider local cultural, logistical, and socio‐economic factors to ensure their effectiveness and inclusivity. Workforce development is equally important in embedding AI in disaster management policies. Emergency management professionals require training to understand and apply AI insights effectively, including interpreting predictive models, integrating AI tools into decision‐making processes, and addressing ethical concerns. Policies should support the creation of educational programmes and professional development opportunities that build this capacity. Lastly, policies must prioritise equitable access to AI technologies to prevent further disparities between resource‐rich and underserved regions. Investment in infrastructure and capacity building for under‐resourced communities can ensure that AI‐related benefits are distributed fairly, enhancing disaster resilience on a broader scale. By addressing these different dimensions, policies can harness AI's potential to revolutionise disaster management, aligning technological advancements with ethical considerations and societal needs while promoting resilience, inclusivity, and equity in disaster response and recovery efforts. Without doubt, these endeavours are generating teething problems in their respective fields, but they are also advancing inquiry meaningfully beyond simple practical adoption. The question then becomes: why does EM lack a multifaceted view of AI relative to other, prior exogenous developments (such as climate change or pandemic threats)? To answer this question, we must of course answer another one: what conditions define the readiness of a field of social science inquiry to orient on disruptive developments, to organise clear research programmes of investigation, and then to translate scholarly know‐how into practice? 4. THE LESSONS OF A PAST SHOCK: THE CASE OF CYBERSPACE AND PUBLIC AFFAIRS To address the EM field's readiness to grapple with the complex issues surrounding AI development, adoption, and societal transformation, we consider the evolution of thinking on a previous exogenous shock with characteristics similar to that of AI: the rise of the internet. From a point of relative disinterest in the internet just three decades ago, scholars have nurtured a burgeoning subfield of political science that, today, is defined by competing theoretical perspectives and established assumptions. 4.1. The internet: why then and why not? It is easy to see why parallels are often drawn between AI and the internet. Both information revolutions are dynamic insofar as they represent the development of general‐purpose technologies around which multifaceted tools might be built (Naughton, 2016 ). The rise of internetworks following the creation of the ARPANET (Advanced Research Projects Agency Network)—the forebear of all modern networks—in 1967 was rapid, with entire computerised industries in the US and Europe moving online by the 1980s. By 1994, when the move to privatise the core networks of the web was formalised by the US Congress, tens of millions of individuals were interfacing with each other via computers every year—a number that increased rapidly (Naughton, 2012 ). The growth of the internet and the development of key underlying systems like the HyperText Transfer Protocol that underwrite the World Wide Web are well‐told tales (Leiner et al., 2009 ). What is less often understood in modern view is the relatively fragmented nature of the scholarly focus on the implications of the internet for traditional research and professional paradigms. In the social sciences, robust work in limited areas like communication studies and sociology often masks the reality that much scholarship prognosticated on the impact of the web to alter traditional conditions of human interaction and cognition (Dunn Cavelty and Wenger, 2020 ). In the field of political science and its subfield focus areas (such as public administration and international relations), the 1990s produced no distinct subfield research programmes on the internet. Today, we can point to discrete areas of research with clear assumptions and theoretical perspectives, such as the programmes on cyber repression, cyberpsychology, disinformation, and propaganda studies (Dunn Cavelty, 2018 ; Gorwa and Smeets, 2019 ). And yet, in the heyday of the internet's early development during the dotcom boom and onwards, there was precious little coordination of web‐oriented research. What changed? 4.2. The domain concept as both chicken and egg As a series of recent exploratory publications suggest, researchers in political science and international relations can trace the contours of what is now a burgeoning, diverse programme on digital issues to around the start of the 2010s (Whyte, 2023a ). Prior to 2011, almost no major political science journals published work specifically centred on the relevance of the internet for political phenomena; there are a handful of exceptions largely in the realm of political theory (see, for example, Manjikian, 2010 ). Most significantly, the early focus on the internet as a transformative dimension of political phenomena was in no way reflective of the shape of the international relations, strategic studies, or foreign policy fields. Through the mid‐2010s, however, a number of notable articles began to bring studies of the internet into the mainstream view of international security, development, and other public affairs researchers. The common strand linking these efforts was not some real‐world incident (such as Russian interference in Western elections) as is sometimes claimed, but rather an idea that gave researchers analytic purchase for engaging