Early Warning Systems, Emerging Technologies, and Inductive Risk

Abstract Early Warning Systems (EWS) are a class of instruments that can be used to predict the risk of hazardous events and issue timely alerts. Even though EWS are increasingly used in various contexts, they are not infallible. In particular, given that empirical claims are always underdetermined by the available evidence, there is always the possibility that they are wrong – for EWS, there is always the possibility of false alarms and failed alerts. Philosophers have discussed this possibility as inductive risk, highlighting how both types of errors may lead to significant epistemic and non-epistemic costs. In this paper, we want to emphasise the need to discuss the implications of the use of emerging EWS technologies for inductive risk. While increasing the quantity and quality of data collected and improving the processing algorithms used to extrapolate predictions can hone accuracy, we argue that these technological improvements involve value-laden judgements and a non-neutral stance with respect to inductive errors. We discuss this in contexts as different as medicine and seismic risk management, where the use of big data and machine learning for EWS is rapidly expanding, and we show that it is important to understand the mechanisms by which different technologies contribute to the improvement of EWS. 1 Introduction The label Early Warning Systems (EWS) applies to a variety of digital systems that are used to monitor, forecast, and predict some form of risk, aiming to ensure timely communication so that the stakeholders can take actions to reduce the risk (UNDRR 2017). They are important tools to work on both mitigation and preparedness, saving lives and money: recognizing this, the United Nations have established the “Early Warnings for All” initiative, with the objective to protect everyone from natural risks (UN 2023). In light of their importance for society, the application of EWS has been on the rise, and the number of countries employing EWS for natural risks has doubled between 2015 and 2023 (UNDRR 2023). To issue an alert, EWS infer from the data that the risk has passed a certain (more or less arbitrary) threshold. However, EWS are not infallible: there is always the possibility of false alarms and failed alerts, i.e. of making wrong claims about whether the established threshold has been passed or not. These errors may be due to (i) malfunctioning of the systems or human mistakes in using them, (ii) misinterpretations of the data analysed, (iii) confusing, noisy, or spurious data, or (iv) inductive errors: given that hypotheses are always underdetermined by the evidence, making empirical claims implies the possibility that they are wrong. Emerging, digital technologies can help to hone the accuracy of EWS by increasing the quantity and quality of data collected and by improving the processing algorithms used to extrapolate predictions. We see this in contexts as different as medicine and seismic risk management, where the use of big data and of machine learning models is rapidly expanding. However, this direction of EWS can raise a number of concerns that are at the centre of reflections on the extensive use of digital technologies in contemporary societies and their political and social consequences, such as surveillance, privacy, data access and justice, sovereignty. In this paper, we focus on a specific set of implications that digital technologies can have for risk management, as we discuss the relationship between emerging EWS technologies and inductive risk. As technological improvements of accuracy involve value-laden judgements and a non-neutral stance with respect to inductive errors (Karaca, 2021), it is important to understand the mechanisms by which different technologies contribute to the improvement of EWS. We focus on the possibility that EWS make incorrect predictions due to inductive errors, and we explore how big data and machine learning models work to reduce the possibility of mistakes. To do so, we first introduce the concept of risk and how its management implies the possibility of errors (Sect. 2). We frame this issue in terms of inductive risk (Sect. 3) as the inevitable possibility of making wrong empirical claims. We then present early warning systems and the role of emerging technologies (Sect. 4), which we illustrate with cases from medical and seismic risk (Sect. 5). We discuss the implications of this role (Sect. 6) before concluding with some considerations about research and application of EWS (Sect. 7). 