Introduction

Predictive policing uses computer systems and algorithms to analyze large datasets, including historical crime data, to inform decisions about where to deploy police and to identify individuals who are more likely to commit or be victims of crime. These tools have been increasingly used by police departments across the United States, yet their usage remains highly contentious. Marketed as objective, efficient, and data-driven tools, these systems often operate under the premise that algorithmic prediction can guide proactive policing and reduce crime. Yet the rise of AI and predictive policing forces a fundamental constitutional question: can statistical probability ever satisfy the Fourth Amendment’s requirement of individualized suspicion?

This paper argues that predictive policing represents a constitutional category error. By collapsing the distinction between probabilistic forecasts and individualized suspicion, predictive policing redefines “probable cause” as “predictable cause,” thereby undermining the Fourth Amendment’s qualitative standard for evidence and for reasoned police judgment. This doctrinal shift directly conflicts with the Supreme Court’s Fourth Amendment jurisprudence. In Terry v. Ohio (1968), the Court held that reasonable suspicion must be grounded in “specific and articulable facts” tied to the individual.^[1] Furthermore, Illinois v. Gates (1983) reaffirmed that the “totality of the circumstances” standard is a holistic, qualitative judgment made by officers and courts.^[2] Predictive algorithms, by contrast, operate on generalized patterns and correlations, not individualized evidence or officer-observed conduct.

Moreover, predictive policing also delegates core law-enforcement discretion to private, proprietary algorithms, creating an accountability vacuum analogous to the nondelegation problem in administrative law. While proponents claim that predictive algorithms enhance efficiency and objectivity, emerging empirical research shows that these systems often reproduce historical policing biases, lack transparency, and expand state surveillance without the individualized justification the Constitution demands.

This not only undermines the Fourth Amendment’s qualitative standard for suspicion but also delegates police authority to unaccountable actors who privately develop algorithms, similar to the nondelegation problem in administrative law. Section I introduces predictive policing and its rise in U.S. law enforcement. Section II defines predictive policing and reviews major case studies, including Los Angeles, New York, and Chicago. Section III outlines arguments in favor, focusing on efficiency and objectivity, while Section IV presents counterarguments on bias, opacity, and constitutional risks. Section V analyzes how predictive policing distorts Fourth Amendment doctrine, referencing Terry v. Ohio and Illinois v. Gates. Section VI explores algorithmic delegation and accountability gaps. Section VII proposes policy and doctrinal reforms, and Section VIII concludes by calling for renewed judicial clarity on individualized suspicion in the algorithmic era.

I. Overview of Predictive Policing

Predictive policing has its roots in the rise of data-driven policing techniques such as CompStat, which evolved into algorithmic systems that promise efficiency and precision. It spread quickly in the 2010s, with “more than 150 police departments nationally deploying predictive policing analytics.”^[3] A Department of Justice (DOJ) report also stated that 70% of agencies planned to implement or increase the use of predictive policing technology within five years.^[4]

As Andrew Ferguson notes, predictive policing “rests on the insight that certain aspects of the physical and social environment encourage quite predictable acts of criminal wrongdoing.”^[5] Police departments frame these systems as modern, objective extensions of traditional hotspot policing. Predictive policing systems generally fall into two categories: place-based models and person-based models. Place-based systems such as PredPol generate “hot spots” using historical enforcement data—not actual crime incidence. Because police disproportionately patrol and arrest in minority and low-income neighborhoods, these systems reproduce and intensify those racialized enforcement patterns.^[6] By contrast, person-based systems evaluate individuals directly. Chicago’s Strategic Subject List (commonly called the “Heat List”) attempted to algorithmically assess the likelihood that specific residents would commit or be victims of gun violence based on social network analysis and prior police interaction.^[7] These approaches extend the logic of profiling into data science, shifting suspicion from observable conduct to probabilistic inference. Ferguson explains that predictive policing has evolved through three major stages.^[8] Early systems focused on predicting locations of property crimes (like burglaries). The next wave expanded to forecasting violent crime in specific areas. The most recent stage attempts to predict which individuals are likely to commit—or be victimized by—crime. This progression reflects a broader trend: predictive tools are becoming increasingly granular and tied to assessments of individual people rather than places.