with pre‐existing concepts in international relations, strategic studies, innovation studies, and more: that of cyberspace as a distinct domain of human interaction (Branch, 2021 ). This foundational metaphor, as Branch ( 2021 ) calls it, offered clarity about the phenomenon for traditionalist researchers despite, ironically, the overtly artificial and overly simplistic nature of the idea. The concept of cyberspace as we know it today, if not the actual word, emerged in the 1990s within the defence community of the US (Whyte and Mazanec, 2023 ). Although networked computer systems had exploded in popularity in certain strata of Western society in the 1970s and 1980s, it was a series of military deployments from the First Gulf War (1990–91) onwards that brought to the fore questions about exactly how this rapidly advancing medium for communication—the internet—should fit into government planning and budget prioritisation (Nakayama, 2022a ). There were a range of questions that needed answers, from what kinds of digital capabilities fit best with 1990s ideas of success among diverse US institutions to what relationships with private enterprise were desirable to buildout infrastructure to support national security missions. As is so often true of complex bureaucracies with entrenched intra‐service rivalries and politicking, the answers to these questions did not emerge via any kind of consensus‐building debate or even top‐down imperatives from political leadership. Rather, they materialised from a small constellation of individuals who found themselves in the right position to see the potential of network technologies for more than just communication and then to act to influence the establishment view of the issue (Whyte, 2023a ). As it happened, these people were mostly US Air Force officers like Kenneth Minihan, Mike McConnell, and Michael Hayden who occupied important positions in the 1990s, before going on to become directors of the National Security Agency, the Central Intelligence Agency, and the Office of the Director of National Intelligence (Lindsay, 2021 ). These officers, who saw the inherent value in the military using the internet to do more than communicate—that is, to launch and deter network activities that could disrupt adversaries—gradually influenced strategic planning and doctrine to standardise the idea that network operations could be more a use of force analogue than a range of informational activities. Owing to the success of their efforts, this exceptional view of what could only now be called a distinct space of operation became the prevailing vision of cybersecurity that drove the development of new cyber‐warfighting units, missions, and, ultimately, commands (in the form of United States Cyber Command). And this view rapidly spread, popular particularly because it offered a framework for thinking about digital insecurity as something only mildly entangled with existing defence missions. 4.3. Resolving uncertainty The rise of the domain conceptualisation of cyberspace had clear value to Western defence establishments. Above anything else, the exceptionalism of cyberspace has justified the creation of discrete cyber‐conflict institutions across the developed and developing worlds. These extend from the standing up of United States Cyber Command at the end of the 2000s to, as Blessing ( 2021 ) illustrates, the development of several dozen nascent cybersecurity defence forces around the world less than a decade later. There is little doubt that the domain concept has enabled this and has thus speedily made global visibility of digital threats and web technology issues a matter of high politics, rather than just a function of underling technology management. The domain concept serves as a unifying framework for organising not only government activities but also academic thinking around the impact of the internet on traditional areas of scholarly focus—as evidenced in early writings by the likes of Gartzke ( 2013 ) and Kello ( 2013 ). Today, there is a diverse array of research programmes in the fields of political science and public affairs that are far from just simplistic investigations of the technological impact on narrow real‐world conditions. The research programme on the internet and national security operations, for instance, is five generations of thought into its existence, having swiftly evolved from early concern about a ‘cyber Pearl Harbor’, through the adoption and testing of Cold War concepts of deterrence and coercion, and ultimately to contemporary multifaceted debates on the nature of digitised competition: a glorified intelligence contest (Chesney and Smeets, 2023 ) versus reflecting a state's strategic preferences (Fischerkeller, Goldman, and Harknett, 2022 ). But what did the cyberspace framework do for scholars interested in the internet and politics? Like others, we argue that it helped to resolve uncertainty about the implications of a broad exogenous technological shock for discrete fields of study. Political science broadly defines uncertainty along several lines (Rathbun, 2007 ). Realist perspectives on the human condition see it in terms of fear of the unknown. Rationalists see it as ignorance endemic to lacking information about possible outcomes. Cognitively‐minded researchers view uncertainty as confusion or dissonance while constructivists see it as indeterminacy, defined largely by missing or ambiguous normative conditions. The domain concept cuts through uncertainty about the impact of the internet on each of these fronts, establishing the idea of a common space within which different interpretations of uncertainty can be contained, dissected, and