Latency,uncertainty,contagion: Epistemologies of risk-as-reform in crime forecasting software

Introduction: From prediction to risk forecasting

In March of 2021, the predictive policing software company PredPol announced that it was changing its name to Geolitica, a move that follows the broadening of their focus from prediction to ‘geographical analytics’ and promises an expansion of the types of data their algorithm uses. Their original model, comprised only of crime type, location, date and time, has come under public scrutiny for its potential to further entrench racial bias on the argument that sending police to ‘predicted crime areas’ will become a self-fulfilling prophecy as more crime data entering the algorithm continues to focus attention on certain areas, the vast majority of which are lower-income and majority non-white. Mainstream media response has been pointedly skeptical or outright condemnatory, and activists from a range of positions critical of the police as well as numerous civil rights organizations have taken public and strong stances against it (American Civil Liberties Union, 2016; Robinson and Koepke, 2016; Tett, 2014). Speaking to why the company is changing from PredPol to Geolitica, PredPol revised the place of prediction in their work, stating, “Even the use of the word predictive itself does not accurately describe our business. Rather than predicting, we were conducting risk assessments” (PredPol, 2021).

What this quote reflects is the current reworking of the impetus for law enforcement to act on algorithmically determined claims about future crime into what is called ‘crime risk forecasting.’ This takes the form of software that combine historic crime data ¹ and a range of geolocated data not specific to crime, such as the locations of bars, pawn shops and mental health facilities, in order to produce future-looking maps of likely ‘criminal activity.’ These algorithms are increasingly used to bring various stakeholders together around a purportedly shared understanding of which areas of a city need attention from police and other municipal agencies, either under the banner of ‘police–community relations’ or in recently developed collaborative initiatives that provide supposedly non-police-centered, but police-inclusive, solutions to ‘crime’ such as the Dallas Office of Integrated Public Safety Solutions (Caplan, n.d.). That these initiatives purport to take a public health approach to community violence and promote city investment in resource-poor areas indicates the cooptation by police reform efforts of claims that arose in the context of more radical calls for defunding the police and for police abolition. ² Reform, as it is understood in what follows, corresponds to the counterinsurgency efforts of state and capitalist interests dedicated to summoning political subjects to their vision of a world that requires police. ³

The technocratic promises of algorithms are called on to exert their insidious pull as flashpoints like the George Floyd uprisings of 2020 throw political perspectives in flux and, if briefly, displace the familiar terms of a public conversation about police. ‘Vulnerable areas’ have since been repeatedly invoked by politicians and police advocates to levy, against calls for abolition, a vision of urban life that in fact reinforces institutional abandonment and the exigencies of racial capitalism (Glueck and Southall, 2022; Rao, 2022). ⁴ Software that use environmental data to forecast ‘criminal risk’ write this vision into code and prop up the reorientation of critical conversations around reformist tropes of ‘accuracy,’ ‘transparency’ and ‘accountability.’ My aim in this article is to update the critical frameworks available for contesting technocratic reforms as they coalesce around policing algorithms through a detailed examination of how the function and premises of ‘data-driven’ policing are altered by the turn to epistemologies of risk. In a risk-driven paradigm, I suggest, the latent possibility of ‘disorder’ replaces visible aberrations on ‘normal’ order; indeterminacies, rather than biased certainties, justify a range of police interventions; and a focus on arenas of contagion supplements the spatial logic of containment and the securitization of borders. I will further schematize these points later on in this introduction, but first I want to make clear that the technical and epistemological structures of these software differ from earlier iterations in two important ways.

The first shift is that crime risk forecasting expands the types of data sets used and rescripts the function of those data sets. The inclusion of ‘environmental data’ goes beyond merely layering visualizations of crime data and common urban features. Environmental features themselves, through an algorithm that associates them temporally and spatially with crime data, are coded as risk factors and their density and proximity to one another are used to forecast which areas will be at ‘higher risk’ and when. Rather than producing information on where crime data (meaning police) aggregate spatially in a city, risk forecasting transforms spatial correlates of the built environment into vectors of future ‘crime’ through their co-occurrence with one another. The list of ‘environmental features’ that are fed into a model are determined largely through existing criminological work that correlates place types with crime data through statistical modeling or police ‘knowledge.’ These criminological papers begin with the assumption that certain types of places, and not others, are both risky and entirely generalizable. A type of place that scholarship on a particular city produced as ‘risky’ becomes a ‘starter feature’ for wherever the software is used (Kennedy et al., 2010). The second, related aspect is that the turn from prediction to risk means that the software is no longer simply trying to accurately match past crime data to future crime data. The calculation and use of ‘environmental risk factors’ mean that a landscape can evince numerous risky sites without ever evincing notable levels of what appears as ‘crime’ in police data. ‘Environmental risk’ supersedes what is constructed by police data as ‘crime’ for determining the deployment of police officers and the scope of police activity, intervening in a present moment that may never have materialized into what was ‘forecast.’

