The idea that data and algorithms could help prevent crime before it happens has long held a powerful appeal, promising to make policing more efficient, more objective, and more effective by directing limited resources to where they are most needed. Beginning in the early years of the last decade, police departments across the United States and beyond adopted predictive policing tools, software that analyzes historical crime data to forecast where crimes are likely to occur or who is likely to be involved in them, in the hope that such forecasts could guide patrols and interventions to stop crime before it took place. The technology was marketed as a way to bring scientific rigor and neutrality to decisions that had previously relied on the intuition and discretion of officers, and for a time it spread rapidly, embraced by departments seeking to do more with less and by officials drawn to the promise of data-driven public safety.
In the years since, however, predictive policing has become one of the most contested applications of algorithms in public life, criticized for failing to deliver on its promises and, more seriously, for reinforcing and obscuring the very biases it was supposed to transcend. Investigations and studies have found that some prominent predictive policing tools were ineffective at actually predicting crime, and that they tended to direct policing disproportionately toward minority and already heavily policed communities, raising concerns that the technology was lending a veneer of scientific objectivity to discriminatory patterns of enforcement. These criticisms, combined with broader movements for police accountability and racial justice, have produced a wave of scrutiny, with some cities banning the technology, lawmakers questioning its funding, and researchers and advocates calling for fundamental reform of how, or whether, such tools should be used.
This article examines the effort to reform predictive policing and to mitigate its biases, discussing both the documented problems with these tools and the approaches being pursued to address historical bias concerns while improving the troubled relationship between police and the communities they serve. It explains what predictive policing is and the problems that have been documented, the mechanisms by which bias enters these systems, and the range of reform approaches, from transparency and bias auditing to community oversight, accountability frameworks, and outright bans. It weighs the considerations for communities, departments, and oversight bodies, and it grounds the discussion in documented cases and developments. The subject is genuinely contested, touching on difficult questions of public safety, fairness, and the proper role of technology in the exercise of state power, and the aim is to convey the issues clearly and fairly rather than to resolve debates on which reasonable people disagree.
Understanding Predictive Policing and Its Documented Problems
To engage with the reform of predictive policing, one must first understand what the technology is and the specific problems that investigations and research have documented. Predictive policing refers to the use of data analysis and algorithms to anticipate criminal activity, and it generally takes two broad forms. Place-based or location-based predictive policing analyzes historical crime data to forecast where crimes are likely to occur, identifying particular areas or times as high-risk and directing patrols accordingly, while person-based predictive policing attempts to identify particular individuals deemed likely to be involved in crime, either as potential offenders or victims, often by scoring them based on various factors. Both forms share the underlying premise that patterns in historical data can predict future crime, and both have been deployed by police departments with the aim of allocating resources more effectively and preventing crime before it occurs.
The appeal of predictive policing rested on the promise of objectivity and efficiency, the idea that data-driven forecasts could replace subjective and potentially biased human judgment with neutral, scientific analysis. Proponents argued that by directing policing to where the data indicated it was most needed, the technology could make limited resources go further, prevent crime more effectively, and remove the influence of individual officers’ biases from decisions about where and whom to police. This promise was attractive to departments under pressure to reduce crime with constrained budgets and to officials seeking to demonstrate a modern, evidence-based approach to public safety, and it drove the rapid adoption of predictive policing tools across many jurisdictions, with the technology presented as a way to bring the rigor of data science to the difficult work of preventing crime.
The documented reality, however, has fallen well short of these promises in ways that have generated serious concern. On the question of effectiveness, investigations have found that some widely used predictive policing tools performed poorly at actually predicting crime, with one analysis of tens of thousands of predictions made by a prominent tool finding that its forecasts were accurate only a tiny fraction of the time, raising fundamental doubts about whether the technology delivered the predictive value it claimed. Beyond ineffectiveness, a growing body of evidence indicated that members of Congress and researchers came to argue that predictive policing technologies did not reduce crime and instead worsened the unequal treatment of communities of color, with investigations finding that some tools directed policing disproportionately toward minority neighborhoods while leaving predictions largely absent for wealthier, whiter areas. These findings suggested that the technology, far from delivering neutral objectivity, was reproducing and amplifying existing disparities in policing.