Case studies in the United States’ biggest cities demonstrate mixed results from the implementation of predictive systems. The Los Angeles Police Department (LAPD) has partnered with academics to predict property crime hotspots, deploying officers in areas projected to experience spikes in crime beginning in 2010.^[9] In New York City, the New York Police Department (NYPD)’s data-driven policing began with CompStat in the 1990s to “compile information on crimes… and spot emerging crime patterns.”^[10] In 2016, the NYPD launched Patternizr, a machine-learning tool to detect patterns across robbery and larceny reports. Molly Griffard, a law professor at NYU, describes Patternizr as a tool “returning a report listing ten potentially related crimes … scored between 0-1” for similarity and highlights NYPD’s widely publicized case where Patternizr helped “crack the case of the needle-wielding shoplifter.”^[11] Meanwhile, Chicago’s “Heat List” attempted to predict individuals likely to become perpetrators or victims of gun violence, identifying hundreds of residents for targeted interventions. The strategy’s framing seemed to mirror “a public health approach … [targeting] both potential shooting victims and offenders.”^[12]

The core statistical premise is that past crime predicts future crime—not actually “predicting a particular crime but predicting an elevated risk of crime based on pre-determined … factors.”^[13] Scholars caution, however, that because these systems rely on historical policing data, concerns about bias and feedback effects arise;^[14] these issues are explored in Section IV. These systems differ in form, but all share the same constitutional flaw: they substitute probabilistic pattern recognition for individualized suspicion.

II. Arguments in Favor

Proponents of predictive policing argue that algorithmic forecasting represents a major advancement in the efficiency, accuracy, and objectivity of modern law enforcement, though these claims ultimately fail to address the Fourth Amendment’s constitutional demands. Supporters emphasize that predictive policing enables police to “do more with less”^[15] by helping departments preposition officers in areas of anticipated criminal activity rather than relying on broad, undirected patrols. Ferguson notes that early implementations in Los Angeles and Santa Cruz were driven by the promise of cost-effective policing during periods of shrinking municipal budgets, with predictive systems advertised as tools that optimize increasingly scarce or limited resources.^[16]

Police chiefs promoting predictive policing depict this as a shift from reactive to proactive policing, enabling departments to cover more ground with fewer officers and to intervene before crime escalates. Empirical results from early pilot programs are frequently cited: the Santa Cruz Police Department reported a 4–11% drop in property crime in the first six months after adopting its predictive policing algorithm.^[17] Although preliminary, these statistics are regularly invoked by advocates as evidence that predictive models yield measurable real-world benefits. Ferguson describes how Santa Cruz’s algorithm flagged micro-locations (500×500 ft) tied to meaningful decreases in burglaries and thefts.^[18] These success stories helped fuel a wave of adoption across cities such as Memphis, Chicago, Dallas, and Minneapolis. Griffard likewise notes that the NYPD promoted Patternizr as a “new, effective… recommendation engine” that could assist investigators in linking related crimes more quickly than manual review.^[19]

Another key argument is that algorithmic guidance can reduce reliance on individual officer hunches—long criticized for allowing racial bias and arbitrary decision-making. Ferguson frames predictive tools as more “objective” than traditional hot-spotting because they rely on mathematically derived probability models rather than discretionary perceptions of danger.^[20] This line of reasoning suggests that when predictive alerts function like standardized “tips” about high-risk locations, they can guide policing in a more even-handed and consistent manner. Similarly, Griffard notes that Patternizr’s developers explicitly designed the NYPD’s system to minimize subjective inputs, claiming that removing “sensitive suspect attributes” (e.g., race, gender) and coarsening location data helps promote neutrality in pattern detection.^[21]

Proponents interpret these design choices as evidence that data-driven policing can be insulated—at least in part—from the biases that affect human decision-makers. Predictive policing is also presented as the logical extension of decades of criminological research on how crimes cluster in space and time. Ferguson points out that predictive models are rooted in theories such as near-repeat theory, routine activity theory, and crime-and-place analysis, all of which show that certain crimes (especially property crimes) tend to recur near prior incidents within predictable time windows.^[22]