Analyses of risk-as-securitization that have arisen in response to the post-911 state integration of crisis management tools and the advancement of modeling techniques for impending climate crises have identified this temporal framework of risk as one of anticipation, by which techniques for composing certainty in the face of uncertain futures direct interventions in the present (Amoore, 2013; Anderson, 2010; Derickson, 2017). The emphasis within crime risk forecasting on ‘environmental data’ able to reveal potential futures that inform present action certainly draws on the temporal logic of risk evidenced in these techniques. I suggest, however, that crime risk forecasting relies on a somewhat different spatial logic because, in order to function as a police reform, proponents of this software need to sell a picture of urban life in which danger is imminent, intimate and diffuse. I trace the ways risk epistemologies enable tactics of policing and proselytization for police reform that rely on activating a social logic in which public life is understood as fundamentally requiring pre-emptive protection from itself.

The orientation I bring to studying the police is indebted to scholarship that examines actual or professed shifts in policing strategies in order to elaborate what aspects of state power can be, and are, contested in the terrain of politics. Studies of community policing (Gilmore, 2016; Herbert, 2006) and broken-windows policing (Hanhardt, 2016; Harcourt, 2001; Williams, 2015), for example, show that policing strategies index civic visions able to move institutions and actors closer or farther apart in ways that matter to sustaining an informed critique of reformist initiatives and tracking how what André Gorz termed ‘reformist reforms’ coopt and divert the aims of ‘non-reformist reforms.’ ⁵ My research begins from an abolitionist perspective and thus within a frame of interpretation that does not take crime statistics as an actual indication of either ‘real’ distributions of criminal activity or of harm caused to others. In carceral responses to situations where harm registers as crime, harm is not necessarily mitigated; in many cases, as Beth Richie and numerous other demonstrate, it is exacerbated, particularly for those already marginalized and lacking resources (Davis et al., 2022; Gilmore, 2017; Kaba and Ritchie, 2022; Richie, 2012; Ritchie, 2017; Vitale, 2017; Wang, 2018). The distinction between crime and harm is crucial to transformative justice and the process of building a world that can address harm without carceral intervention, and thus it is important to attend to the ways that technocratic projects devise new knowledge-making practices that seek to foreclose this distinction and extend the framework of policing into other public sectors and ‘community initiatives.’

In what follows, I put forth a series of arguments that build towards an understanding of the vision of governance enacted by this software as one in which spaces where people publicly come together are both what cause ‘crime’ and what must be protected from it. In the first section, I demonstrate that the strategies of differentiating risk from prediction construct a framework that promotes the need for police in the absence of what is defined as ‘crime’ in police data. This reorients the order maintenance paradigm of policing from a focus on visible ‘disorder’ to algorithmic outcomes that translate a lack of perceptible aberrations into maps of a latent, potential ‘disorder.’ The second section argues that, owing to this vision, criminal risk forecasting is able to operate by making meaning out of what is difficult to prove, and thus what is difficult to disprove, in ways that circumvent political pushback anchored in concerns of objectivity and bias. Similar to the persistence of risk in the absence of crime data, the inability to show exactly what ‘worked’ to shift crime numbers is used to justify the need for further risk management interventions. The third section argues that the difficulty of proving efficacy provides the grounds for multiplying possible interventions initiated by police and further extending the analytic premises of criminal risk into other civic spheres, such as public transit routes and infrastructure projects. This paradigm likens the spatial study of ‘criminal risk’ to epistemologies of contagion and rearticulates ‘crime’ as both the symptom and disease necessitating the proliferation of management activities.

These contributions are built on analyses of interviews with developers, industry professionals and law enforcement practitioners as well as white papers, published statements and promotional materials. ⁶ The interviews I conducted were semi-structured, lasted an hour to two hours each, and often involved follow-up questions over email. I reached out describing my project as an examination of the rise of algorithmic technology in municipal policing. Interviewees tended to assume that I am not in opposition to their work, in large part because they imagine themselves to be fixing a broken system in response to recent critiques of policing and their own skewed and parochial definitions of ‘community needs.’ When I asked questions about the racial dimensions of their data sets and the underlying systemic issues of urban inequity, they commonly responded by assuring me that awareness of these problems is central to their work, before going on to describe the actuality of that work – how data is used, how risk is defined, and how interventions are staged – as fundamentally oriented around naturalizing crime, police, and the indelibility of structural inequity. Interviews allowed me to see what happens in the background of published materials that spend months in production with a team of strategists crafting the language that they want associated with their product. Insight into the social and political premises of this software’s knowledge-making aims emerged from attending to the common concepts and presuppositions people reached for to more colloquially tell the stories of perceived successes and challenges. My aim is to understand what epistemic forms are at work here and thereby, as I elaborate in the conclusion of this article, lay the groundwork for approaches that can broaden contestations from concerns with what a software does to the visions of civic society and political belonging that a software relies on and enacts.