2 Risk The notion of risk has been defined in several ways (Aven et al., 2011). In this paper, we adopt the conceptualisation proposed by the Society for Risk Analysis, the main scientific society in the multi-disciplinary field of risk science. This definition frames risk as two-dimensional: risk is composed of consequences (C) and related uncertainty (U). The consequences are the result of events or activities, they concern something that humans value, and at least one of the possible outcomes is undesirable (Aven & Renn, 2009; SRA, 2017). The uncertainty can concern whether the relevant event/activity will occur or whether the consequences will result from it. From this perspective, we can understand the risks introduced by emerging technologies and discussed in the philosophy of technology (van de Poel, 2016) as contributions to changes of risk levels, for instance because of novelty of emerging technologies: the widespread use of these technologies may have serious consequences (C) on society, but due to the lack of operational experience, these consequences are highly uncertain (U). Note that the uncertainty involved in this characterisation it is not uncertainty due to limits in our knowledge or understanding, but uncertainty as a property of a reality that is, for all practical purposes, impossible to foresee. Given these specifications, we can see that this concept identifies risk as something which exists independently of the assessor (Aven et al., 2011). Considering the possibility of bad outcomes, there is an incentive to reduce risk: following the (C, U) conceptualisation, this consists in reducing either the severity of the consequences or their likelihood. However, while we want to intervene on the actual risk, we can do so only guided by our description of it – we want to act on reality, but we do so based on our knowledge. Any description of risk is therefore relative and contextual, at least in the sense of being related to a specific body of knowledge. While the concept of risk is defined by the pair (C, U), a description of risk is defined by (C’, Q, K), where C’ is the set of consequences specified in the assessment, Q is a (typically probabilistic) measure of the uncertainty, and K is the body of knowledge on which C’ and Q are based (Aven, 2023). Given that any description of risk is based on a specific and necessarily limited state of knowledge, there is a gap between the actual risk that is the target of the description and the risk that is modelled in the description. The existence of this gap opens the door to errors, as it is always possible that our description is inaccurate. This possibility of error concerns both components of risk: the consequences specified in the descriptions may not correspond to the actual consequences, i.e. C’ and C may be different, and Q may be an inaccurate measure of U. Furthermore, the error can go in two directions: it can be either an overestimation or an underestimation of the severity of the consequences or of their likelihood. In the philosophical debate, the possibility of making one such error is known as inductive risk. 3 Inductive Risk Inductive risk refers to the possibility of wrongly accepting or rejecting a hypothesis H on the basis of a body of evidence and knowledge K, and this is a possibility that can never be completely eliminated. Hypotheses can never be verified beyond doubt, as scientific knowledge and evidence by themselves are not direct proofs for whether hypotheses are true or false (Hempel, 1970; Rudner, 1953). As a result, when making inferences, we always run the risk of accepting a false hypothesis or rejecting a true one. These errors correspond to two types, which are usually identified as false positives (wrongly accepting a false hypothesis) and false negatives (wrongly rejecting a true hypothesis). For example, in medicine the study of the state of the patient may lead to a diagnosis of e.g. COVID-19 when the patient does not have COVID-19 (false positive), as well as to the lack of diagnosis of COVID-19 when the patient indeed suffers from COVID-19 (false negative). The philosophical literature has shown that both types of errors have significant costs. From an epistemic point of view, accepting a false hypothesis or rejecting a true one is problematic because it leads away from knowledge. For example, wrongly accepting a false warning about heart anomalies such as atrial fibrillation is epistemically problematic because our resulting beliefs on the health of the patient are fallacious (Biddle, 2016). But inductive risk can also carry significant social and ethical costs. Accepting a false positive from EWS can lead to overdiagnosis, for instance, and unnecessary procedures that expose patients to superfluous risks, which are problematic from an ethical point of view; more generally, they can be a waste of the money, time, and resources of healthcare services, and thus have significant social