The accumulation of these documented problems transformed predictive policing from a promising innovation into a deeply contested practice and set the stage for the reform efforts that are the subject of this article. The combination of evidence that the tools often did not work as claimed and concern that they reinforced discriminatory patterns of enforcement, lending those patterns a misleading appearance of scientific legitimacy, drew sustained criticism from civil liberties organizations, racial justice advocates, researchers, and increasingly lawmakers. This criticism gained force alongside broader movements demanding police accountability and an end to discriminatory policing, and it produced concrete responses, including bans in some jurisdictions, the discontinuation of particular programs, scrutiny of the federal funding that supported the technology, and calls to either fundamentally reform predictive policing or abandon it altogether. Understanding both the original promise of the technology and the specific, documented ways it has fallen short is essential to grasping the reform debate, which centers on whether and how the genuine aspiration to use data for public safety can be reconciled with the imperative to avoid reinforcing bias and injustice.
It is worth acknowledging at the outset that the underlying problem predictive policing sought to address is real and serious, which is part of why the technology was embraced despite the risks. Communities affected by violent crime, including many of the same communities later subjected to disproportionate algorithmic policing, have a genuine and urgent interest in effective public safety, and the desire to use every available tool to prevent harm is understandable and legitimate. Police departments operate with finite resources and face real pressure to reduce crime, and the prospect of directing those resources more intelligently held an honest appeal. Recognizing the legitimacy of these motivations matters, because the critique of predictive policing is not that the goal of preventing crime is illegitimate but that the particular means failed to achieve it fairly or even effectively, and because any constructive path forward must take seriously both the communities harmed by biased enforcement and the communities harmed by the crime that public safety efforts aim to prevent. Framing the debate as a simple contest between technology’s defenders and its critics obscures this shared underlying concern for safety, and a fair treatment of the subject must hold together the genuine need for public safety, the documented failures of these particular tools, and the difficult question of how the former might be pursued without repeating the latter.
How Bias Enters Predictive Policing Systems
To understand the reform of predictive policing, it is essential to understand how bias enters these systems, since the problem is not generally a matter of deliberate discrimination encoded into the algorithms but of subtler mechanisms by which existing inequities are absorbed and amplified. The core issue is that predictive policing tools learn from historical data about crime and policing, and that data reflects the patterns of a policing system that has itself been shaped by bias and disparity, so that the tools inherit and reproduce those patterns even when they are designed to be neutral. The mechanisms by which this occurs are important to grasp, because they determine what reform must address and reveal why simply removing explicit references to race or other protected characteristics is insufficient to make these systems fair.
The two subsections that follow examine the principal mechanisms by which bias enters predictive policing. The first concerns the problem of biased training data and the self-reinforcing feedback loops it creates, how data generated by biased policing practices, when used to predict future crime, produces predictions that direct more policing to the same places and people, generating more biased data in a cycle. The second concerns the specific problems of person-based targeting, how systems that score individuals as likely offenders or victims can produce disparate impacts and serious harms to those they target. Understanding both the data-driven feedback loops and the dynamics of individual targeting is necessary to grasp how bias operates in these systems and what reform efforts must confront.
Dirty Data and Feedback Loops
The foundational mechanism by which bias enters predictive policing is the use of historical crime data that is itself a product of biased policing, a problem sometimes referred to as dirty data. Predictive policing tools learn from records of past crimes and police activity, but these records do not neutrally reflect where crime actually occurs; rather, they reflect where police have looked for and recorded crime, which has been shaped by decades of enforcement patterns that concentrated policing in particular, often minority, communities. Because more policing in an area produces more recorded incidents regardless of the true underlying rate of crime, the data overrepresents heavily policed communities and underrepresents others, so that a tool trained on this data learns to associate those communities with crime not because crime is necessarily more prevalent there but because that is where policing and recording have been concentrated. The result is that the data on which the predictions are built embeds the biases of the policing that generated it.
This problem is compounded by the creation of self-reinforcing feedback loops, in which biased predictions lead to actions that generate more biased data, perpetuating and amplifying the original bias. When a tool predicts that crime is likely in a particular area based on historical data, police are directed to patrol that area more intensively, and because more policing produces more recorded incidents, the increased patrols generate more crime data for that area, which the tool then uses to predict even more crime there, drawing still more policing in a self-perpetuating cycle. This feedback loop means that the predictions can become self-fulfilling, directing escalating attention to the same communities regardless of the actual distribution of crime, and creating a spiral in which the tool’s outputs increasingly reflect its own past directions rather than any independent reality. The dynamic is particularly insidious because it operates automatically and can appear to validate the tool’s predictions, since the increased policing does indeed find more incidents, masking the circularity of the process.