Part one: Latency

This article focuses on two software that undertake crime risk forecasting: the product called ResourceRouter sold by SoundThinking and the approach known as risk terrain modeling (RTM) as it is packaged and sold by the company Simsi. Similar to the rebranding vignette that opened this article, SoundThinking is the latest iteration of the gunshot detection software company formerly known as Shotspotter, and ResourceRouter is the renamed version of their product previously called Shotspotter Connect. ⁷ The synopsis of the creation of both ResourceRouter and Simsi is that federal money allowed academic, police and industry partnerships to develop software then or now owned by private companies who currently sell that software to police departments. RTM is a theoretical approach that focuses on human-made environmental features as indices of criminal risk. Two criminologists at Rutgers, Leslie Kennedy and Joel Caplan, began developing RTM in the first decade of the 2000s funded by grants from the National Institute for Justice, an agency of the US Department of Justice dedicated to the development of ‘knowledge tools to aid law enforcement.’ They brought on developers from the B Corporation ⁸ Azavea to help with the technical side and conducted on the ground experiments through partnering with police departments in multiple mid-sized cities in the United States. CEO of Azavea Robert Cheetham was also in the process of developing a crime forecasting software through grants from the National Science Foundation and partnerships with criminologists at Temple University. Azavea developed Hunchlab 2.0 into a crime forecasting software that uses multiple crime theories including RTM and sold it to SoundThinking, who have since developed it into the patrol management software called ResourceRouter. Back at Rutgers, Caplan and Kennedy turned their work on RTM into a company called Simsi that sells RTM software to police agencies, government sectors and non-profits. ⁹

My analysis focuses on these two products because of their common usage of particular forms of environmental data to calculate risk. Although ResourceRouter incorporates multiple crime theories, the dashboard that police view when carrying out directed patrols is designed to show them the top three environmental features drawn from the calculations of RTM. RTM works by combining environmental feature data on the locations of things like liquor licenses, pawn shops, methadone clinics, bus stops and laundromats with historic crime data for the city. For ResourceRouter, this is a standard 5 years of crime data, and for Simsi the time span can vary. The algorithm gleans from this data which features are most often associated with clusters of crime, and these are the features that become important to the calculation of risk. The process might begin with a data set of environmental features that includes a variety of types of places, as the suggested starter list on riskterrainmodeling.com has 78 feature types, but the algorithm could easily find that only 5 of those are statistically meaningful. The algorithm determines what environmental features matter across the landscape of the city as a whole, divides the city into cells that are smaller than a city block, and assigns each cell a risk score given the proximity and density of the environmental features that have been determined to be high risk (Caplan and Kennedy, 2016; HunchLab, 2015).

Developers from ResourceRouter and Simsi were intent upon communicating that their focus on risk factors does not reproduce the spatial biases of hot spot policing. Hot spot policing refers to the practice of sending police to areas determined to be ‘high crime’ based on the fact that crime has occurred there in the past, and it is the backbone of the original models of predictive policing developed by companies like PredPol (Braga et al., 2012). The issue with hot spot policing, from the perspective of developers, is not so much that it is inaccurate or biased but rather that it requires crime to keep happening in particular geographic ways in order for an algorithm to operate well. According to Joel Caplan, “The biggest one [difference] is that prediction requires the problem to continue in order for it to occur where predicted” (2022, personal communication). Rather than returning again and again to problem areas – and thus, as critics of predictive policing have emphasized, creating problem areas – the inclusion of environmental context data is supposed to disperse the clustering of police by showing places that could become hot spots in the future. This means that the focus of the algorithm becomes areas that are not high crime yet but might become so according to the presence of risk factors (Caplan and Kennedy, 2011; Lester Wollman, 2022, personal communication; Simen Oestmo, 2022, personal communication). Once ‘environmental risk factors’ have been determined by the algorithm, their co-location and proximity to one another feed the calculation of geographically dispersed risk scores.

Developers’ articulations of the potential for risk forecasting to ‘correct’ the ills of earlier predictive software are rooted in the assertion that the key analytic structure underlying risk-based policing is not strictly tied to crime data. While crime data is initially used to identify risk factors from publicly available data sets, thus excluding, for example, anything that happens in and around middle- and upper-class private homes but including structures like public housing, risk factors then point to “levels of criminogenesis” that require attention even in the absence of crime (Caplan and Kennedy, 2011: 115). ¹⁰ This idea is succinctly illustrated in Leslie Kennedy’s contention that, “Crime does not have to actually occur in a place for it to be a risky place” (2022, personal communication). What Kennedy’s quote reflects is a worldview in which police and policing structures are promoted as necessary even when what they understand as ‘crime’ has not emerged to a level that registers as notable in police data. To use a blunt hypothetical, this paradigm promotes the idea that, even if there is no crime, there will always be risk and therefore the need for police to keep risk from translating into crime. This inverts a commonsense notion that upholds and justifies the police as an entity entailed by the existence of crime, instead reconfiguring them as stewards of a society in which crime, though possibly absent, is always latent: first police, then (maybe not) crime. If the forecasted area does experience a criminal event, it is proven that the situation was risky; but if such an event does not happen, the lack of this event does not disprove the idea that the situation is risky.