Supporters also point to the rapid growth of predictive policing vendors—such as PredPol, Palantir, and HunchLab—as evidence of strong institutional confidence in the technology’s potential. What began as a limited academic experiment quickly evolved into a multimillion-dollar enterprise supported by federal grants and national law-enforcement conferences. For proponents, this commercial and governmental investment signals that predictive policing is not a speculative or fringe idea but an innovation that major agencies view as central to the future of public safety.

 III. Arguments Against Predictive Policing

The core statistical premise is that past crime predicts future crime—not actually “predicting a particular crime, but predicting an elevated risk of crime based on pre-determined … factors.”^[23] However, predictive policing has become one of the most controversial developments in contemporary criminal law, with the most common criticisms being systemic bias and “dirty data”—inaccurate, incomplete, or inconsistent data, particularly within computer systems or databases. Since algorithms are trained on historically biased data, this reproduces racial disparities in policing.  Manipulating crime statistics to meet quotas or produce ambitious crime-reduction results can also lead to increased policing in the neighborhoods where those statistics are concentrated. Some critics have labeled predictive policing as a form of “tech-washing,” allowing racially biased policing methods to seem acceptable in society due to the appearance of objectivity.^[24] Although proponents frame algorithmic systems as neutral, efficient, or data-driven, the empirical and legal evidence reveal a very different reality.

As Lum and Isaac write: “Police records do not measure crime. They measure some complex interaction between criminality, policing strategy, and community–police relations.”^[25] Their Oakland case study evaluates how biased police data sets are by combining a demographically representative synthetic population of Oakland with population-based drug use estimates to compare underlying drug use with police records. The results show that drug crimes known to police systematically misrepresent the true distribution of drug use in the population. While drug use was “relatively uniform across the city,” arrests were concentrated almost exclusively in historically non-white and low-income neighborhoods.^[26] When the PredPol algorithm was applied to this data, “Black people would be targeted at roughly twice the rate of whites,” despite equivalent estimated drug-use patterns.^[27] This demonstrates that predictive systems do not predict crime—they predict police behavior, embedding past enforcement disparities into future action.

Rotaru et al. likewise find that event-level models encode a “signature of enforcement bias.”^[28] Their event-level crime model revealed that models were susceptible to “perturbations caused by policing intensity rather than actual crime-pattern changes.”^[29] Across eight major metropolitan areas, the researchers found that predictive models were strongest at forecasting where police would go, not where crime would occur. As a result, these findings undermine any claim that predictive policing increases fairness. Instead, predictive models create feedback loops: the more heavily a neighborhood is policed, the more “crime” appears in the data, the more predictions target it, and the cycle repeats—confirming and amplifying historic racialized over-policing.

Opacity compounds these issues. Predictive policing tools are almost always proprietary, developed by private vendors such as PredPol, Palantir, or IBM. This means the algorithmic logic that directs state coercive power is hidden from courts, defendants, and the public. The Yale Media and Freedom Information Access (MFIA) Clinic emphasizes that information laws have serious shortcomings with respect to algorithmic accountability because companies claim that algorithms are “trade secrets that cannot be disclosed to the public,”^[30] despite the algorithms operating as black boxes. This makes it difficult, if not impossible, to assess the accuracy and fairness of their performance. Compare this with administrative procedures to promote the accountability of human decisions, such as freedom of information laws, due process protections, and appeals procedures.