The deeper difficulty is that this bias cannot be eliminated simply by removing explicit references to protected characteristics like race from the data or the model, because the bias is embedded in the patterns of the data itself. Even a tool that never considers race directly can produce racially disparate outcomes if it relies on data shaped by racially disparate policing, since factors like location, prior arrests, and other variables can serve as proxies that correlate with race and carry the embedded bias forward. This means that achieving genuine fairness requires confronting the bias in the underlying data and the patterns it reflects, a far more difficult task than removing explicit discriminatory factors, and it reveals why the promise of algorithmic objectivity was so misleading, since the apparent neutrality of the algorithm masked the deeply non-neutral character of the data it learned from. The recognition that predictive policing tools absorb and amplify the biases of the policing that generated their training data, through dirty data and feedback loops that resist simple technical fixes, is central to understanding both why these systems have produced biased outcomes and why reforming them is so challenging, since the problem reaches beyond the algorithms into the historical patterns of policing and the data those patterns produced.
Person-Based Targeting and Disparate Impact
A distinct and especially fraught form of predictive policing attempts to identify particular individuals deemed likely to be involved in crime, and this person-based targeting raises its own serious concerns about bias and harm. Rather than forecasting locations, person-based systems generate lists or scores of specific people, ranking them by their assessed likelihood of being involved in crime as offenders or victims based on factors such as their criminal history, their associations, and other data. The most prominent example was a major city’s effort to create a list scoring individuals by their risk of involvement in violence, assigning each person a numerical score intended to identify those most likely to be involved in a shooting. The premise was that focusing attention on the highest-risk individuals could prevent violence, but the practice of singling out specific people for heightened police attention based on algorithmic predictions raised profound questions about fairness, due process, and the consequences for those identified.
The concerns about person-based targeting center on the disparate impact it can produce and the serious harms it can inflict on the individuals targeted. Because these systems rely on data such as prior arrests and associations that reflect biased policing patterns, they can disproportionately identify members of minority communities, subjecting them to heightened scrutiny based on predictions derived from biased data. For the individuals placed on such lists, the consequences can be significant, since being labeled by an algorithm as likely to be involved in crime can lead to increased surveillance, more frequent police contact, and a presumption of dangerousness, all based on a prediction rather than any specific wrongdoing, raising questions about whether people are being treated as suspects based on statistical associations rather than their actual conduct. The opacity of these systems often left those affected unaware that they had been scored or why, denying them any meaningful ability to understand or challenge their inclusion, which compounded the due process concerns.
The documented experience of person-based predictive policing has been troubling enough that prominent programs were discontinued after scrutiny revealed both their questionable effectiveness and their concerning effects. A major city’s individual risk-scoring program, after operating for years, was decommissioned following criticism that included questions about its efficacy raised by oversight officials and concerns about its fairness and impact, illustrating a pattern in which person-based systems, subjected to examination, were found wanting and abandoned. The fundamental tension in person-based targeting is between the legitimate goal of focusing prevention efforts on those genuinely at risk and the danger of subjecting individuals to heightened policing based on biased predictions and statistical associations, treating them as suspects before any wrongdoing, in ways that can deepen the very mistrust and injustice that undermine effective policing. This tension, and the documented harms of person-based systems, has made individual targeting one of the most criticized forms of predictive policing and a focus of reform efforts that question whether such targeting can ever be made fair, or whether the harms to those wrongly identified and the corrosive effects on community trust outweigh any preventive benefit it might offer.
Approaches to Reform and Bias Mitigation
In response to the documented problems with predictive policing, a range of reform approaches has emerged, spanning technical measures to mitigate bias, governance and oversight mechanisms to ensure accountability, and in some cases the outright prohibition of the technology, and understanding these approaches clarifies the landscape of efforts to address the concerns. At one end are technical and procedural reforms that aim to make predictive policing fairer and more accountable while continuing to use it, including efforts to address the bias in training data, to test and audit systems for discriminatory outcomes, and to make the technology more transparent. At the other end are more fundamental responses that question whether the technology can be reformed at all, including moratoria and bans. The spectrum of approaches reflects the genuine disagreement about whether predictive policing can be made acceptable or should be abandoned.