assessed. Notably, the domain concept is a framework that points researchers towards established views of domain‐specific human interaction, encouraging thinking about a logic of digital engagement that parallels that which can be found elsewhere. This can then be grafted on top of conventional assumptions and concepts of political behaviour to modify them and allow for fruitful conversation about the digital version of a phenomenon. It is important to point out, however, that there is a notable shortcoming of cyberspace as an organising framework for scholarship and practice relevant to the exogenous shock that is the internet. As a range of work has documented in just the past half decade, the artificial creation of the idea of a space within which use of force analogues are possible biased both research and practice in the US towards overly‐militaristic views of the utility of the technology. This posture has been roundly blamed for an inability on the part of scholars and policymakers to foresee the rise of alternative forms of digital threats to Western society, particularly influence campaigns, social subversion, and counter‐society hacking by foreign state proxies that falls beyond the purview of traditional interstate security frameworks (Nakayama, 2022b ; Whyte, 2023a ). The result has been a swift reassessment of the core assumptions of the cyber‐conflict studies subfield in political science since 2016 and exploration of alternative models of political behaviour in the age of the internet. Cyberspace, in other words, served its purpose as an organising framework for political science and public affairs research but incentivised thinking that was ultimately narrower than, in hindsight, should have been desired (Whyte, 2023a ). 5. THE TITAS FRAMEWORK: HOW TO ORGANISE EM EFFORTS ON AI RESEARCH AND PRACTICE The case of the domain concept and the internet suggests that fields of study in the social sciences adapt to an exogenous shock given appropriate frameworks for organising knowledge and interaction of the details of the novel phenomena with the underlying assumptions of their theoretical and empirical programmes. The most effective frameworks are likely those that reconcile the most uncertainty—whether it be concern about disruptive effects, lack of information, unclear conditions, ambiguous normative circumstances, inter alia—by bridging the character of an emergent phenomenon with core field assumptions about human behaviour. As the cyberspace example illustrates, though, the trick is in adopting a framework for organising knowledge‐building efforts that interacts with existing field theories and missions without supplanting any of them. At present, we see no obvious EM framework which should holistically guide the focus on AI, 4 in no small part (quite understandably) because subfield focus areas like EM are so often dominated by narrow models that describe narrow activities or more conventional macro‐theoretical distinctions in the social sciences (rationalist versus constructivist, for instance). Consequently, we suggest a framework for organising EM efforts to problematise AI, understand its impact on society, model its utility across EM issue areas, and adapt core assumptions for a phenomenon that is certain to evolve over time. We label this the TITAS Framework, which, as noted, stands for the Tool/Intermediary/Talent pool/Attack Surface Framework. Our goal in developing this for knowledge‐producing efforts around AI was principally to replicate the organisational effects of a framework like the cyberspace domain concept without oversimplifying the nature of AI's transformative effects on human society and the EM mission. The TITAS Framework thus reflects the technical characteristics of AI technologies as a multidimensional phenomenon with a clear relation to elements of the EM enterprise. Most importantly, the framework we suggest is loosely coupled. By this, we mean that it parses AI's impact on the EM field (and related areas, such as the homeland security subfield) into distinct domains of transformation and presentation without adopting a holistic axiom about what AI is . In other words, we see no cyberspace corollary for AI, in part because we do see danger in suggesting any unitary conceptualisation of machine intelligence. An organising framework must act as a bridge that facilitates conversation about the objective attributes of AI and the diverse underlying uncertainties that characterise knowledge‐creation efforts in the EM field. It must be ontologically relevant to the field rather than the emergent variable, the technology itself. By suggesting a four‐part typology, as set out earlier, we accomplish this while avoiding a situation in which theoretical developments invalidate other lines of thought: AI as a tool describes the most oft‐cited dimension of conceptual and empirical focus on machine intelligence for EM, namely that of AI as a toolkit or set of applications that stand to upgrade or otherwise augment processes of disaster mitigation, response, recovery, and more. AI as an intermediary , by contrast, describes EM functions, systems, or interlocutors that are the result of AI agency absent direct human inputs. AI as a talent pool describes AI priors, including training data, algorithmic development and related input processes, supply chain functionality, and more, all of which functions as a set of latent resources distinct from the more immediate manifestations of AI in the first two categories. AI as an attack surface describes the process by which EM issue areas both within and without apparatuses of hazard engagement are shaped by AI