This black-box structure makes it impossible for defendants to challenge algorithmic suspicion. Yet predictive systems still shape encounters that may escalate into stops, searches, or arrests. Griffard’s empirical evaluation of the NYPD’s Patternizr reinforces this concern. She finds that Patternizr’s pattern matching process is opaque and reproducibly inconsistent, with significantly different outputs depending on human analyst decisions and data-selection choices.^[31] According to the Brennan Center, after a lengthy legal battle, the NYPD finally disclosed some documentation about its use of algorithms and predictive policing software.^[32] However, the NYPD remains reluctant to disclose the source of their data sets. Despite claiming to not use enforcement data (such as arrest data) for predictive policing purposes, the department does not keep audit logs of the creation, access, and storage of the predictions it generates.^[33] This makes it difficult for independent auditors or policymakers to evaluate these tools properly. Similarly, in Los Angeles, major details about predictive policing only emerged after years of activism demanding more information from the LAPD about the nature of its operations.^[34] This concern reflects a shift in the locus of discretion, with accountability shifting from the police chief to the algorithm designer. This raises a structural constitutional issue, as the state is delegating core policing authority to actors beyond public oversight.

 IV. Doctrinal Layer

Even if predictive policing were unbiased and transparent—which it is not—it faces a doctrinal problem: the Fourth Amendment requires individualized suspicion, not statistical generalizations. Predictive systems do not produce evidence tied to a specific person or specific act. They produce probabilistic area-based forecasts.

Supreme Court precedent makes clear that this equivalence is impermissible. As seen in Terry v. Ohio (1968), reasonable suspicion requires specific, observed facts. In Terry, the Court required that reasonable suspicion be based on “specific and articulable facts”^[35] about a person’s conduct, not on general profiles or inferences drawn from broad trends. The Court emphasized that a stop must be justified by the facts available to the officer at the moment of the seizure, and those facts must “warrant a man of reasonable caution” to suspect that this individual is engaged in crime.^[36] Predictive algorithms provide no specific, individualized, or moment-specific facts about any person. They supply only statistical risk scores, which are the exact kind of generalized, unparticularized hunch Terry prohibits. Treating a probabilistic risk score as if it were a present-tense fact collapses Terry’s objective standard into what the Court expressly forbids: “intrusions… based on nothing more substantial than inarticulate hunches.”^[37] Terry also makes clear that constitutional protections apply most stringently in precisely the contexts where predictive policing is deployed. The Court emphasized that the Fourth Amendment “protects people, not places,”^[38]  and applies fully to the citizens on the streets. Predictive policing tools overwhelmingly guide officers’ actions in public spaces—hotspots, neighborhoods, intersections—yet the individuals encountered there retain full Fourth Amendment protections. When officers rely on algorithmic risk to initiate encounters, they trigger constitutional scrutiny because “whenever a police officer accosts an individual and restrains his freedom to walk away, he has ‘seized’ that person.”^[39] And any resulting frisk is a “serious intrusion upon the sanctity of the person… not to be undertaken lightly.”^[40] Crucially, Terry authorizes such intrusions only when the officer has reason to believe that this individual is “armed and dangerous,” grounding suspicion in specific, situational facts.^[41] Predictive policing flips this logic entirely: individuals become the focus of police attention not because of their observed conduct but because they fall within a statistically “high-risk” group or geographic area. Area-level or group-level predictions are not—and cannot be—“specific and articulable facts” about an individual’s behavior. They are, at best, correlations about spaces, populations, or past enforcement patterns. As scholars like Lum & Isaac and Rotaru et al. show, these correlations often reflect “dirty data”^[42] and enforcement bias, meaning risk scores frequently amount to formalized, data-driven hunches. Predictive policing, therefore, violates Terry at its core: it replaces individualized, moment-specific observation with statistical generalization, converting the constitutional standard of reasonable suspicion into an algorithmic approximation.