Transparency is widely regarded as a foundational reform, since the opacity of predictive policing systems has been a major obstacle to accountability and to identifying and addressing bias. Advocates and emerging policies call for agencies to disclose information about the predictive tools they use, including summaries of the training data, the logic of the models, and the results of audits, so that the public, oversight bodies, and affected communities can understand and scrutinize how the systems work and what effects they have. Transparency requirements can extend to the vendors who supply the technology, mandating that they disclose how their systems function and submit to independent testing, addressing the problem that proprietary systems were often black boxes that even the departments using them did not fully understand. The principle is that systems affecting people’s lives and liberty should not operate in secret, and that openness is a prerequisite for the accountability and correction that fairness requires, making transparency a cornerstone of most reform proposals. The importance of transparency is underscored by the fact that nearly all of the documented problems with predictive policing came to light not through the routine operation of the systems but through external investigation, as journalists obtained and analyzed data, researchers examined outcomes, and oversight officials probed the programs, repeatedly finding flaws that the operating agencies had not disclosed or perhaps not themselves recognized. Without such scrutiny, the ineffectiveness and bias of these tools would have remained largely invisible, and the fact that it took determined outside investigation to reveal what was happening demonstrates how thoroughly the opacity of these systems shielded them from accountability, and why building transparency into their governance from the start, rather than relying on after-the-fact exposure, is so central to any meaningful reform.
Bias auditing and testing represent a second key technical reform, aiming to identify and measure discriminatory outcomes so they can be addressed. The approach involves systematically testing predictive policing systems for biased effects, examining whether their predictions and the policing they direct fall disproportionately on particular communities, and conducting these audits regularly to monitor for bias over time. Pre-deployment testing seeks to assess systems before they are used, while ongoing audits monitor their effects in operation, and the results can inform decisions about whether and how to use the technology. Frameworks for managing the risks of artificial intelligence, including those emphasizing fairness, accountability, and transparency, provide structured approaches to such testing, and policies have begun to mandate independent testing and impact assessments for predictive policing applications that affect people’s rights. The recognition that bias must be actively measured and monitored, rather than assumed away, underlies these auditing approaches, though they face the deep challenge that the bias often lies in the underlying data in ways that are difficult to fully correct.
Community oversight and governance, along with the more fundamental option of prohibition, complete the spectrum of reform. Many reform proposals emphasize involving affected communities in decisions about whether and how predictive policing is used, through community oversight boards or committees with genuine authority, sometimes including the power to halt the use of systems found to discriminate, reflecting the principle that the communities most affected should have a voice in the technology deployed upon them. Broader governance reforms include requirements for cost-benefit analyses, impact assessments, public feedback, and accountability mechanisms that allow the auditing, explanation, and correction of algorithmic decisions. At the most fundamental level, some jurisdictions have concluded that predictive policing cannot be adequately reformed and have banned it outright, with a number of cities prohibiting the technology, and lawmakers have questioned the federal funding that supports it, reflecting the view of some that the harms and the entanglement with biased data are so deep that the only adequate response is to stop using the technology altogether. The full range of reform approaches, from transparency and auditing through community oversight to outright bans, reflects the contested nature of the debate and the absence of consensus on whether predictive policing can be made acceptable, with different jurisdictions and advocates reaching different conclusions about how to balance the aspiration for data-informed public safety against the documented risks of bias and harm.
A further direction within the reform conversation concerns shifting the focus of any predictive tools away from individuals and toward environmental and situational factors that can be addressed without targeting people. Some researchers and practitioners have argued that if prediction is used at all, it should aim at identifying conditions, such as features of the physical environment or situational circumstances associated with crime, that can be changed through means other than aggressive enforcement, for instance by improving lighting, addressing neglected spaces, or directing social services rather than police to areas of need. This approach attempts to separate the analytical insight that data might offer from the punitive enforcement that has caused harm, redirecting any predictions toward interventions that support communities rather than subjecting them to heightened policing. Whether such approaches can avoid the data biases that plague enforcement-focused tools, and whether they represent a genuine improvement or merely a softer framing of the same surveillance logic, are themselves debated, but they illustrate that the reform conversation extends beyond making enforcement-oriented prediction fairer to questioning whether prediction should serve enforcement at all, or whether it might instead inform a broader, less punitive, and more community-centered conception of public safety. This reimagining of what data-informed public safety could mean represents one of the more ambitious strands of reform, moving beyond mitigating the harms of existing tools toward rethinking the purposes they serve.
Benefits and Challenges Across Stakeholders
The reform of predictive policing involves considerations for many parties, and a balanced assessment requires weighing the potential benefits of reform against its limitations and the unresolved questions across communities, police departments, and oversight bodies. Reform efforts aim to reduce bias and harm, to restore accountability and public trust, and to ensure that any use of these tools serves rather than undermines justice, yet they face deep challenges, including the difficulty of fully eliminating bias, disagreement about whether the technology can be salvaged, and the broader tensions between public safety and civil liberties. The debate is genuinely contested, and a clear-eyed view must consider the perspectives and concerns of the different stakeholders involved.