and which recursively shape EM in turn. It is important to note that these typological areas are inherently linked and that these linkages are the key to drawing accurate lines around emergent areas of research in EM with AI. To illustrate the utility in using this form of differentiating framework to guide thinking, we might consider a broad empirical prompt relevant to the EM field and the tasks involved in fleshing out a programme of multifaceted narrow questions closer to the point of clear research design. For instance, how might EM professionals expect algorithmic bias to impact existing hazard response missions? Here, we might think about bias in the EM mission profile existing along several distinct lines. It might emerge from the pressure on the part of an operator or decision‐maker to interact with a new AI technique embodied in a software interface, a physical tool, or information fed into the decision‐making loop. As a broad literature has already illustrated, a range of dispositional and experiential factors shape users' trust in technology and their willingness to deploy novel technological solutions in place of legacy alternatives (Whyte, 2023b ). Likewise, these cognitive or affective tendencies can be altered by institutional structure or the perception of human inputs to a technologically‐aided process. People have, for example, been shown to trust technology more when there is an apparent human validation of technological outcomes in the decision‐making loop (Whyte, 2023b ). Here, bias may be generated around the use of AI systems entirely from the socio‐cultural, political, and institutional contexts within which they are embedded. Consequently, there exists the probability that these deviations from objective baselines might be fed back into underlying models to impact the autonomous activities of AI models. Particularly where AI models are undertaking roles traditionally overseen by a human, such as checking red flags during a cyber‐intrusion event or authorising overtime for responders during a crisis, this produces the possibility of bias leading to disruption or inequity entirely due to AI agency. Naturally, these possible pathological outcomes interact with potential existing bias in the design of initial AI tools, given that the latent resources on which we rely to develop novel applications (that is, the talent pool of AI data, data curation, integration processes, inter alia) are prone to both historical misrepresentations of reality and the same inherent complexity of active management conditions that affect the use of the tools themselves. Lastly, furthermore, all of this will occur—or is already occurring—in the context of AI's shaping of society in its current form. The continued growth of the Internet of Things, the automation of global economic and related social processes, and more all portend changes in the fundamentals of citizen engagement with the hazards that threaten society. In short, AI will alter the landscape of response that EM professionals must address. Here, bias may impact the EM enterprise via a more diffuse influence on its priors, helping drive public opinion and elite preference in ways that (like major exogenous shocks in the past for the most part, such as the 11 September 2001 attacks or the COVID‐19 pandemic) preferentially boost certain mission priorities over those that are objectively similar in potential impact. More formally, we present the utility of the TITAS Framework as an ability to encode a set of analytic assumptions about what kinds of knowledge EM requires for what to organise, when, and for what purpose(s), in addition to its simpler use for operational mapping previewed in Table 1 . Table 2 makes that planning utility explicit by translating each phase into (i) the dominant TITAS elements likely to shape risk and readiness, (ii) the governance and coordination problems that follow, and (iii) the types of questions agencies should treat as priority planning prompts. TABLE 2. TITAS Framework for knowledge organisation and strategic planning. Knowledge/strategic planning by TITAS prioritisation Dominant TITAS elements Driving applications/research questions Governance/risk imperatives EM phase Mitigation Tool, talent pool (with attack surface as a persistent background element) How do AI tools alter risk identification and mitigation targeting? What ‘priors’ (datasets, vendor ecosystems) lock in future mitigation capacity? Hidden bias in priors; brittle assumptions in long‐horizon models; political/communication manipulation shaping mitigation priorities. Preparedness Talent pool, intermediary What data‐sharing and QA regimes are needed for ‘centaur’ planning and exercising? Where should AI agents be allowed to mediate coordination and alerting? Procurement/supply chain fragility; insufficient testing/red teaming; over‐automation of escalation and warning pathways. Response Tool, intermediary (with attack surface becoming acute only under time pressure) Where can AI decision support speed up operations without displacing human judgement? Which response functions are most vulnerable to agentic automation errors? Adversarial manipulation of inputs; automation bias; misinformation/rumour operations exploiting AI‐enabled communication and decision loops. Recovery Tool, talent pool (with attack surface narrowly focused on fraud/trust and institutional legitimacy) How should AI support benefits triage, case management, and rebuilding prioritisation while preserving transparency and equity? How do recovery datasets feed back into future models? AI‐amplified fraud/scams; inequitable automation at scale; long‐tail legitimacy/trust