While Terry enshrines individualized suspicion at the street level, Illinois v. Gates (1983) illustrates how probable cause requires contextual, qualitative judgment. In Gates, the Court rejected rigid or quantitative tests for probable cause, emphasizing that the Fourth Amendment requires a “totality of the circumstances” analysis grounded in the quality, not the quantity, of the evidence. This cuts directly against the logic of predictive policing. Gates rejects “rigid” or “quantitative” tests for probable cause.^[43] Instead, it argues that probable cause is a “practical, nontechnical conception” assessed through a “fluid concept—turning on the assessment of probabilities in particular factual contexts—not readily, or even usefully, reduced to a neat set of legal rules.”^[44] Probable cause, the Court explains, does “not deal with hard certainties, but with probabilities… [and] practical considerations of everyday life on which reasonable and prudent men act,”^[45] emphasizing that the inquiry is fundamentally human and situational rather than mathematical. The Court emphasizes that probable cause should not be “reduced to a neat set of legal rules,”^[46] and explicitly warns that “finely tuned standards” or “numerically precise” thresholds have “no place” in the magistrate’s decision.^[47] Predictive policing does exactly what Gates forbids: it translates probable cause and reasonable suspicion into finely tuned statistical thresholds, operationalized through an algorithm that encodes rigid, pre-set weights and rules. Instead of evaluating a suspect based on “particular factual contexts”^[48] and a fact-rich, individualized narrative, officers are invited to rely on abstract risk scores and hotspot outputs that strip away context and replace it with automated correlation. In doctrinal terms, predictive policing inverts the Gates standard. Where Gates insists on a holistic, circumstance-specific judgment grounded in human reasoning, predictive systems reduce the probable cause inquiry to algorithmic suspicion and substitute a numerical output for the “totality of the circumstances”^[49] analysis. Together, Terry and Gates reject the very logic on which predictive policing depends: suspicion can be mechanically derived from generalized probabilities.

V. Structural and Governance Layer

Beyond doctrinal shortcomings, predictive policing raises a structural constitutional problem: law enforcement discretion is being delegated to private, proprietary algorithm developers. These systems are often created and controlled by entities that are not democratically accountable and operate outside traditional due-process safeguards.

This accountability failure is akin to the logic of the nondelegation doctrine in the U.S. government,^[50] which prevents one branch of government from delegating to another entity the power or function it is constitutionally authorized to exercise itself. It is explicit in Article 1, Section 1, of the U.S. Constitution, which states that “all legislative Powers herein granted shall be vested in a Congress of the United States, which shall consist of a Senate and House of Representatives.”^[51] Therefore, Congress cannot cede its lawmaking authority to external actors, such as the executive or judicial branches or private parties. By ensuring that decisions affecting rights and obligations are made by more public institutions than opaque or private entities, the doctrine aims to protect democratic legitimacy.

Predictive policing raises a parallel structural concern: police departments increasingly rely on proprietary, privately developed algorithms to exercise core policing discretion, effectively outsourcing the constitutionally significant task of determining suspicion. Just as Congress cannot hand its legislative power to private regulators, law enforcement cannot delegate the constitutional judgment required by the Fourth Amendment to companies like PredPol, Palantir, or ShotSpotter, whose operations are shielded from the Freedom of Information Act (FOIA) by trade secrecy and contractual nondisclosure.

This creates what scholars and policymakers identify as an accountability vacuum, as the public, and sometimes even the agency, doesn’t know if “government algorithms are functioning properly or [how] to identify sources of ineffectiveness or bias.”^[52] For example, police involved in the Chicago Heat List invoked the algorithm to bolster its credibility, despite its inaccurate results. In effect, it was a way for cops to target people. A review by the Inspector General of the Los Angeles Police Department found significant disparities in the administration of the LAPD’s predictive policing algorithms, PredPol and LASER, by different departments across the city. As a result, the LAPD stopped using LASER in 2019. This demonstrates how, in such contexts, the use of an algorithm may lend unwarranted legitimacy to discriminatory and baseless practices, ironically failing to constrain human discretion despite being thought to improve accuracy and reduce bias.

Case studies in Connecticut demonstrate that agencies using high-stakes algorithms are unaware of, or do not monitor, what those tools are doing.^[53] A child welfare algorithm produced by Mindshare Technology used predictive analytics to estimate the chance that a child faced neglect, danger, or death, though with anomalous results. The Department of Children and Families (DCF) was unable to produce the algorithm’s source code and “complete a robust formal evaluation to assess [ERSF’s] efficacy.”^[54] Therefore, nobody evaluated whether the algorithm exhibited bias.