The analysis below organizes these considerations by stakeholder and by category, first examining the potential benefits that reform offers to communities, departments, and oversight, then turning to the risks, limitations, and unresolved questions that the reform effort confronts. Keeping these perspectives distinct helps engage seriously with a difficult and divisive issue, presenting the considerations on different sides fairly rather than presuming a single correct answer to questions on which thoughtful people, including those genuinely committed to both public safety and justice, hold differing views.
Benefits for Communities, Departments, and Oversight
For communities, particularly those that have borne the brunt of biased policing, the central benefit of reform is the prospect of being protected from the harms that biased predictive policing has inflicted and of having a voice in the technologies deployed upon them. Reforms that mitigate bias, increase transparency, and provide community oversight offer the possibility that policing technology will not reinforce the disproportionate scrutiny and enforcement that have damaged the relationship between police and minority communities, and that the communities affected will be able to understand, scrutinize, and influence how such tools are used. For residents who have experienced or feared the consequences of being subjected to algorithmically driven policing, reforms that constrain or eliminate biased systems, that make them accountable, and that give communities genuine input represent a meaningful protection of their rights and dignity, and a step toward policing that treats them fairly rather than as statistical risks. The restoration of fairness and the provision of a voice are benefits of real significance to communities long disadvantaged by the patterns these technologies reflected.
For police departments and the broader goal of effective public safety, reform offers the possibility of rebuilding the community trust that is essential to legitimate and effective policing, and of avoiding the harms that discredited the technology. Effective policing depends heavily on the trust and cooperation of the communities served, and the use of biased technology that alienated those communities undermined that trust, so reforms that address the bias and restore accountability can help repair the relationship between police and the public, which is itself important for public safety. Departments also have an interest in not deploying tools that are ineffective or that expose them to legal and reputational risk, so reforms that ensure any technology used is genuinely effective, fair, and defensible serve the legitimate interests of policing as well as the communities it serves. The recognition that public safety depends on legitimacy and trust, and that biased and ineffective technology undermines both, aligns the interests of thoughtful departments with the goals of reform, suggesting that fairer and more accountable approaches can serve genuine public safety rather than oppose it.
For oversight bodies, lawmakers, and the broader project of democratic accountability over the use of state power, reform offers mechanisms to ensure that powerful technologies affecting people’s lives and liberty are subject to scrutiny, control, and correction. The transparency, auditing, oversight, and accountability mechanisms at the heart of reform give legislators, inspectors, courts, and the public the means to examine how these technologies are used, to identify and address their harms, and to make democratic decisions about their deployment, addressing the troubling reality that predictive policing often operated with little oversight or accountability. For a society governed by the rule of law, the principle that the exercise of state power, especially the coercive power of policing, should be transparent and accountable is fundamental, and reforms that bring algorithmic policing tools within the reach of democratic scrutiny serve this principle. The development of frameworks and requirements for the governance of these technologies represents an effort to extend the accountability that should attend the use of state power into the domain of algorithmic decision-making, a benefit that serves the broader health of democratic governance and the protection of rights against the unaccountable exercise of power.
Risks, Limitations, and Unresolved Questions
The most fundamental challenge facing reform is the deep difficulty of fully eliminating bias from systems that learn from biased data, which raises the question of whether technical reform can ever make predictive policing genuinely fair. Because the bias is embedded in the historical data and the patterns of policing that generated it, and because it persists through proxies even when explicit discriminatory factors are removed, mitigating it is far harder than it might appear, and some argue that it cannot be adequately accomplished, that any system trained on the data of a biased policing system will reproduce that bias in some form. This is why many advocates have concluded that reform is insufficient and that the technology should be abandoned, while others maintain that careful attention to data, rigorous auditing, and strong oversight can reduce bias to acceptable levels. The unresolved technical question of whether bias can be sufficiently mitigated, or whether it is inherent in the enterprise of learning from biased data, lies at the heart of the disagreement about whether predictive policing can be reformed or must be rejected.
Beyond the technical difficulties lie deeper questions about effectiveness, legitimacy, and the proper role of prediction in policing that reform does not fully resolve. Evidence that some predictive policing tools were ineffective at predicting crime raises the question of whether the technology offers enough genuine value to justify its use even if its biases could be addressed, since a tool that does not work well is hard to defend regardless of its fairness. There are also profound questions about the legitimacy of policing people based on predictions and statistical associations rather than their conduct, particularly in person-based systems, that touch on fundamental principles of justice and individual responsibility and that technical reforms do not answer. The tension between the legitimate goal of preventing crime and the dangers of subjecting people to heightened policing based on algorithmic predictions reflects deeper disagreements about the proper balance between public safety and individual rights, disagreements that the reform of predictive policing engages but cannot settle, since they involve values about which a democratic society must deliberate.