The fact that an absence of crime data does not indicate an absence of risk renders the use of ‘high crime’ designations inadequate for police decision-making, thus opening space for the forecasting algorithm to be formulated as showing police what they cannot see from the spatialization of ‘crime’ alone and extending, to use Caren Kaplan and Andrea Miller’s phrase, the “police sensorium” into a latent space of present but not-yet-actualized ‘threat’ (2019). This rendering of risk through environmental data reflects familiar tenets promoted by broken windows theory: criminalized behavior is linked to the built environment, and police similarly justify showing up to ‘risky’ environments through the idea that their presence mitigates the development of potential ‘crime’ into actual ‘crime.’ What is notably different here is what actually shows up as ‘risky’ in the built environment. Broken windows theory relies on the criminological argument that visible signs of ‘disorder’ encourage and escalate criminal activity, but the framework of crime risk forecasting does not rely on perceived aberrations of ‘order’ to identify ‘crime-generating’ areas. Rather, the data sets used to locate ‘risk’ reflect and visualize normative order-making practices of cities, in all their classed and racialized understandings of what should be where: zoning requirements for businesses, urban planning decisions around the placement of parks, food and resource deserts, histories of displacement around the construction of highways, and the priorities of public transit routes, to name a few.

This distinction is important to attend to because it updates accounts that expose the disciplinary, life-curtailing function of the distinction between order and disorder at work in broken-windows policing and its legacies. The approaches of Bernard Harcourt (2001) and Christina Hanhardt (2016), for example, have shown that ‘visible disorder’ is defined in the same stroke that its associated populations are controlled and subdued, allowing police to surveil and attack the lifeways of anyone outside the narrow window of white, middle-class heteronormative subjecthood. Policing practices developed in the 1990s that relied on a zero-tolerance policy for ‘quality of life’ infractions promoted a vision of disorderly city life that is indelibly racialized and intended to protect the interests of middle- and upper-class urban dwellers (Jefferson, 2020: 117). In the case of crime risk forecasting, however, the normalizing line between order and disorder is no longer between perceptible categorizations of place and human behavior. Instead, the algorithm purports to see in perceptible order the forces of disorder that have not yet been actualized. The line between order and disorder is perceptible only through the algorithm’s ability to visualize distributions of potential futures hiding under cover of apparent order.

A story told to me by Jonas Baughman, captain at the Kansas City Police Department (KCPD) and major proponent of RTM, is illustrative here. Kansas City was one of the seven cities that participated in the federally funded RTM experiments from 2012 to 2013. The risk terrain model identified bus stops, vacant buildings and liquor stores as important criminogenic features which, according to Baughman, allowed the department to see ‘crime’ that had been hiding in plain sight. Referring to a particular bus stop, Baughman said this site was “the focal point that allowed the loitering.” He explained, “We assumed that people were waiting for the bus for legitimate transport purposes, but they were actually using and conducting drug deals” (2022, personal communication). The line between order and disorder, here, lies between how the world appeared – a bus stop with people waiting for the bus – and what the algorithm was able to suggest about what was ‘actually’ happening – loitering that enabled criminalized drug sales. While police harassment at transit stops is not in itself new, the epistemological restructuring attendant to risk means that crime forecasting joins the array of algorithmic technologies that remake the present through perceiving the threatening and unstable futures the present is thought to obscure. In departing from a line between order and disorder that can be contested through exposing its qualities, the algorithm rewrites ‘disorder’ as a latent state that could (theoretically) be anywhere at any time and look like anything.