costs too. Hence, the inductive risk of making false negative and false positive claims cannot be completely avoided and it comes with significant epistemic, ethical, and social costs. The converge of these two factors – the inevitability and costs of inductive risk – has important consequences (Biddle & Kukla, 2017). As the possibility of being wrong is always present and inferences cannot be settled by the strength of K only, it is necessary to establish a threshold for when K is sufficiently strong to accept or reject H. This decision also impacts the potential epistemic, ethical, and social costs of possible errors: lower thresholds of acceptance imply a higher risk of false positives, and thus a higher risk of the epistemic and non-epistemic costs related to false positives, and conversely for higher thresholds and false negatives. In other words, setting a threshold for K effectively establishes how many and which negative consequences we consider tolerable at epistemic, ethical, and social levels – i.e. it implies value judgements about the consequences of being wrong (Douglas, 2000). Therefore, the inevitability of inductive risk implies the inevitability of value judgements, even when it comes to supposedly technical and methodological choices such as the selection of parameters for statistical significance and exemplifies one of two main directions through which science and values are intertwined (Ratti & Russo, 2024), where values enter science and shape these and other choices throughout the scientific process. For instance, consider the choice of significance levels in EWS. Choosing lower significance levels leads to fewer false negatives and more false positives, meaning that less probably problematic conditions will be missed and more probably wrong warnings will be issued. If we are wrong – inevitably, it is always possible that we are – the consequences of errors can include the delivery of unsafe and wrong warnings, which is problematic from both epistemic and non-epistemic points of view. Thus, even decisions that may appear mostly technical inevitably involve value judgements. 4 Early Warning Systems As we have seen, the notion of risk implies the possibility of negative impacts. For this reason, there are several instruments that have been developed to manage risk at different stages, acting before, during, or after the occurrence of potentially harmful events or activities. Among these, EWS have the function of detecting a significant increase in the possibility that the negative event is about to occur and using this information to put stakeholders in the position to reduce risk, either by preventing the event or by taking actions to limit harmful impacts. Even though these systems can vary greatly in structure and mechanism, they all share a common function and a common scaffolding involving four stages: - i. Collection of data coming from the constant monitoring of the system at risk; - ii. Analysis of the data to forecast possible future evolutions of the system; - iii. Assessment of the risk level for each possible scenario; - iv. If the likelihood and impact of the event surpass a certain threshold of acceptable risk, then an alarm is issued. The mechanism of EWS is therefore based on a trade-off between accuracy and rapidity: to be effective, the system should be both reliable in its predictions and fast in assessing risk and issuing the alert. Any approach to improve an EWS, including emerging, digital technologies such as big data and machine learning, can therefore work on either of these desiderata – they can reduce the time required by different steps of the process, or they can hone the accuracy of the predictions. Here, we discuss technologies that focus on the improvement of accuracy, leaving improvements of rapidity aside. The trade-off between rapidity and accuracy is still salient, even with these technologies: while technological advancements may lead to shorter warning times, the processing e.g. of big datasets may instead require longer times. However, the implementations we discuss are aimed at accuracy, and it is the implications of this explicit goal that we discuss, regardless of positive or negative impacts on rapidity. As a starting point, it is important to highlight that the process described in steps i-iv is an inductive process, where empirical observations are elaborated through theories and models to generate unobserved predictions. Being the result of an inductive process, as we have seen, it is always possible that the prediction is wrong – i.e. predictions issued by EWS are subject to inductive risk. Improving the accuracy of an EWS therefore means reducing the inductive risk involved in its mechanism, which however can never be eliminated