The remaining challenges concern the practical and political difficulties of implementing reform and the broader context in which the debate occurs. Implementing meaningful transparency, auditing, and oversight requires resources, expertise, and political will that may be lacking, and the proprietary nature of vendor systems, the resistance of some departments, and the complexity of the technology can all impede effective reform. The debate is also politically charged, occurring within broader and contentious disputes about policing, crime, and racial justice, which can make reasoned deliberation difficult and can lead to outcomes driven by political dynamics rather than careful assessment of evidence and values. There is, moreover, the risk that reforms could provide a false reassurance, allowing biased or ineffective systems to continue under the cover of nominal accountability measures that do not genuinely address the underlying problems. These challenges do not mean that reform is futile, and meaningful progress has been made in some places through transparency, oversight, and the discontinuation of harmful programs, but they make clear that the reform of predictive policing is a difficult and contested undertaking, that the fundamental questions of whether the technology can be made fair and whether it should be used at all remain genuinely unresolved, and that the path forward requires grappling honestly with hard questions of fairness, effectiveness, and the proper limits of algorithmic power in the exercise of justice.
Real-World Implementations and Measured Outcomes
The trajectory of predictive policing and its reform is illustrated by documented cases that span the rise and fall of prominent tools, the discontinuation of controversial programs, and the policy responses that have constrained the technology. Three examples in particular, drawn from a leading place-based tool, a major city’s person-based program, and the wave of bans and policy reforms, demonstrate how the documented problems with predictive policing led to concrete consequences and reform efforts. Each is grounded in documented developments, showing the real-world course of a technology that moved from enthusiastic adoption to serious scrutiny and, in many cases, abandonment or constraint.
The tool known as PredPol, later rebranded as Geolitica, exemplifies the place-based approach and the trajectory from prominence to discredit and discontinuation. One of the most widely adopted predictive policing tools, it analyzed historical crime data to forecast where crimes were likely to occur and direct patrols accordingly, and it was for years a leading example of the technology. Investigations, however, documented serious problems, with an analysis of tens of thousands of its predictions finding that they were accurate only a tiny fraction of the time, raising fundamental doubts about its effectiveness, and earlier reporting indicating that it directed policing disproportionately toward minority neighborhoods while leaving wealthier, whiter areas largely absent from its predictions. These documented failures of both effectiveness and fairness contributed to growing criticism, and the company behind the tool ceased operations at the end of 2023, with its assets reportedly acquired by another firm. The arc of this prominent tool, from a leading exemplar of predictive policing to a discredited and discontinued product, illustrates how investigation and evidence transformed the technology’s reputation and led to concrete consequences.
A major city’s individual risk-scoring program exemplifies the person-based approach and its discontinuation following scrutiny. The program created a list that scored individuals by their assessed risk of involvement in violence, assigning numerical scores intended to identify those most likely to be involved in a shooting as either offender or victim, and it operated for years as one of the most prominent examples of person-based predictive policing. The program drew sustained criticism over questions about its effectiveness, raised by oversight officials, and over concerns about its fairness and its effects on the individuals it identified, who could be subjected to heightened police attention based on algorithmic predictions. After years of operation and mounting scrutiny, the program was decommissioned, reflecting a judgment that its problems outweighed its value. The trajectory of this person-based program, from an ambitious effort to predict and prevent violence to a discontinued program criticized for its efficacy and its impact, illustrates the particular difficulties of individual targeting and the pattern in which person-based systems, subjected to examination, were found wanting and abandoned. Independent reviews of the program also raised the troubling finding that a large share of the people on its list had never been arrested for or charged with a serious offense, yet had nonetheless been flagged by the algorithm, underscoring how such systems can sweep in individuals on the basis of statistical association rather than conduct. This kind of finding, emerging only because oversight bodies and researchers examined the program, exemplifies both why scrutiny matters and why the opacity of these systems was so dangerous, since without external examination the basis on which people were being singled out for police attention remained hidden from the public and often from the individuals themselves.