Part two: Uncertainty

Towards making an argument about indeterminacies, I want to first clarify the distinction between accuracy and efficacy as well as the logics of assessment and justification within each. Accuracy refers to how well the algorithm was able to match where it thought crime data would be collected with where crime data was actually collected as compared to the matches of randomly generated maps. ResourceRouter uses a standard ‘area under curve’ assessment method, and from a purely statistical perspective their software is able to generate significantly more matches than random allotments (Shapiro, 2020: 171–173; Wollman and Oestmo, 2022, personal communication). These are the types of measurements that become the accuracy statistics used to advertise such products to police departments and to support arguments for the scientific veracity and rigor of this approach in white papers, compendiums and user guides (Caplan and Kennedy, 2011: 19, 77, 86; Hunchlab, 2015; ResourceRouter, 2021). Efficacy, however, refers to how well the product is positioned to achieve its stated goal – the reduction of what registers in police data as ‘crime.’ Efficacy determines whether the activities undertaken based on a software’s outputs are having their intended effect, and assessing efficacy from the vantage of data science is quite difficult. Developers compare crime data across time periods in search of reductions and spatial redistributions, but it is not possible to compare crime data that was collected to ‘crime’ that would have happened in the absence of police presence (Caplan and Kennedy, 2011: 86, 95; Wollman, 2022, personal communication). Comparing reductions or redistributions has virtually no explanatory power to determine whether actions instigated by the algorithm’s outcomes were the primary cause of statistical spatial changes. As evidenced in Aaron Shapiro’s ethnographic work with the early developers of Hunchlab (the product sold to SoundThinking and now called ResourceRouter), once the software begins directing patrols and police activities, there is no longer an operative control category against which to compare its effects (Shapiro, 2020,:175–176). As the algorithm gets going in the world, what would happen in its absence is statistically unobservable.

Strategies oriented by risk are tasked with deterring events that may or may not happen, and thus the metrics used to evaluate accuracy are not necessarily meaningful to the question of whether or not risk was ‘efficaciously’ managed. Statistical evaluations of efficacy are often not possible because they rely on comparing new data to a hypothetical – what, speculatively, would have happened – or to a situation that has so many outlying variables that it would be impossible to isolate the purported effects of the software. Risk management, as the deterrence of a possible future, relies on comparing what did or did not happen to what could have, operating in an epistemological structure that detaches metrics of assessment from the realm of proof. Uncertainty, here, is not about whether the algorithm can generate matches between forecasted and collected data. What is constitutively indeterminate is the role the software plays in shaping particular outcomes as both a participatory agent in those outcomes and a lens for diagnosing them. There is no way to prove or disprove the ‘success’ of risk management from within a framework that continually generates risk as its only form of assessment. Rather than defining a future and testing whether it happened (as did prediction), risk forecasting enables the ongoing conversion of indeterminate statistical worlds into actionable operations.

What I refer to as uncertainty lies not only in the world, in the sense of an unknowable future, but in the evolutive relation between the algorithm and the world it participates in. Louise Amoore’s work on the epistemology of cloud computing algorithms has similarly argued that “particular actions that might appear as errors or aberrations are in fact integral to the algorithm’s form of being” (2020: 23). She shows that the excision of ‘bias’ is not an operative horizon for algorithmic software because the work that an algorithm does lies in acts of discerning, discriminating, grouping, and determining pathways (2020: 8, 19–20). I want to extend this insight towards parsing the ways that the polysemic nature of the term ‘bias’ has tended to mystify political reckonings with policing algorithms. A critical lens has developed within scholarship on the data technologies of policing that focuses on revealing and refuting claims to objectivity made by advocates of these technologies. The idea that proponents of algorithmic governance understand algorithms to be unbiased, or specifically antithetical to racial bias, has informed numerous critical accounts and political programs that reckon with the racially biased outcomes of algorithms. ¹¹ Ruha Benjamin’s influential account of racism and technology critiques digital forms that “aim to fix racial bias but end up doing the opposite” and “shroud racist systems under the cloak of objectivity” (2019: 8). To be sure, crime risk forecasting is marketed using terms like ‘objective data’ and ‘unbiased outcomes,’ and it participates in the larger technocratic paradigm that Benjamin diagnoses. What I want to point to, as a contribution to this diagnosis, are the ways that critique oriented by refuting objectivity and revealing bias often understands those terms quite differently from their usage in data science. That these terms do not mean the same thing across knowledge communities means we risk misunderstanding how the creators of software understand the software to work, which in turn has important implications for the political claims that can be usefully levied against it.

Rather than pointing to the fact that a perceived spatial pattern is the result of systemic inequity and racialized institutional abandonment, as we might understand it, developers’ use of the term ‘bias’ adjudicates how well a data set captures what they understand as a pattern in the world. For Oestmo, developer on ResourceRouter, the main issue with the software is that people living in certain neighborhoods do not call the police as often as those in other areas. He asserted that the incident data used to correlate risk factors is skewed by this because these areas have higher incidences of ‘crime,’ based on arrest data, but people living there do not trust the police. For him, this constitutes an issue of ‘bias’ that is one of the primary hurdles the software faces vis a vis ‘underserved areas.’ ¹² When I asked the about the potential for certain data sets to correlate strongly with racial disparities, Oestmo responded that their team is always searching data for potential correlates that might overdetermine the algorithm’s outcomes and that many of them do correlate with race. He mentioned that “even geography [topography] correlates with race” but that, in his view, it is useful to include variables that might have racial correlates in order to get “a bigger, fuller context of what’s going in the city” (2022, personal communication). He explained that adding more information can help wash out biases, rather than excise or correct them. Here, the fact that spatial data in US cities are skewed along racial lines translates into an understanding of bias that reasserts crime forecasting as a tool that works in spite of the difficulty of producing objective or unbiased calculations.