When a FOIA request was submitted to the Department of Education (DOE) requesting information about algorithms in the administration of school lotteries, the DOE withheld “documents detailing coding of the placement system [and] specific data inputs and protocols,” invoking trade secret protection.^[55] Its description of factors was rudimentary, and the DOE “provided no information about any procedures it has used to assess the performance of the algorithm.”^[56] Similarly, the Department of Administrative Services (DAS) plans to withhold “anything substantive related to the actual algorithm”^[57] used in the hiring process for state employees and state contractors, citing trade secret and exam exemptions.

All three agencies indicated that their respective algorithms were proprietary and not subject to disclosure, suggesting that the state is increasingly deploying high-stakes automated decision-making systems without understanding, monitoring, or evaluating them. Agencies cannot explain how these systems operate, verify that they function, or assess whether they reproduce bias—yet they continue to rely on them in contexts with profound consequences for children, students, and public-sector workers. Constitutional and administrative judgment is effectively transferred to private vendors whose proprietary interests trump transparency. The Connecticut examples thus illustrate the core structural flaw of predictive policing: once the state relies on algorithms it cannot see, test, or challenge, meaningful accountability is impossible.

Unlike the nondelegation doctrine, predictive policing is even more troubling because the delegated function is not regulatory—it is coercive policing power. Individuals subjected to algorithmically informed stops or heightened surveillance cannot meaningfully challenge their classification, both because the algorithm is a black box and because there are no legal protections that compel police to disclose, justify, or scrutinize algorithmic inputs.

This structure also poses serious due process risks. When a person becomes the subject of police attention because of proprietary algorithmic labeling, they face consequences without notice, without the ability to contest, and without access to the underlying evidence. Predictive policing thus inverts basic procedural norms: the state acts first, and the justification is never revealed.

VI. Policy and Doctrinal Implications

The combined doctrinal and structural flaws of predictive policing reveal the need for a comprehensive governmental response. Courts, legislatures, and administrative agencies all have roles to play in reasserting constitutional limits on algorithmic policing.

Courts should clarify that predictive policing outputs cannot independently establish reasonable suspicion or probable cause. Judicial doctrine must reaffirm that the Fourth Amendment requires individualized, contextualized evidence, and that algorithmic “risk scores” or hotspot predictions do not meet the threshold articulated in Terry and Gates. Judges can also require full disclosure of algorithmic methodologies when such systems are invoked in criminal proceedings, aligning algorithmic policing with existing rules of evidence and confrontation.

Regarding legislative action, state legislatures and Congress should adopt statutory frameworks requiring:

1. Mandatory algorithmic audits assessing disparate impact, data bias, and accuracy.
2. Transparency requirements, including disclosure of vendors, training data, and system logic.
3. Prohibitions on sole reliance on predictive scores for stops, arrests, or investigative decisions.
4. Clear retention and oversight rules for data used in predictive tools.

Such statutes would align with emerging AI governance standards and directly address the accountability concerns raised by the Yale MFIA Clinic.

Finally, increasing public accountability and democratic oversight is needed as a balancing power.  Predictive policing systems must be subject to public scrutiny equivalent to that afforded to other exercises of coercive state power. Agencies should be required to publish contracts with algorithmic vendors, evaluation metrics, accuracy, disparate-impact reports, and community-oversight mechanisms. These reforms reflect a broader shift toward algorithmic governance frameworks that treat predictive systems as public actors rather than private tools. Adopting such measures would help prevent the constitutional erosion that occurs when predictive policing functions unchecked by law.

Conclusion

Quite paradoxically, predictive policing—far from its premise of being scientific, precise, and impartial—reveals a familiar problem: to what extent can state power be untethered from constitutional constraint? In a digitized world now increasingly surrounded by yet simultaneously concerned with the rise of artificial intelligence permeating every corner of society, the issue of predictive policing is a hotspot that heavily influences how AI and other algorithmic systems are regulated in the future. It sits at the intersection of constitutional doctrine, technological governance, and racialized enforcement—and it fails on all three fronts.