problems from opaque AI decisions. Open in a new tab Note: QA = quality assurance. Source: authors. Table 2 should be read as a practical heuristic for preparedness as organisation. It clarifies why ‘AI readiness’ cannot be reduced to tool adoption, that is, because the central tasks are often upstream (such as building shared vocabularies, data and procurement standards, validation routines, training pipelines, and escalation rules for human–AI teaming) of those time pressures that collapse deliberation. It also highlights the persistent tension that makes centaur capacity hard in practice—in essence, the same intermediating systems and talent pools that enable speed and scale can also introduce new failure cascades and expand the attack surface, especially under response conditions when verification and accountability are hardest to maintain. From this vantage point, preparedness is best understood as designing the socio‐technical arrangements that keep AI useful and governable across the full lifecycle. A framework like TITAS is a critical pivot for this act of design. 6. AI AND INSIGHTS FROM THE WEB: HOW TO BUILD EM CENTAURS For any discipline, the task of cultivating a workforce of responders, researchers, and planners that effectively blends the intrinsic strengths of both machine and human being is a daunting one. For a practice‐centred subfield like EM, the challenge is greater still. After all, the emergence of AI demands that we simultaneously observe trends at the lowest level of analyses (that is, at the level of individuals, their communities, and our organisations) while accounting for rapid and prospectively unexpected transformations at the highest level (that is, our society, our climate, the global economy, and world politics). AI is transforming EM policies by enabling data‐driven decision making and proactive disaster preparedness strategies. AI‐powered tools, such as supervised models, neural networks, and reinforcement learning algorithms, process vast datasets—including real‐time weather updates, social media feeds, and satellite imagery—to predict hazards, assess vulnerabilities, and develop mitigation strategies. For practitioners, supervised models have been utilised to create hazard susceptibility maps for landslides and avalanches, while convolutional neural networks are applied to flood zone prediction and early storm detection. These advancements allow policymakers to implement more targeted instruments, such as hazard‐specific building codes, enhanced evacuation protocols, and optimised resource allocation plans. Predictive modelling powered by AI fosters dynamic, adaptable policies that address evolving threats like climate change and urbanisation. Yet, the integration of AI into policy frameworks requires overcoming challenges such as data quality issues, the computational intensity of AI systems, and concerns around data privacy and equity. Embedding AI in EM policies can not only enhance disaster resilience but also reduce response times and mitigate socio‐economic impacts, fundamentally reshaping how communities prepare for and recover from disasters. This article has undertaken two tasks. First, it has provided context and insight with respect to the current push to think more seriously about AI. In particular, we have offered a view of AI as a phenomenon with a clear logic of evolution, both in decades past and in recent years. We suggest that this paper's opening sections are a resource for general thinking about AI that presently does not exist in the EM field. Second, and arguably more importantly, the article has undertaken the task of analytically gauging EM's readiness for grappling with the implications of an exogenous shock like the emergence of AI. Our insights drawn from the example of the internet and the rise of ‘cyberspace’, now the preeminent concept driving much thinking about the web among social scientists, suggest clear opportunities and pitfalls with respect to any attempt to apply legacy frameworks to novel phenomena like AI. We argue that bridging frameworks are required for organising knowledge‐building efforts, meaning those that interact with existing field theories and missions without supplanting any of them. Our TITAS Framework, which is representative of such a bridging framework, illustrates the potential value of such an approach, allowing researchers and practitioners alike to link conceptual issues surrounding AI with the extended implications of AI development, deployment, societal integration, and societal contextualisation at the point of contact. Clearly, there is value in such an approach, which the EM field would be well advised to harness as it moves forward into the age of AI. Practically, as Table 2 suggests, the utility of the TITAS Framework is not only classificatory or research‐oriented, but is also a way to discipline how EM organisations sequence knowledge building and strategic planning across mitigation, preparedness, response, and recovery. Looking forward, the broad implication of this is the need for bounded experimentation rather than wholesale adoption. Mirroring what is already under way in more specialised national security settings, this means piloting tool use cases in auditable workflows with explicit human sign‐off, while setting boundary conditions and cut‐off rules in areas where AI might act as an intermediary. By contrast, the harder, longer‐run work exists in the talent pool space, where the development of data governance standards, validation pipelines, or procurement requirements that preserve documentation and update transparency reflects choices that lock in or otherwise foreclose future capacity. And finally, because AI inevitably expands attack surfaces, the intrinsic structure of the TITAS Framework implies that agencies must normalise adversarial testing and exercises that stress AI‐enabled decision loops quickly or otherwise face an extending tail of trust erosion challenges for response and recovery missions. CONFLICT OF INTEREST STATEMENT The authors declare no conflicts of interest. Endnotes 1 We use the term exogenous shock here in line with social science inquiry on ‘outside of context’ disruptions that come to redefine areas of industrial, societal, academic, scientific, or other function. See, for instance, Savun and Tirone ( 2012 ) and Miklian and Hoelscher ( 2022 ). 