The wave of municipal bans and policy reforms exemplifies the more fundamental responses that have constrained predictive policing at the level of policy. A number of jurisdictions concluded that the technology could not be adequately reformed and prohibited it, with one California city becoming the first in the nation to ban predictive policing, a decision notable because that city had been an early home of the technology, and the ban, supported by a coalition of civil liberties and racial justice organizations, reflected the view that the technology fostered discriminatory policing. At the federal level, members of Congress publicly called for halting the funding of predictive policing programs, citing evidence that the technology did not reduce crime and worsened unequal treatment, and broader governance efforts emerged to require transparency, independent testing, and impact assessments for predictive policing applications affecting people’s rights, alongside frameworks for managing the risks and biases of artificial intelligence. These policy responses, ranging from outright bans through funding scrutiny to governance requirements, demonstrate the concrete reforms that the documented problems with predictive policing have produced, and they reflect the contested judgment, reached differently in different places, about how to respond to a technology whose promise of data-driven public safety collided with evidence of ineffectiveness and bias. Taken together, these cases illustrate the real-world trajectory of predictive policing and its reform, showing a technology that moved from enthusiastic adoption to serious scrutiny, and that has been variously discontinued, banned, and subjected to new requirements as communities and policymakers have grappled with its documented harms.
Final Thoughts
The story of predictive policing and its reform is a cautionary tale about the promise and peril of applying algorithms to the exercise of state power, and about the danger of mistaking the appearance of objectivity for its substance. The technology was embraced on the promise that data could make policing more neutral and effective, removing human bias, but the documented reality revealed that algorithms trained on the data of a biased policing system inherit and amplify its biases, lending those biases a misleading veneer of scientific legitimacy while too often failing to deliver the predictive value they claimed. The recognition that the apparent neutrality of the algorithm masked the non-neutral character of the data it learned from is the central lesson of this experience, with implications far beyond policing for any effort to use historical data to guide consequential decisions about people’s lives.
The broader significance of this reckoning lies in what it teaches about the relationship between technology, power, and justice. The experience demonstrates that technology is not neutral, that systems built on biased data and deployed within unjust structures will reproduce and may amplify that injustice, and that the promise of algorithmic objectivity can obscure rather than eliminate bias, making it harder to see and to challenge. It also demonstrates the importance of transparency, accountability, and the voice of affected communities, since it was investigation, scrutiny, and advocacy that exposed the problems and drove reform. The intersection of technology and social responsibility is starkly visible here, in the recognition that deploying powerful predictive tools in the coercive domain of policing demands the highest standards of fairness, accountability, and democratic control, and that the failure to meet those standards can deepen rather than remedy long-standing injustices.
The honest assessment must acknowledge that the fundamental questions remain genuinely unresolved and contested. Whether predictive policing can be reformed to be acceptably fair, or whether the bias embedded in its data makes it irredeemable, is a question on which thoughtful people disagree, as is the deeper question of whether policing people based on predictions and statistical associations can ever be reconciled with principles of justice and individual responsibility. These are not merely technical questions but matters of values about which a democratic society must deliberate, weighing the legitimate goal of protecting communities against the dangers of bias, the erosion of civil liberties, and the treatment of people as risks rather than individuals. The reform of predictive policing engages these questions but cannot settle them, and the different conclusions reached in different places reflect the absence of consensus and the genuine difficulty of the issues.
The most constructive understanding is that the predictive policing experience, for all its troubling history, has advanced the collective recognition of what the responsible use of algorithms in consequential domains requires, even as it has left the hardest questions open. The development of transparency requirements, bias auditing, community oversight, and accountability frameworks represents genuine progress in establishing that powerful technologies affecting people’s lives must be subject to scrutiny and control, lessons that extend well beyond policing. Whatever the future of algorithmic public safety, the enduring lesson is that the pursuit of public safety through technology must be inseparable from the pursuit of justice, fairness, and accountability, and that any tool that sacrifices the latter in the name of the former ultimately undermines both. The ongoing effort to grapple with these questions, with attention to the communities most affected, is the difficult but necessary work of ensuring that the power of technology serves rather than subverts the cause of just and equitable public safety.
FAQs
- What is predictive policing?
Predictive policing is the use of data analysis and algorithms to anticipate criminal activity. It generally takes two forms: place-based predictive policing analyzes historical crime data to forecast where crimes are likely to occur and direct patrols there, while person-based predictive policing attempts to identify particular individuals deemed likely to be involved in crime as offenders or victims, often by scoring them. Both share the premise that patterns in historical data can predict future crime, and both have been deployed by police departments aiming to allocate resources more effectively and prevent crime before it occurs. - Why was predictive policing originally adopted?