Although the vision of a decision-making practice that corrects human prejudices still defines discursive strategies linking software companies to the rebranding efforts of police, it is important to understand that when software developers invoke objectivity they refer to an organizing principle that is about striving to match data sets to a world understood in terms of those data categories. While we, rightly, levy claims that the categories themselves are built on bias, ‘objectivity’ operates for developers as a way of delineating an operative threshold of indeterminacy: it decides how much of what is not known by a data set should be considered acceptable for initiating an action plan. Crime risk forecasting adjudicates the statistical threshold at which the biases inherent to spatial data on US cities will become irrelevant to what an initiated action plan looks like. To argue that this threshold ought to be ‘never’ is to misunderstand how algorithms work, but knowing this can inform political claims that take on the software’s core work of converting indeterminacies into actionable directives. If developers themselves understand bias to play an operational role in the algorithm, we should consider shifting our critical framework away from a paradigm that relies on revealing the biases entrenched by professed objectivity and towards intervening in the techniques and imaginaries that allow ambiguity itself to be operationalized. As I will show in the following section, the constitutive indeterminacies of crime risk forecasting allow the aims of ‘criminal risk management’ to justify a proliferation of actions that could matter. Because the algorithm doubts, a multiplication of interventions arises to fill up the space of certainty required for curtailing future ‘danger.’ ¹³

Part three: Contagion

Thus far I have argued that the epistemology of risk that defines crime forecasting relies on the visualization of latent ‘disorder’ hiding within mundane urban life and the conversion of indeterminate statistical processes into plans for intervention. In what follows, I focus on what risk means for how those interventions are constructed, in terms of what they actually are and how proponents understand them to relate to a particular vision of urban problems. This vision likens the emergence of ‘crime’ to the spread of disease and proposes interventions into interfaces of contagion that are interspersed throughout a given landscape. Rather than a modality of securitization that strives for the provisional fixing of an inside versus an outside, risk indexes a social logic in which the fabric of immediate social life is understood to contain the potential for its own unraveling. Demonstrating that this logic is at work here requires staying with, briefly, the question of efficacy in order to understand how particular interventions are rendered necessary.

The two products I examine diverge in the actions they are able to initiate, such that claiming confidence that the software ‘works’ correlates to the breadth of interventions the software can enable. In other words, the more types of things the algorithm can justify doing, the more confident developers are that it is successful at managing risk. ResourceRouter produces crime risk forecasts primarily for police departments, and the outcomes generated by the software initiate actions by telling police officers where to go and for how long. Oestmo and Wollman, the ResourceRouter developers, readily admitted that a reduction in crime does very little to actually prove whether the software is performing as it is supposed to (2022, personal communication). They referred to a report in production that ties crime data to deployment locations, but, as Wollman expressed, they understand that “crime is affected by a hell of a lot more stuff than this” (2022, personal communication). The RTM product sold by Simsi, however, initiates a range of suggestions beyond directing police deployment, and developers here expressed certainty that this software accomplishes what it is meant to. Software outcomes lead to environmental changes such as more comprehensive LED lighting, the mitigation of barriers to vision, boarding up vacant properties, and the removal of bus stops, to name several that came up in interviews (Caplan, 2022, personal communication, 2022; Kennedy, 2022, personal communication). Baughman, Kansas City police captain mentioned above, was able to effect the removal of the ‘risky’ bus stop that was deemed to be providing “cover for our bad guys” (2022, personal communication). Kennedy shared a similar story about the removal of a bus stop in Dallas as well as the closure of rooming houses in Atlantic City that were flagged at risk for “prostitution” and drug dealing (2022, personal communication). Kevin Oden, director of the Dallas Office of Integrated Public Safety Solutions, referred to projects for “cleaning up” vacant lots and properties to “avoid people using them as stash or use houses” (2023, personal communication).

Because it is structurally difficult to assess efficacy in relation to risk, the assertion that this software does function as it should relies on multiplying the possible actions that could be meaningful. Indeed, not being able to prove exactly what it was that ‘worked’ justifies the expansion of knowledge produced by police analytics into other civic arenas. RTM remains oriented by perceived ‘criminal risk,’ but its users extend beyond police departments to non-profits and other government agencies that are in collaboration with police departments in initiatives such as the Newark Public Safety Collaborative and The Dallas Office of Integrated Public Safety Solutions. This diagnosis of the expansion of police influence effected by crime risk forecasting resonates with analyses of broken-windows policing as a program that “has vastly broadened the capacities of police both nationally and globally” (Camp and Heatherton, 2016: 2). Indeed, it may be that this move is definitive of any initiative for ‘improving’ the police. The bets being hedged here are not primarily concerned with making the police more palatable but, rather, less abolish-able because they are more integrated and diffuse. The Dallas Office of Integrated Public Safety Solutions is founded on “utilizing officers in non-traditional roles across the board to serve our residents” and is comprised of nuisance abatement officers and code enforcement officers who, in the vision Oden shared with me, will one day be able to make economic and housing development recommendations based on the outcomes of RTM (2023, personal communication).