2 These challenges are already taken up in other fields of study. See, for example, Collins et al. ( 2021 ). 3 For perhaps the best overview of the development of AI from antiquity through the modern era, see Nilsson ( 2010 ). 4 Despite work on developing frameworks being of clear significance. See, for example, Pearson and Mitroff ( 2018 ). Contributor Information Christopher Whyte, Email: [email protected]. Brittany ‘Brie’ Haupt, Email: [email protected]. DATA AVAILABILITY STATEMENT Data sharing not applicable to this article as no datasets were generated or analysed during the current study. REFERENCES Al‐Hashimi, H.M. (2023) ‘Turing, von Neumann, and the computational architecture of biological machines’. Proceedings of the National Academy of Sciences of the United States of America. 120(25). Article number: e2220022120. 10.1073/pnas.2220022120. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Al‐Qizwini, M. , Barjasteh I., Al‐Qassab H., and Radha H. (2017) ‘Deep learning algorithm for autonomous driving using GoogLeNet’. In IEEE (Institute of Electrical and Electronics Engineers) (ed.) 28th IEEE Intelligent Vehicles Symposium: June 11–14, 2017, Redondo Beach, California, USA. IEEE Publications, Piscataway, NJ. pp. 89–96. [ Google Scholar ] Alom, M.Z. et al. (2018) ‘The history began from AlexNet: a comprehensive survey on deep learning approaches’. arXiv:1803.01164 . September. 10.48550/arXiv.1803.01164. [ DOI ] Baum, K. et al. (2023) ‘From fear to action: AI governance and opportunities for all’. Frontiers in Computer Science. 5(May). Article number: 1210421. 10.3389/fcomp.2023.1210421. [ DOI ] [ Google Scholar ] Blessing, J. (2021) ‘The global spread of cyber forces, 2000–2018’. In Jančárková T. et al. (eds.) 2021 13th International Conference on Cyber Conflict: Going Viral. NATO (North Atlantic Treaty Organisation) Cooperative Cyber Defence Centre of Excellence, Tallinn. pp. 233–255. [ Google Scholar ] Branch, J. (2021) ‘What's in a name? Metaphors and cybersecurity’. International Organization. 75(1). pp. 39–70. [ Google Scholar ] Buchanan, B. and Imbrie A. (2022) The New Fire: War, Peace, and Democracy in the Age of AI. MIT Press, Cambridge, MA. [ Google Scholar ] Chen, J.X. (2016) ‘The evolution of computing: AlphaGo’. Computing in Science & Engineering. 18(4). pp. 4–7. [ Google Scholar ] Chen, N. , Liu W., Bai R., and Chen A. (2019) ‘Application of computational intelligence technologies in emergency management: a literature review’. Artificial Intelligence Review. 52(3). pp. 2131–2168. [ Google Scholar ] Chesney, R. and Smeets M. (2023) Deter, Disrupt, or Deceive: Assessing Cyber Conflict as an Intelligence Contest. Georgetown University Press, Washington, DC. [ Google Scholar ] Collins, C. , Dennehy D., Conboy K., and Mikalef P. (2021) ‘Artificial intelligence in information systems research: a systematic literature review and research agenda’. International Journal of Information Management. 60(October). Article number: 102383. 10.1016/j.ijinfomgt.2021.102383. [ DOI ] [ Google Scholar ] Comfort, L.K. (2007) ‘Crisis management in hindsight: cognition, communication, coordination, and control’. Public Administration Review. 67(S1). pp. 189–197. [ Google Scholar ] Comfort, L.K. and Rhodes M.L. (2022) ‘Collective cognition in complex systems’. In Comfort L.K. and Rhodes M.L. (eds.) Global Risk Management: The Role of Collective Cognition in Response to COVID‐19. Routledge, Abingdon. pp. 1–21. [ Google Scholar ] Creswell, A. et al. (2018) ‘Generative adversarial networks: an overview’. IEEE Signal Processing Magazine. 35(1). pp. 53–65. [ Google Scholar ] Dunn Cavelty, M. (2018) ‘Cybersecurity research meets science and technology studies’. Politics and Governance. 6(2). pp. 22–30. [ Google Scholar ] Dunn Cavelty, M. and Wenger A. (2020) ‘Cyber security meets security politics: complex technology, fragmented politics, and networked science’. Contemporary Security Policy. 41(1). pp. 5–32. [ Google Scholar ] El Samad, M. , Nasserddine G., and Kheir A. (2023) ‘Introduction to artificial intelligence’. In Hemachandran K. et al. (eds.) Artificial Intelligence and Knowledge Processing: Improved Decision‐Making and Prediction. CRC Press, Boca Raton, FL. pp. 1–14. [ Google Scholar ] Fischerkeller, M.P. , Goldman E.O., and Harknett R.J. (2022) Cyber Persistence Theory: Redefining National Security in Cyberspace. Oxford University Press, Oxford. [ Google Scholar ] Floridi, L. (2020) ‘AI and its new winter: from myths to realities’. Philosophy & Technology. 33(1). pp. 1–3. [ Google Scholar ] Fotovatikhah, F. et al. (2018) ‘Survey of computational intelligence as basis to big food management: challenges, research directions and future work’. Engineering Applications of Computational Fluid Mechanics. 12(1). pp. 411–437. [ Google Scholar ] French, R.M. (2000) ‘The Turing Test: the first 50 years’. Trends in Cognitive Sciences. 