While marketed as an objective, data-driven innovation, predictive policing fundamentally reconfigures the Fourth Amendment by substituting statistical risk for individualized suspicion. As Terry v. Ohio and Illinois v. Gates make clear, the Constitution requires police to ground intrusions in “specific and articulable facts”^[58] and the “totality of the circumstances”^[59]—a contextual, qualitative judgment anchored in particularized evidence. Predictive systems, however, generate suspicion from correlations, not conduct; from populations, not persons. Transforming probability into suspicion collapses the doctrinal boundary that the Fourth Amendment was designed to preserve. Furthermore, by outsourcing core policing judgments to proprietary, opaque, and private algorithms, police departments effectively delegate constitutional decision-making to entities beyond democratic or judicial oversight. Agencies often lack access to the source code, cannot evaluate algorithmic effectiveness or bias, and cannot disclose the algorithmic logic due to trade-secret claims. The authority to initiate stops, surveillance, and enforcement against real people without notice, transparency, or recourse remains unchecked and runs rampant throughout the United States. Predictive policing fails because it transforms constitutional judgment into statistical output and places that transformation beyond accountability.

Given these limitations, predictive policing is incompatible with foundational Fourth Amendment principles. Courts should reject attempts to treat algorithmic forecasts as reasonable suspicion or probable cause, and legislatures must adopt accountability frameworks that restore transparency, oversight, and due process. Until policing algorithms can meet constitutional standards—and until the public can meaningfully scrutinize their design, operation, and effects—their deployment threatens not only individual rights but the legitimacy of the criminal legal system itself.

---

^[1] Terry v. Ohio, 392 U.S. 1 (1968).

^[2] Illinois v. Gates,462 U.S. 213 (1983). 

^[3] Ali Winston, All Tomorrow’s Crimes: The Future of Policing Looks a Lot Like Good Branding, S.F. Weekly (Oct. 30, 2013), http://archives.sfweekly.com/sanfrancisco/all-tomorrows-crimes-the-future-of-policing-looks-a-lot-likegood-branding/Content?oid=2827968&showFullText=true.

^[4] U.S. Dep’t of Justice, Office of Community Oriented Policing Services, Predictive Policing: The Role of Crime Forecasting in Law Enforcement Operations (2014), https://portal.cops.usdoj.gov/resourcecenter/content.ashx/cops-p282-pub.pdf.

^[5] Andrew Guthrie Ferguson, Policing Predictive Policing, 94, Wash. U. L. Rev. 1109, 1116 (2017),  https://openscholarship.wustl.edu/law_lawreview/vol94/iss5/5.

^[6] Nazgol Ghandnoosh, One in Five: Disparities in Crime and Policing, The Sentencing Project (Nov. 2, 2023), https://www.sentencingproject.org/reports/one-in-five-disparities-in-crime-and-policing/.

^[7] Jeremy Gorner, Chicago Police Use ‘Heat List’ As Strategy to Prevent Violence, Chi. Trib. (Aug. 21, 2013), https://www.chicagotribune.com/2013/08/21/chicago-police-use-heat-list-as-strategy-to-prevent-violence/.

^[8] Ferguson, supra note 5, at 1115.

^[9] Mark Puente, LAPD Pioneered Predicting Crime With Data. Many Police Don’t Think It Works, The L.A. Times, (July 3, 2019). https://www.latimes.com/local/lanow/la-me-lapd-precision-policing-data-20190703-story.html.   

^[10] Molly Griffard, A Bias-Free Predictive Policing Tool?: An Evaluation of the NYPD’s Patternizr, 47 Fordham Urb. L.J. 43, 47 (2019), https://ir.lawnet.fordham.edu/ulj/vol47/iss1/2.

^[11] Griffard, supra note 10, at 60.

^[12] Ferguson, supra note 5, at 1116.

^[13] Ferguson, supra note 5, at 1137.

^[14] See: George Mohler et al., A Penalized Likelihood Method for Balancing Accuracy and Fairness in Predictive Policing, in 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC) (2018), https://doi.org/10.1109/SMC.2018.00421; Frank R. Baumgartner et al., Racial Disparities in Traffic Stop Outcomes, 9 Duke F. L. & Soc. Change 21 (2017).