completely. This can happen symmetrically, i.e. reducing the possibility of false positives and false negatives equally, or asymmetrically, i.e. reducing exclusively or primarily the possibility of either false positives or false negatives. Improvements in the accuracy of EWS can come either from improvements in the data on which the predictions are based (step i) or from improvements in the data processing procedures (step ii), and current technological advancements can go in either of these directions. Among the most significant technological advancements that have improved data collection is the advent of big data, i.e. of technologies that allow the collection of very large sets of fine-grained data (Kitchin, 2025, pp. 20–21). Among those that have improved the accuracy of data processing there are machine learning models, which extrapolate predictions about new data from datasets of previous cases thanks to statistical algorithms that identify patterns and similarities (Kitchin, 2025, p. 164). In the next section, we illustrate these mechanisms by looking at two different applications of EWS. 5 Seismic and Medical Early Warning Systems Let us consider the case of EWS used for earthquakes. The impact of earthquakes on society can hardly be underestimated, both in terms of structural damages and in terms of cultural, environmental, and psychological harms (an earthquake may e.g. destroy historical heritage, tamper cultural practices by tearing a community, trigger chain effects leading to pollution and contaminations). And yet, seismic shocks are an entirely natural phenomenon that it is impossible to prevent, and their aleatory nature makes seismic risk management a particularly challenging field. While it is currently impossible to predict earthquakes, and therefore to work on prevention and on the reduction of their probability, the dual nature of seismic waves provides some time to potentially reduce their impacts. Primary (P) waves are longitudinal, compressional waves that travel through any material; secondary (S) waves are transverse, shear waves that only travel through solids. Most importantly, P-waves are non-destructive but travel significantly faster than S-waves, which means that they reach the areas around the epicentre before the destructive S-waves. This allows a time lag between the beginning of the earthquake and the most dangerous shaking, which can be exploited to reduce the impacts. Even though the time lag is very short – typically a matter of seconds or few minutes at most – effective warnings can still alert the relevant civil protection bodies; urge the population to take shelter; automatically halt specific systems to avoid chain disasters, like nuclear plants or trains at risk of derailment (Allen & Melgar, 2019). Given the limited time available to collect data, assess the risk, dispatch the alert, and take action, the trade-off between accuracy and rapidity is particularly relevant for seismic EWS. Recently, machine learning technologies have started being fruitfully applied in many fields of seismology (Mousavi & Beroza, 2022), including EWS (Jozinović et al., 2020; Otake et al., 2020; Saad et al., 2021). These employ deep learning algorithms that are trained on large datasets of historical earthquakes, usually expanded artificially to better account for the rarity of high magnitude shocks, so that they can use this information to predict plausible evolutions of new seismic events. Münchmeyer et al. (2021), for instance, propose a deep-learning based alert model that predicts future behaviours on initial waveforms detected at arbitrary locations. Crucially, in their tests, the model outperforms traditional systems in the number of both false positives and false negatives produced, and similar trends are identified in other proposed deep-learning based EWS, where using the model leads to high classification accuracy. Instead, the second case we consider – medical EWS – is based on the possibility of monitoring individual users continuously and remotely, for instance trough digital technologies such as wearable devices, which can be worn directly on the body and can track various health parameters continuously (Friend et al., 2023). The large volumes of data collected in this way are analysed by EWS on device or – more often – through Cloud services to identify possible anomalies in the monitored parameters and provide warnings on this basis. Often, warnings are the result of machine learning models, which can identify some correlations as anomalies even in the absence of a strong theoretical understanding of the possible aetiology behind them. Currently, wearable devices are increasingly used to monitor heart health (Friend 2024), to the extent that some of the most famous