4(3). pp. 115–122. [ DOI ] [ PubMed ] [ Google Scholar ] Gartzke, E. (2013) ‘The myth of cyberwar: bringing war in cyberspace back down to earth’. International Security. 38(2). pp. 41–73. [ Google Scholar ] Goldfarb, A. and Lindsay J.R. (2022) ‘Prediction and judgment: why artificial intelligence increases the importance of humans in war’. International Security. 46(3). pp. 7–50. [ Google Scholar ] Gonsalves, T. (2019) ‘The summers and winters of artificial intelligence’. In Khosrow‐Pour M. (ed.) Advanced Methodologies and Technologies in Artificial Intelligence, Computer Simulation, and Human–Computer Interaction. IGI Global Scientific Publishing, Hershey, PA. pp. 168–179. [ Google Scholar ] Gorwa, R. and Smeets M. (2019) ‘Cyber Conflict in Political Science: A Review of Methods and Literature’. Working Paper prepared for the ‘ISA (International Studies Association) 60th Annual Convention’, 27–30 March 2019, Toronto, ON, Canada. 10.31235/osf.io/fc6sg. [ DOI ] [ Google Scholar ] Gupta, S. , Modgil S., Kumar A., Sivarajah U., and Irani Z. (2022) ‘Artificial intelligence and cloud‐based collaborative platforms for managing disaster, extreme weather and emergency operations’. International Journal of Production Economics. 254(December). Article number: 108642. 10.1016/j.ijpe.2022.108642. [ DOI ] [ Google Scholar ] Hashemi, M.R. , Spaulding M.L., Shaw A., Farhadi H., and Lewis M. (2016) ‘An efficient artificial intelligence model for prediction of tropical storm surge’. Natural Hazards. 82(1). pp. 471–491. [ Google Scholar ] Hsu, F‐h ., Anantharaman T., Campbell M., and Nowatzyk A. (1990) ‘A grandmaster chess machine’. Scientific American. 263(4). pp. 44–51. [ Google Scholar ] Huang, D. , Wang S., and Liu Z. (2021) ‘A systematic review of prediction methods for emergency management’. International Journal of Disaster Risk Reduction. 62(August). Article number: 102412. 10.1016/j.ijdrr.2021.102412. [ DOI ] [ Google Scholar ] Iavazzo, C. , Gkegke X‐E.D., Iavazzo P‐E., and Gkegkes I.D. (2014) ‘Evolution of robots throughout history from Hephaestus to Da Vinci robot’. Acta Medico‐Historica Adriatica. 12(2). pp. 247–258. [ PubMed ] [ Google Scholar ] Ilievski, A. , Zdraveski V., and Gusev M. (2018) ‘How CUDA powers the machine learning revolution’. In IEEE (Institute of Electrical and Electronics Engineers) (ed.) Proceedings of Papers: 2018 26th Telecommunications Forum (TELFOR), Belgrade, Serbia, November, 20–21, 2018. IEEE Publications, Piscataway, NJ. pp. 420–425. [ Google Scholar ] Janis, I.L. and Mann L. (1977) ‘Emergency decision making: a theoretical analysis of responses to disaster warnings’. Journal of Human Stress. 3(2). pp. 35–48. [ DOI ] [ PubMed ] [ Google Scholar ] Jensen, B.M. , Whyte C., and Cuomo S. (2020) ‘Algorithms at war: the promise, peril, and limits of artificial intelligence’. International Studies Review. 22(3). pp. 526–550. [ Google Scholar ] Johnson, J. (2019) ‘Artificial intelligence & future warfare: implications for international security’. Defense & Security Analysis. 35(2). pp. 147–169. [ Google Scholar ] Kello, L. (2013) ‘The meaning of the cyber revolution: perils to theory and statecraft’. International Security. 38(2). pp. 7–40. [ Google Scholar ] Kramer, S. (1996) ‘Mind, symbolism, formalism: is Leibniz a precursor of artificial intelligence?’. Knowledge Organization (KO). 23(2). pp. 83–87. [ Google Scholar ] Leiner, B.M. et al. (2009) ‘A brief history of the internet’. ACM SIGCOMM Computer Communication Review. 39(5). pp. 22–31. [ Google Scholar ] Lindsay, J.R. (2021) ‘Cyber conflict vs. Cyber Command: hidden dangers in the American military solution to a large‐scale intelligence problem’. Intelligence and National Security. 36(2). pp. 260–278. [ Google Scholar ] Manjikian, M.M. (2010) ‘From global village to virtual battlespace: the colonizing of the internet and the extension of realpolitik’. International Studies Quarterly. 54(2). pp. 381–401. [ Google Scholar ] McCarthy, J. , Minsky M.L., Rochester N., and Shannon C.E. (2006) ‘A proposal for the Dartmouth summer research project on artificial intelligence, August 31, 1955’. AI Magazine. 27(4). pp. 12–14. [ Google Scholar ] Miklian, J. and Hoelscher K. (2022) ‘SMEs and exogenous shocks: a conceptual literature review and forward research agenda’. International Small Business Journal: Researching Entrepreneurship. 40(2). pp. 178–204. [ Google Scholar ] Moor, J. (2006) ‘The Dartmouth College artificial intelligence conference: the next fifty years’. AI Magazine. 27(4). pp. 87–91. [ Google Scholar ] Morozov, E. (2023) ‘The true threat of artificial intelligence’. The New York Times website. 30 June. https://www.nytimes.com/2023/06/30/opinion/artificial-intelligence-danger.html (last accessed on 8 April 2026). [ Google Scholar ] Nakayama, B. (2022a) ‘Information vs the cyberspace domain’. Journal of Cyber Policy. 7(2). pp. 213–229. [ Google Scholar ] Nakayama, B. (2022b) ‘Democracies and the future of offensive (cyber‐enabled) information operations’. The Cyber Defense Review. 7(3). pp. 49–65. [ Google Scholar ] Naughton, J. (2012) From Gutenberg to Zuckerberg: What You Really Need to Know About the Internet. Quercus, London. [ Google Scholar ] Naughton, J. (2016) ‘The evolution of the internet: from military experiment to general purpose technology’. Journal of Cyber Policy. 