It was embraced on the promise of objectivity and efficiency, the idea that data-driven forecasts could replace subjective, potentially biased human judgment with neutral, scientific analysis. Proponents argued it could make limited resources go further, prevent crime more effectively, and remove individual officers’ biases from decisions about where and whom to police. This was attractive to departments under pressure to reduce crime with constrained budgets and to officials seeking to demonstrate a modern, evidence-based approach, and it drove rapid adoption across many jurisdictions during the last decade. - What problems have been documented with predictive policing?
Investigations and studies have documented two main problems. On effectiveness, an analysis of tens of thousands of predictions by a prominent tool found they were accurate only a tiny fraction of the time, raising doubts about whether the technology delivered real predictive value. On fairness, evidence indicated that some tools directed policing disproportionately toward minority neighborhoods while leaving wealthier, whiter areas largely absent from predictions, and lawmakers and researchers argued the technology did not reduce crime and worsened the unequal treatment of communities of color, reproducing rather than eliminating disparities. - What is dirty data and why does it matter?
Dirty data refers to historical crime data that is itself a product of biased policing. Crime records do not neutrally reflect where crime occurs but where police have looked for and recorded it, shaped by enforcement patterns that concentrated policing in particular communities. Because more policing produces more recorded incidents regardless of the true crime rate, the data overrepresents heavily policed communities. A tool trained on this data learns to associate those communities with crime, embedding the biases of the policing that generated the data, which is why removing explicit references to race is insufficient to make these systems fair. - What is a feedback loop in predictive policing?
A feedback loop is a self-reinforcing cycle in which biased predictions lead to actions that generate more biased data. When a tool predicts crime in an area based on historical data, police patrol it more intensively, and because more policing produces more recorded incidents, those patrols generate more crime data for the area, which the tool uses to predict even more crime there, drawing still more policing. This makes the predictions self-fulfilling, directing escalating attention to the same communities regardless of the actual distribution of crime, and it can appear to validate the predictions even as it masks the circularity of the process. - What concerns surround person-based predictive policing?
Person-based systems generate lists or scores ranking specific people by their assessed likelihood of involvement in crime, raising profound concerns. Because they rely on data like prior arrests that reflect biased policing, they can disproportionately identify members of minority communities. For those targeted, the consequences can include increased surveillance, more frequent police contact, and a presumption of dangerousness based on a prediction rather than any specific wrongdoing, raising questions about due process and whether people are treated as suspects based on statistical associations. The opacity of these systems often left affected individuals unaware they had been scored or able to challenge it. - How can transparency help reform predictive policing?
Transparency is regarded as a foundational reform because the opacity of these systems obstructed accountability. Reforms call for agencies to disclose information about the tools they use, including summaries of training data, model logic, and audit results, so the public, oversight bodies, and affected communities can scrutinize how the systems work and what effects they have. Transparency can extend to vendors, requiring them to disclose how systems function and submit to independent testing. The principle is that systems affecting people’s liberty should not operate in secret, and that openness is a prerequisite for the accountability and correction that fairness requires. - What is bias auditing?
Bias auditing involves systematically testing predictive policing systems for discriminatory outcomes, examining whether their predictions and the policing they direct fall disproportionately on particular communities, and conducting these audits regularly to monitor for bias over time. Pre-deployment testing assesses systems before use, while ongoing audits monitor their effects in operation, informing decisions about whether and how to use the technology. Frameworks for managing AI risks that emphasize fairness and accountability provide structured approaches, though auditing faces the deep challenge that the bias often lies in the underlying data in ways difficult to fully correct. - Have any places banned predictive policing?
Yes. A number of jurisdictions concluded the technology could not be adequately reformed and prohibited it. One California city became the first in the nation to ban predictive policing, a decision notable because it had been an early home of the technology, and the ban was supported by a coalition of civil liberties and racial justice organizations. At the federal level, members of Congress have called for halting funding of predictive policing programs, citing evidence the technology did not reduce crime and worsened unequal treatment, reflecting the view of some that the harms are too deep for reform to address. - Can predictive policing be made fair, or should it be abandoned?
This is genuinely unresolved and contested. Because bias is embedded in the historical data and persists through proxies even when explicit discriminatory factors are removed, some argue it cannot be adequately mitigated and the technology should be abandoned, while others maintain that careful attention to data, rigorous auditing, and strong oversight can reduce bias to acceptable levels. There are also deeper questions about effectiveness and about the legitimacy of policing people based on predictions rather than conduct. These are matters of values about which a democratic society must deliberate, and different jurisdictions have reached different conclusions, from outright bans to governance requirements.