The actual interventions initiated can be read as of a piece with policing tactics that use the requirements of civic bureaucracy, such as the enforcement of city code, to either criminalize houselessness, public behaviors and physical disrepair (Bloch and Meyer, 2019; Christensen and Albrecht, 2020; Ramírez, 2019) or divert city resources through naturalizing the presence of ‘disrepair’ in marginalized communities (Bartram, 2022; Graziani et al., 2021). Scholarship on the ‘data-driven’ technologies used to securitize urban spaces has tended to understand the spatial modality in which they operate as one of containment and sequestration, evincing a geographic mode that scales the mechanics of national borders down into the city and analogizes spaces of control such that US-occupied foreign territories require similar tactics to ‘high-crime’ municipal areas (Jefferson, 2017: 103; Jefferson, 2020; Miller, 2019). This modality is certainly at work here, particularly in the legal grey area around whether entering a ‘risky area’ can construct reasonable suspicion enabling police to question and search or detain individuals. ¹⁴ The fact that features that tend to come up as ‘risk factors’ are facilities and resources integral to low-income communities, such as laundromats and pawn shops, reflects a familiar story in which police operations work to entrench drastically inequitable access to resources and public amenities.

However, part of the project of crime risk forecasting is constructing an image of ‘crime’ that in fact does not rely on the spatial logic of its being contained to particular geographic areas. The logic of controlled containment is supplemented by a framework of contagion that informs the construction of ‘risky spaces.’ The technical structure of these software shifts from the practice of spatial demarcation at work in predictive policing, one that might produce maps targeting whole neighborhoods, to a practice of pinpointing micro-spaces of contact and conditions of ‘spread.’ An environmental ‘risk factor’ is thought not to signal a part of the city per se but rather a likely point of emergence, a place where the environmental conditions are revealed to attract ‘crime,’ much like one would understand that stagnant water increases the risk of malaria. One of Kennedy and Caplan’s papers on the usefulness of RTM applies it to the cholera outbreak of 1854 London and shows that the model would have accurately identified the contaminated pump. This means, they argue, that the spatiality of ‘criminal risk’ is able to be analyzed in the same terms as the spread of disease (Caplan et al., 2020). This framework seeks to explain a problem, be it disease or ‘crime,’ through attending to the conditions thought to enable that problem to occur in new places and move from one place to another. This is evident even in the discursive construction of police activities, as ResourceRouter operationalizes risk scores by directing three 15-minute patrol intervals referred to as ‘dosage.’ As a risk map visualizes where the next ‘outbreak’ will occur, police presence is likened to the preventative treatment of a disease that may have not yet emerged.

This framework is less interested in defining and maintaining the boundaries of a geographic area than in understanding risk as something that arises at public interfaces. Here, we see another of the epistemological tricks of police reform. The structure of risk has digested the claim that police are harmful because they go to the same areas again and again and thus shifted the spatial modality of justifications for police to a vision in which everywhere may hold potential and latent danger. Referring to the ways that he seeks to dispel concerns over the potential of RTM to concentrate police in low-income communities of color, KCPD captain Jonas Baughman said, “Even in the most affluent areas of our city, where we have less violent crime, you will still see hits in the RTM models of high-risk factors.” This is because, he explained, “risk is throughout our city” (2022, personal communication). This spatial pseudo-democratization of risk is meant to index the software’s departure from problematic geographies as the arena of police activity, but also what reasserts the necessity of police presence in light of this departure. The importance of an epidemiological framework to this shift lies in the purported focus on sites of intervention as ‘conditions’ rather than areas to be contained. This turn to environments does not disarticulate risk from individuals entirely, as Andrea Miller has argued attends the environmental control logics of drones (2019: 97), but, rather, is imagined as arising between individuals and one another and between individuals and their surroundings. The sites identified as ‘risky’ are understood less as a container in which threat must be kept and more as an interface between a series of material micro-factors that can be adjusted, tweaked, and infrastructurally modulated to discourage the ‘catching’ or ‘spreading’ of a criminal impetus. The spatial logic of quarantine is joined by the spatio-temporal logic of communicability, wherein the conditions for the communication of ‘criminal impulses’ are understood as a locus of intervention that could pre-empt the need for quarantine.