1(1). pp. 5–28. [ Google Scholar ] Newborn, M. (2012) Kasparov Versus Deep Blue: Computer Chess Comes of Age. Springer Science & Business Media, New York City, NY. [ Google Scholar ] Nilsson, N.J. (2010) The Quest for Artificial Intelligence: A History of Ideas and Achievements. Cambridge University Press, New York City, NY. [ Google Scholar ] Patterson, D. (1990) Introduction to Artificial Intelligence and Expert Systems. Prentice‐Hall, Upper Saddle River, NJ. [ Google Scholar ] Pearson, C.M. and Mitroff I.I. (2018) ‘From crisis prone to crisis prepared: a framework for crisis management’. In Mars G. and Weir D.T.H. (eds.) Risk Management: Volume II: Management and Control. Routledge, Abingdon. pp. 185–196. [ Google Scholar ] Photis, Y.N. and Grekousis G. (2012) ‘Locational planning for emergency management and response: an artificial intelligence approach’. International Journal of Sustainable Development and Planning. 7(3). pp. 372–384. [ Google Scholar ] Qerimi, Q. (2020) ‘Imagination, invention and internet: from Aristotle to artificial intelligence and the “post‐human” development and ethics’. In Popkova E.G. and Sergi B.S. (eds.) The 21st Century from the Positions of Modern Science: Intellectual, Digital and Innovative Aspects. Springer International Publishing, Cham. pp. 360–371. [ Google Scholar ] Rathbun, B.C. (2007) ‘Uncertain about uncertainty: understanding the multiple meanings of a crucial concept in international relations theory’. International Studies Quarterly. 51(3). pp. 533–557. [ Google Scholar ] Ruff, K.M. and Pappu R.V. (2021) ‘AlphaFold and implications for intrinsically disordered proteins’. Journal of Molecular Biology. 433(20). Article number: 167208. 10.1016/j.jmb.2021.167208. [ DOI ] [ PubMed ] [ Google Scholar ] Russell, S.J. and Norvig P. (2022) Artificial Intelligence: A Modern Approach. Pearson Education Limited, Harlow. [ Google Scholar ] Savun, B. and Tirone D.C. (2012) ‘Exogenous shocks, foreign aid, and civil war’. International Organization. 66(3). pp. 363–393. [ Google Scholar ] Sharma, S. et al. (2019) ‘Artificial intelligence agents for crowd simulation in an immersive environment for emergency response’. IS&T International Symposium on Electronic Imaging. 31(2). pp. 176‐1–176‐8. [ Google Scholar ] Struckman, I. and Kupiec S. (2023) ‘Why they're worried: examining experts’ motivations for signing the “pause letter”’. arXiv:2306.00891 . June. 10.48550/arXiv.2306.00891. [ DOI ] Sun, W. , Bocchini P., and Davison B.D. (2020) ‘Applications of artificial intelligence for disaster management’. Natural Hazards. 103(3). pp. 2631–2689. [ Google Scholar ] Tan, L. , Guo J., Mohanarajah S., and Zhou K. (2021) ‘Can we detect trends in natural disaster management with artificial intelligence? A review of modeling practices’. Natural Hazards. 107(3). pp. 2389–2417. [ Google Scholar ] Toosi, A. , Bottino A.G., Saboury B., Siegel E., and Rahmim A. (2021) ‘A brief history of AI: how to prevent another winter (a critical review)’. PET Clinics. 16(4). pp. 449–469. [ DOI ] [ PubMed ] [ Google Scholar ] Vinyals, O. et al. (2017) ‘StarCraft II: a new challenge for reinforcement learning’. arXiv:1708.04782 . August. 10.48550/arXiv.1708.04782. [ DOI ] Whyte, C. (2020) ‘Deepfake news: AI‐enabled disinformation as a multi‐level public policy challenge’. Journal of Cyber Policy. 5(2). pp. 199–217. [ Google Scholar ] Whyte, C. (2023a) ‘Cyberspace is what states make of it: the social construction (and deconstruction) of strategic concepts’. Comparative Strategy. 42(1). pp. 128–151. [ Google Scholar ] Whyte, C. (2023b) ‘Learning to trust Skynet: interfacing with artificial intelligence in cyberspace’. Contemporary Security Policy. 44(2). pp. 308–344. [ Google Scholar ] Whyte, C. and Mazanec B. (2023) Understanding Cyber Warfare: Politics, Policy and Strategy . Routledge, Abingdon. [ Google Scholar ] Wiener, N. (2019) Cybernetics or Control and Communication in the Animal and the Machine. MIT Press, Cambridge, MA. [ Google Scholar ] Wu, T. et al. (2023) ‘A brief overview of ChatGPT: the history, status quo and potential future development’. IEEE/CAA Journal of Automatica Sinica. 10(5). pp. 1122–1136. [ Google Scholar ] Yuan, C. , Zhang Y., and Liu Z. (2015) ‘A survey on technologies for automatic forest fire monitoring, detection, and fighting using unmanned aerial vehicles and remote sensing techniques’. Canadian Journal of Forest Research. 45(7). pp. 783–792. [ Google Scholar ] Zarghami, S.A. and Dumrak J. (2021) ‘Implications of artificial intelligence for bushfire management’. Australian Journal of Emergency Management. 36(1). pp. 84–91. [ Google Scholar ] Zhang, F. , Qiao Q., Wang J., and Liu P. (2022) ‘Data‐driven AI emergency planning in process industry’. Journal of Loss Prevention in the Process Industries. 76(May). Article number: 104740. 10.1016/j.jlp.2022.104740. [ DOI ] [ Google Scholar ] Zhang, L. , Lv X., and Dhakal S. (2019) ‘A reinforcement learning‐based stakeholder value aggregation model for collaborative decision making on disaster resilience’. In Cho Y.K. et al. (eds.) Computing in Civil Engineering 2019: Smart Cities, Sustainability, and Resilience. American Society of Civil Engineers, Reston, VA. pp. 490–497. [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement Data sharing not applicable to this article as no datasets were generated or analysed during the current study. Articles from Disasters are provided here courtesy of Wiley ACTIONS View on publisher site PDF (1.1 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• State-of-the-art and future developments in dosimetry for SSTR- and PSMA-targeting radioligand therapies.#1016098
• American Association for Thoracic Surgery Quality Gateway: A surgeon case study of its application in adult cardiac surgery for quality assurance.#1016099
• Metabolic Reprogramming of Cancer Stem Cells: Targeting Lipid Flux and Mitochondrial Plasticity to Overcome Therapeutic Resistance.#1016100
• Symptom Bother Is Associated With Medical Care‐Seeking for LUTS in Japanese Women: A Survey of 1800 Women.#1016101