studies using wearables and the largest medical studies ever have focused on heart health (Lubitz et al., 2022; Perino et al., 2021). Many of these applications function as EWS: for instance, most smartwatches available in the market can monitor constantly key heart parameters such as heart rate, as a basis to provide early warnings directed at users about anomalies that can lead to specific episodes such as heart attacks or more general conditions such as atrial fibrillation. Some systems that are presented more explicitly as EWS are currently under testing for the early detection of organ failures (Hoche et al., 2024; Yèche et al., 2022): machine learning models have been trained on the basis of data from intensive care patients to develop more accurate and timely identification of possible organ failures and alert healthcare professionals to intervene. 6 Discussion In both seismic and medical EWS emerging technologies are used to improve the accuracy of the system. However, the two cases show a crucial difference. So far, seismic EWS rely on machine learning models that are not used to increase the quantity of the data harvested, but mostly to improve their interpretation. On the other hand, currently medical EWS are mostly based on the use of technologies that allow the constant monitoring of patients, thus producing a wealth of big data that can then be processed via machine learning algorithms. While machine learning algorithms can potentially improve accuracy symmetrically – and indeed, this is what we observe in seismic EWS – the use of big data shifts the threshold of inductive risk in favour of false positives (Canali et al., 2026). Constant and more fine-grained observation leads to a reduction in the number of false negatives, as it is less likely that relevant symptoms or significant correlations go unnoticed; but, at the same time, the higher number of variables observed means an increase in the number of spurious correlations and unrelated noise, which lead to false positives. Of course, data processing algorithms can learn from their mistakes and tend to reduce the absolute numbers of both false positives and false negatives – but the ratio between the two will remain in favour of the former. In addition, the grounding of these systems on large datasets and the tendency towards over-detection push towards improvements of accuracy that have little clinical utility and are highly problematic (Green et al., 2025). As more conditions are detected as anomalous and diagnosed as belonging to increasingly expanded disease categories, even true positives that remain inconsequential can unnecessarily burden patients and use healthcare resources unfairly, weighing on society in general. Thus, the use of big data reflects an attitude towards risk that prioritises the reduction of false negatives. The fact that the use of emerging technologies is geared towards avoiding failed alerts is coherent with the purposes of EWS. However, in itself this is not a neutral stance, as it reflects a value-laden, precautionary attitude towards risk which prioritises the reduction of false negatives. While it certainly makes sense to prefer a false alarm over a failed alert for both insurgent health conditions and earthquakes, we have seen that both errors can have significant social and ethical costs. Moreover, EWS only work insofar as they can be trusted – but in a typical “boy who cried wolf” scenario, high levels of false alarms may lead to distrust in the alarm source, which will not be believed when necessary. Interestingly, even true positives can lead to similar consequences if they are inconsequential and unactionable, as we have mentioned earlier in relation to EWS that correctly classify conditions as pertaining to a poorly defined disease category, which does not lead to clinical actions and thus contributes to overdiagnosis (Green et al., 2025). As a result, the management of inductive risk requires a system of accountabilities and responsibilities. Setting the inductive threshold is a consequential choice with potentially significant non-epistemic impacts, and it requires value judgements and the assessment of priorities. Whose value judgements, and whose priorities – and therefore, whose responsibility in the management of inductive risk – is a crucial question, and one that needs to be addressed explicitly for those like policy-makers that set the rules and determine what to prioritise in case of emergency. This question is particularly salient when the management of inductive risk happens through automated and possibly opaque procedures, such as in the case of EWS based on emerging, digital technologies, with consequences for the responsibility of those that apply risk management protocols. Here, systems