There are numerous historical and ongoing cases of the translation of public health analytics into racialized security measures (Fluri, 2014; King, 2003), and crime risk forecasting supplements measures to sequester the ‘risky’ from the protected with an understanding that the interfaces of urban life, wherever they exist, are themselves able to attract and facilitate ‘risk.’ I do not mean to suggest that spatial containment is undone or replaced, and the material processes of containment identified by Brian Jordan Jefferson, Emily Kaufman and others participate actively in what ‘risk-based’ policing looks like on the ground (Jefferson, 2017; Kaplan and Miller, 2019; Kaufman, 2016). But an environment, here, operates as both an enforceable container and a “series of operations,” as described by Zala Volčič and Mark Andrejevic, in the sense that what is manifested by a ‘risk factor’ is already a thesis about the steps necessary to intervene in the relations taking place there (2021). Where Andrejevic and Volčič continue to focus their analysis of ‘smart’ cameras on the spatiality of enclosure, however, crime risk forecasting supplements the focused monitoring of demarcated spaces by formulating a data visualization of what is explicitly called a ‘risk terrain,’ simulating a broad and variegated landscape that pinpoints micro-locations where the processes of urban life are thought to mask latent threats and criminogenic conditions.

Conclusion

By way of conclusion, I want to sketch two premises of policing and civic society that come into view through crime risk forecasting and that work to shift political perspectives towards a world in which, to quote the director of the Dallas Office for Integrated Public Safety Solutions, “you’re always going to need law enforcement” (Kevin Oden, 2023, personal communication). These claims, in concert with other strategies, are integral to the contemporary coalescing of a reformist disposition and bipartisan program intended to reassert the status quo by coopting and neutralizing tenets of more radical projects that have arisen through decades-long organizing in communities, the international George Floyd uprisings of 2020, and the rise of mutual aid projects in the wake of these uprisings as well as the state abandonment that attended the Covid-19 pandemic. Defining these core premises can inform future work identifying and challenging them across reformist initiatives and technologies, as they show up in both implicit and explicit ways.

The first premise is that ‘crime’ can be understood as a ‘symptom’ without substantive investigation or intervention into what it might be a symptom of. Maps forecasting criminal ‘risk’ enable the category of ‘crime’ to operate as both the symptom (what signals a problem) and the disease (what causes the symptoms and could be present even in their absence). The software’s framework collapses the distinction between ‘symptom’ and ‘disease’ such that the solutions it proposes are unable to consider what systemic conditions actually shape the lives that come in contact with police through its outcomes. This works to foreclose perspectives that use a public health approach to understand community violence as the result of structural, spatial and economic inequities and frameworks that refuse an analytic of ‘crime’ in demands for resource redistribution and care infrastructures. Understandings of ‘crime as symptom,’ here, direct attention towards micro-infrastructural changes and away from locating root causes in systemic institutional abandonment and the consistent state prioritization of capital accumulation over human lives.

The second premise is that police are no longer tasked with policing potential ‘criminals’ but, following the turn to arenas of contagion discussed above, shift their focus to the spatial conditions that allow ‘crime’ to emerge. These software render individuals available for police contact through understanding them as vectors, in the epidemiological sense. The framework of risk management rearticulates policing as the staging of pre-emptive interventions into arenas of ‘communicability,’ where public life, as the space that arises only between and among individuals, is imagined as carrying within it the potential for its own undoing. In this vision, public life must be protected from public life for the sake of public life, and reformist perspectives are summoned through invoking collectivities as simultaneously precious and volatile. The potential of public life – a quality felt all the more poignantly in the long wake of the ongoing pandemic – is imagined as separable from the risks of public life in ways that re-assert the iterative processes of this separation as the proper purview of the state and its armed officers. ¹⁵ A brief, illustrative anecdote: Wendy Ethridge, director of analytic solutions for ResourceRouter, shared that one city saw the algorithm bring up churches as risk factors once they started providing services to people experiencing homelessness (2022, personal communication).

This reformist program takes an understanding of collective life as a collective resource, one that has informed much of the political organizing around building communities without police, and rescripts collective life as a resource only when ‘protected’ by the state from its own potential conflicts. In this paradigm, only peaceful protests should be sanctioned and only ‘innocent,’ victimized subjects should be exempted from policing, incarceration and police violence (Gilmore, 2017). To quote Jackie Wang, a political program rooted in “removing all elements of risk and danger reinforces a politics of reformism that often reproduces the existing social order” (2018). What I hope to work towards through parsing the architecture of ‘criminal risk’ are further tools for differentiating the threat imaginary the state works to impose upon communities from frameworks that take seriously the potential for violence in ways able to support transformation and ties of accountability that continue to bring people together in the face of risk (Pérez, 2018). The conditions for violence rendered by ‘criminal risk’ are steeped in the fantasies of banishment that inform the spatiality of imprisonment, augmented with a futuristic bent: the state can purify the conditions of public life before they are even violated. But the worlds that police and prison abolitionism enact are not purified worlds that shy away from the reality of the potential for harm. They are worlds that understand the resources of collective life to lie in its ability to address this potential as an element of community mitigated only by the ongoing, everyday tending of the conditions of care.