^[15] Charlie Beck & Colleen McCue, Predictive Policing: What Can We Learn from Wal-Mart and Amazon About Fighting Crime in a Recession?, Police Chief, (2009), 18. https://www.ojp.gov/ncjrs/virtual-library/abstracts/predictive-policing-what-can-we-learn-wal-mart-and-amazon-about.

^[16] Andrew G. Ferguson, Predictive Policing and Reasonable Suspicion, 62 Emory L. J. 259, 269 (2012), https://scholarlycommons.law.emory.edu/elj/vol62/iss2/1.

^[17]  Josh Koehn, Algorithmic Crimefighting, SanJose.com (Feb. 22, 2012), https://www.sanjose.com/2012/02/22/sheriffs_office_fights_property_crimes_with_predictive_policing/. 

^[18] Ferguson, supra note 16, at 268.

^[19] Griffard, supra note 10, at 45.

^[20] Ferguson, supra note 16, at 270.

^[21] Id. at 44.

^[22] Ferguson, supra note 16, at 277–284.

^[23] Ferguson, supra note 5, at 1137.

^[24] Will Douglas Heaven, Predictive Policing Algorithms Are Racist. They Need To Be Dismantled., Mit Tech. Rev. (July 17, 2020), https://www.technologyreview.com/2020/07/17/1005396/predictive-policing-algorithms-racist-dismantled-machine-learning-bias-criminal-justice/.

^[25] Lum, K. and Isaac, W. (2016), To Predict and Serve? Significance, 13: 14-19. https://doi.org/10.1111/j.1740-9713.2016.00960.x.

^[26]  Id.

^[27]  Id.

^[28] Victor Rotaru, Yi Huang, Timmy Li, James Evans & Ishanu Chattopadhyay, Event-Level Prediction of Urban Crime Reveals a Signature of Enforcement Bias in U.S. Cities, 6 Nat. Hum. Behav. 1056 (2022), https://doi.org/10.1038/s41562-022-01372-.

^[29] Rotaru et al., supra note 28..

^[30] Media Freedom & Information Access Clinic, Algorithmic Accountability: The Need for a New Approach to Transparency and Accountability When Government Functions Are Performed by Algorithms 2 (Jan. 18, 2022), Yale Law School, https://law.yale.edu/sites/default/files/area/center/mfia/document/algorithmic_accountability_report.pdf. 

^[31] Griffard, supra note 10.

^[32] Tim Lau, Predictive Policing Explained (Apr. 1, 2020), Brennan Center for Justice, https://www.brennancenter.org/our-work/research-reports/predictive-policing-explained.

^[33] Id.

^[34] Id.

^[35] Terry v. Ohio, 392 U.S. 1, 21 (1968).

^[36]  Id. at 22.

^[37]  Id. at 22.

^[38]  Id. at 9.

^[39] Id. at 16.

^[40] Id. at 16–17.

^[41] Id. at 27.

^[42] Madison Blevins, When Dirty Data Leads to Dirty Policing, 29 Rich. J.L. & Tech. 166 (2024), https://scholarship.richmond.edu/jolt/vol29/iss2/4.

^[43] Illinois v. Gates, 462 U.S. 213, 237 n.11 (1983).

^[44] Id. at 230–32.

^[45] Id. at 231–32.

^[46] Id. at 232.

^[47] Id. at 235.

^[48] Id. at 232.

^[49] Id. at 272.

^[50] Cong. Rsch. Serv., Overview of Nondelegation Doctrine, Constitution Annotated, https://constitution.congress.gov/browse/essay/artI-S1-5-1/ALDE_00000014/.

^[51] U.S. Const. art. I, § 1.

^[52] MFIA Clinic, Algorithmic Accountability, supra note 30, at iii.

^[53] Id. at 17.

^[54] Id. at 20–21.

^[55] Id. at 22.

^[56] Id at 23.

^[57] Id. at 24.

^[58] Terry v. Ohio, 392 U.S. 1, 21 (1968).

^[59]  Illinois v. Gates,462 U.S. 213,272 (1983).