using big data and machine learning techniques often make human control and the allocation of responsibility to human persons very difficult if not impossible – an issue usually referred to with the notion of “responsibility gap” (see e.g. Santoni De Sio & Mecacci, 2021). Moreover, responsibility for the consequences of inductive errors introduces a further asymmetry between false positives and false negatives. This asymmetry surfaces in the legal context, which allows the observation of social attitudes towards responsibility. Societal aversion to false negatives is reflected in a tendency to search for someone to hold responsible for the consequences of some event that has not been adequately predicted (this search is particularly difficult in the face of the aforementioned responsibility gap). In medicine, this can lead to defensive medicine, i.e. “a way of practicing medicine that is designed to minimize the chances of being sued” (Church 2025). Doctors face the incentive to follow standardized procedures and conduct diagnostic tests not on the basis of their expected predictive power, but rather in order to avoid exposure to legal liabilities in the case of an undiagnosed problem. Similarly, after natural disasters there are often legal investigations into possible responsibilities behind the impacts. In a paradigmatic example, after a Mw 6.3 earthquake hit the Italian city of L’Aquila in 2009, leading to the death of 309 people, the scientific experts who failed to predict the event were processed (and initially convicted) for manslaughter (Alexander 2014; Demichelis & Ongaro 2024). As we have discussed, false positives are not free of consequences either. And yet, responsibilities for false positives and their consequences are less prominently investigated than those for false negatives. In 1985, an alert was issued in the Italian region of Garfagnana, in anticipation of a major shock after some increased seismic activity. The population was largely evacuated, before going back home after 48 h, once the situation calmed down. While the Minister for the Civil Protection was put on trial for procured alarm, he was acquitted, and the population accepted the evacuation with no clamour (INGV 2025). In the medical context, the consequences of a high rate of false positives become particularly relevant at the large scale. As we have seen in the previous cases of overdiagnosis, the ethical costs and personal harms of the single false positive may be significant. At a broader societal level, a high rate of false positives leads to significant social and economic costs, as well as to the risk of consequential distrust in diagnostic procedures. And yet the identification of responsibilities for these consequences is not obvious, often not prioritised, and increasingly difficult with emerging technologies. This is not to say that big data should not be used in EWS, or that seismic EWS are in any sense better than medical EWS. Our discussion rather shows how emerging technologies can improve the accuracy of EWS either symmetrically or asymmetrically with respect to false negatives and false positives. Whether symmetry is positive or not depends on our stance on inductive risk. Ideally, we would want to avoid both types of error altogether; but as this is impossible, then we may prioritise our efforts to reduce those errors that carry the heavier costs – which, as we have seen throughout our discussion, need to include cultural, psychological, environmental, and social impacts, beyond the purely economic costs. As long as false alarms are preferred over failed alerts, then an asymmetrical improvement that reduces the risks of false negatives may be considered a positive outcome. 7 Conclusions By making predictions from observations, early warning systems are open to the possibility of errors, i.e. of either false alarms or failed alerts. Philosophers have discussed this possibility as inductive risk, and many have argued that, given that the two types of errors may lead to social and ethical consequences, apparently technical choices on how to regulate the inductive process involve value-laden considerations. Emerging technologies like big data and machine learning can help improve the accuracy of EWS. But while machine learning seems to do that symmetrically for false positives and false negatives, the use of big data favours false positives. The expansion of variables and the consequent increase in correlations observed with the use of big data, while reducing the possibility that some warning sign is neglected, also increases the possibility of identifying spurious correlations and misinterpreting unrelated signs, thus leading to an increase in the number of false positives. Thus, the application of emerging technologies to EWS provides another example of how societal reliance on technology requires value judgements that should not remain hidden behind technical advancements – while we have made this point in relation to inductive risk, it is equally relevant to many value-laden implications of the widespread application of digital technologies to socially impactful instruments like EWS. Given their function, EWS apply a precautionary approach that tend to privilege false positives over false negatives, i.e. false alarms over failed alerts. In this respect, the asymmetrical improvements of big data can be desirable. However, it is important to keep in mind that this is not a neutral stance and also that an increase in the probability of false positives comes with significant costs. What this means for the practice of EWS is twofold. First, it means that, while recognising priority to avoiding failed alerts, efforts should be made to improve accuracy with respect to false positives too. Second, it means that the costs produced by the occurrence of false positives should be taken into consideration with an adequate system of accountabilities that ensures that they are mitigated and compensated and properly identified and distinguished between those who set the rules and those who apply them. A proper risk management requires also managing the risks created by risk management itself. Data Availability Not applicable. References Alexander, D. E. (2014). Communicating earthquake risk to the public: the trial of the “L’Aquila Seven”. Natural Hazards, 72(2), 1159-1173. https://doi.org/10.1007/s11069-014-1062-2 Allen, R. M., & Melgar, D. (2019). Earthquake Early Warning: Advances, Scientific Challenges, and Societal Needs. Annual Review of Earth and Planetary Sciences, 47(1), 361–388. https://doi.org/10.1146/annurev-earth-053018-060457 Aven, T. (2023). On the gap between theory and practice in defining and understanding risk. Safety Science, 168, 106325. https://doi.org/10.1016/j.ssci.2023.106325 Aven, T., & Renn, O. (2009). On risk defined as an event where the outcome is uncertain. Journal of Risk Research, 12(1), 1–11. Aven, T., Renn, O., & Rosa, E. A. (2011). On the ontological status of the concept of risk. Safety Science, 49(8–9), 1074–1079. https://doi.org/10.1016/j.ssci.2011.04.015 Biddle, J. B. (2016). Inductive risk, epistemic risk, and overdiagnosis of disease. Perspectives on Science, 24(2), 192–205. https://doi.org/10.1162/POSC_a_00200 Biddle, J. B., & Kukla, R. (2017). The geography of epistemic risk. In K. Elliot C., & T. Richards (Eds.), Exploring Inductive Risk: Case Studies of Values in Science. Oxford University Press. https://doi.org/10.1093/acprof:oso/9780190467715.003.0011 Canali, S., Schiaffonati, V., & Aliverti, A. (2026). Evidence-based medicine and the promises and limits of digital health and wearable technology. Studies in History and Philosophy of Science, 116. https://doi.org/10.1016/j.shpsa.2025.102097 Demichelis, A., & Ongaro, M. (2024). Public understanding and scientific uncertainty: The communication of risk in the L'Aquila earthquake. AIMS GEOSCIENCES, 10(3), 540–552. https://doi.org/10.3934/geosci.2024028 De Sio, S., F., & Mecacci, G. (2021). Four Responsibility Gaps with Artificial Intelligence: Why they Matter and How to Address them. Philosophy & Technology, 34(4), 1057–1084. https://doi.org/10.1007/s13347-021-00450-x Douglas, H. (2000). Inductive Risk and Values in Science. Philosophy of Science, 67(4), 559–579. https://doi.org/10.1086/392855 Friend, S. H., Ginsburg, G. S., & Picard, R. W. (2023). Wearable Digital Health Technology. New England Journal of Medicine, 389(22), 2100–2101. https://doi.org/10.1056/NEJMe2303219 Green, S., Haase, C. B., & Spalletta, O. (2025). Precision prevention and the temporal disruption of evidence: The case of heart rate notifications from wearables. Medicine Health Care and Philosophy. https://doi.org/10.1007/s11019-025-10308-0 Hempel, C. G. (1970). Aspects of scientific explanation: And other essays in the philosophy of science. Free Press; Collier Macmillan Hoche, M., Mineeva, O., Burger, M., Blasimme, A., & Rätsch, G. (2024). FAMEWS: A Fairness Auditing tool for Medical Early-Warning Systems [Preprint]. Health Informatics. https://doi.org/10.1101/2024.02.08.24302458 INGV - Istituto Nazionale di Geofisica e Vulcanologia (2025). L'allarme terremoto in Garfagnana del 1985. https://ingvterremoti.com/2025/01/24/lallarme-terremoto-in-garfagnana-del-1985/ Jozinović, D., Lomax, A., Štajduhar, I., & Michelini, A. (2020). Rapid prediction of earthquake ground shaking intensity using raw waveform data and a convolutional neural network. Geophysical Journal International, 222(2), 1379–1389. https://doi.org/10.1093/gji/ggaa233 Karaca, K. (2021). Values and inductive risk in machine learning modelling: The case of binary classification models. European Journal for Philosophy of Science, 11(4), 102. https://doi.org/10.1007/s13194-021-00405-1 Kitchin, R. (2025). Critical data studies: An A to Z guide to concepts and methods. Polity Lubitz, S. A., Faranesh, A. Z., Selvaggi, C., Atlas, S. J., McManus, D. D., Singer, D. E., Pagoto, S., McConnell, M. V., Pantelopoulos, A., & Foulkes, A. S. (2022). Detection of Atrial Fibrillation in a Large Population Using Wearable Devices: The Fitbit Heart Study. Circulation, 146(19), 1415–1424. https://doi.org/10.1161/CIRCULATIONAHA.122.060291 Mousavi, S. M., & Beroza, G. C. (2022). Deep-learning seismology. Science, 377(6607), eabm4470. https://doi.org/10.1126/science.abm4470 Münchmeyer, J., Bindi, D., Leser, U., & Tilmann, F. (2021). The transformer earthquake alerting model: A new versatile approach to earthquake early warning. Geophysical Journal International, 225(1), 646–656. https://doi.org/10.1093/gji/ggaa609 Otake, R., Kurima, J., Goto, H., & Sawada, S. (2020). Deep Learning Model for Spatial Interpolation of Real-Time Seismic Intensity. Seismological Research Letters, 91(6), 3433–3443. https://doi.org/10.1785/0220200006 Perino, A. C., Gummidipundi, S. E., Lee, J., Hedlin, H., Garcia, A., Ferris, T., Balasubramanian, V., Gardner, R. M., Cheung, L., Hung, G., Granger, C. B., Kowey, P., Rumsfeld, J. S., Russo, A. M., True Hills, M., Talati, N., Nag, D., Tsay, D., & Desai, S. (2021). Arrhythmias Other Than Atrial Fibrillation in Those With an Irregular Pulse Detected With a Smartwatch: Findings From the Apple Heart Study. Circulation: Arrhythmia and Electrophysiology, 14(10), e010063. https://doi.org/10.1161/CIRCEP.121.010063. on behalf of the Apple Heart Study Investigators. Ratti, E., & Russo, F. (2024). Science and values: A two-way direction. European Journal for Philosophy of Science, 14(1), 6. https://doi.org/10.1007/s13194-024-00567-8 Rudner, R. (1953). The Scientist Qua Scientist Makes Value Judgments. Philosophy of Science, 20(1), 1–6. Saad, O. M., Hafez, A. G., & Soliman, M. S. (2021). Deep Learning Approach for Earthquake Parameters Classification in Earthquake Early Warning System. IEEE Geoscience and Remote Sensing Letters, 18(7), 1293–1297. https://doi.org/10.1109/LGRS.2020.2998580 SRA. (2017). Glossary. Society for risk analysis. https://www.sra.org/resources Accessed December 17, 2024 UN – United Nations (2023). Early warnings for all. https://www.un.org/en/climatechange/early-warnings-for-all UNDRR – United Nations Office for Disaster Risk Reduction (2023). Global Status of Multi-Hazard Early Warning Systems. https://www.undrr.org/media/91954/download?startDownload=20240503 UNDRR – United Nations Office for Disaster Risk Reduction (2017). Report of the open ended intergovernmental expert working group on indicators and terminology relating to disaster risk reduction. https://www.undrr.org/publication/report-open-ended-intergovernmental-expert-working-group-indicators-and-terminology Van de Poel, I. (2016). An ethical framework for evaluating experimental technology. Science and engineering ethics, 22(3), 667–686. https://doi.org/10.1007/s11948-015-9724-3 Yèche, H., Kuznetsova, R., Zimmermann, M., Hüser, M., Lyu, X., Faltys, M., & Rätsch, G. (2022). HiRID-ICU-Benchmark—A Comprehensive Machine Learning Benchmark on High-resolution ICU Data (arXiv:2111.08536). arXiv. http://arxiv.org/abs/2111.08536 Funding Open access funding provided by Politecnico di Milano within the CRUI-CARE Agreement. Financial support for SC was provided by the FAIR (Future Artificial Intelligence Research) project, funded by the NextGenerationEU program within the PNRR-PE-AI scheme (M4C2, Investment 1.3, Line on Artificial Intelligence). Financial support for MO was provided by Fondazione Cariplo, grant n° 2024–1211. Author information Authors and Affiliations Contributions Both authors contributed equally. Corresponding author Ethics declarations Ethical Approval The research for this paper did not involve any studies on human subjects or animals. Informed Consent Not applicable. Conflict of Interest The authors declare no conflict of. The funding agency had no role in study design, analysis, writing the article and in the decision to submit the article for publication. Additional information Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Rights and permissions Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/. About this article Cite this article Canali, S., Ongaro, M. Early Warning Systems, Emerging Technologies, and Inductive Risk. Digit. Soc. 5, 48 (2026). https://doi.org/10.1007/s44206-026-00289-9 Received: Accepted: Published: Version of record: DOI: https://doi.org/10.1007/s44206-026-00289-9