The previous three chapters examined algorithms that operate at defined procedural moments: COMPAS at sentencing, the PSA at bail, Clearview AI at the identification stage. Predictive policing systems operate earlier still — not at a point where a crime has been committed and a suspect identified, but before any of that, at the stage where the state decides where to look and whom to watch.
That earlier intervention is precisely what makes predictive policing the most constitutionally unsettled form of criminal justice AI this book examines. When an algorithm influences a sentence, the Fourth Amendment has already run its course. When an algorithm influences where officers patrol, whom they stop, and which communities receive concentrated enforcement attention, the Fourth Amendment is the question — and courts have not yet given a settled answer.
I. The two-model landscape: place-based prediction and acoustic detection
Two systems have dominated the legal and policy debate over predictive policing in the United States: PredPol, now absorbed into SoundThinking’s infrastructure, and ShotSpotter, the gunshot detection network operated by that same company.
PredPol / Geolitica. PredPol was developed beginning in 2010, in part through a collaboration between the Los Angeles Police Department and UCLA researchers, using a mathematical model originally designed to predict earthquake aftershocks. The system ingested three data points — crime type, location, and time — and produced daily maps identifying locations where specific crimes were statistically more likely to occur. Officers were directed to patrol those areas during the predicted time windows. By the early 2020s PredPol, renamed Geolitica in 2021, was the most widely used predictive policing algorithm in the United States.
The system’s operational record did not match its theoretical promise. The Markup examined 23,631 predictions generated by Geolitica for the Plainfield, New Jersey police department between February and December 2018, analyzing the accuracy of each prediction against later-reported crimes. The success rate was less than half of one percent — fewer than 100 of the predictions corresponded to a reported crime in the predicted category at the predicted location. A separate Gizmodo/Markup investigation published in 2021, drawing on data downloaded from an unprotected cloud server linked from the LAPD’s own website, found that Geolitica’s predictions disproportionately targeted Black and Latino neighborhoods across 38 cities. The LAPD, which had been the earliest institutional adopter of PredPol, dropped the product in 2020. Geolitica ceased operations at the end of 2023. SoundThinking acquired key intellectual property and Geolitica’s engineering team, integrating elements of the technology into its ResourceRouter patrol management platform, though SoundThinking stated explicitly that it did not acquire Geolitica’s crime prediction source code.
ShotSpotter. ShotSpotter operates differently. Instead of predicting crime from historical data, it deploys acoustic sensor networks — microphones mounted on street infrastructure — to detect impulsive sounds consistent with gunfire. When the system identifies a possible gunshot, it alerts police dispatchers within approximately sixty seconds. SoundThinking has claimed accuracy rates of ninety-seven percent in internal audits. Independent analyses have produced sharply different numbers.
The MacArthur Justice Center examined ShotSpotter deployments in Chicago between July 2019 and April 2021 and found that 88.72 percent of incidents flagged by ShotSpotter ended with police finding no evidence of gun-related crime. Chicago city data from January 2021 through July 2024 — covering approximately 145,000 ShotSpotter alerts — showed that only 7.1 percent resulted in finding a shooting victim. The Cook County State’s Attorney’s Office analyzed approximately 12,000 ShotSpotter incidents and found that the system contributed to arrests in only 1 percent of cases. Chicago announced in January 2024 that it would not renew its SoundThinking contract, decommissioned ShotSpotter on September 22, 2024, and has budgeted for a replacement gunshot detection technology.
The legal problem ShotSpotter creates is not limited to misclassification. False alerts — triggered by fireworks, vehicle backfires, construction equipment, or other impulsive sounds — can direct officers to locations where no crime has occurred and generate investigative stops, searches, and enforcement contacts based on an acoustic inference. The constitutional question those contacts raise is the same one that runs through all predictive policing: what is the legal status of a police action initiated by an algorithmic output rather than by an officer’s direct observation?
II. The feedback loop: when data confirms itself
U.S. Constitution, Amendment XIV — Washington v. Davis, 426 U.S. 229 (1976)
Predictive policing systems are trained on historical criminal justice data — arrest records, incident reports, enforcement contacts. That data is not a neutral record of where crime occurs. It is a record of where police have historically concentrated their presence, which in turn reflects prior decisions about which communities to surveil, which offenses to prioritize, and which behaviors to treat as criminally relevant. A system trained on that history learns to predict where enforcement has been concentrated — not necessarily where crime is most prevalent.
The feedback loop operates as follows. Historical arrest data shapes the algorithm’s predictions. The algorithm directs patrol resources to neighborhoods with historically high enforcement rates. Increased patrol presence in those neighborhoods produces more enforcement contacts, which generate more data. That new data feeds back into the algorithm, which interprets the additional enforcement activity as confirmation that the targeted area is a high-crime zone. The system reinforces itself.
For criminal defense lawyers, the feedback loop is a litigation argument, not just a policy concern. If an algorithm is trained on data that reflects prior patterns of racially or geographically concentrated policing, and if that algorithm then directs future enforcement in ways that perpetuate those patterns, the question is whether the resulting enforcement actions are constitutionally defensible. As established in Chapter 14, Washington v. Davis, 426 U.S. 229 (1976), requires a showing of discriminatory intent for a constitutional equal protection violation — statistical disparate impact alone is insufficient. But where a department is on documented notice that its predictive tool systematically targets communities of color through a demonstrable feedback loop, and continues deploying the tool without modification, the intent analysis becomes less straightforward.
III. The Fourth Amendment framework
U.S. Constitution, Amendment IV Terry v. Ohio, 392 U.S. 1 (1968) Carpenter v. United States, 585 U.S. 296 (2018) Katz v. United States, 389 U.S. 347 (1967)
The Fourth Amendment’s prohibition on unreasonable searches and seizures establishes the constitutional framework for evaluating police stops and searches. Terry v. Ohio, 392 U.S. 1 (1968), established that a brief investigative stop — a Terry stop — is constitutionally permissible if the officer has reasonable suspicion, defined as specific and articulable facts that criminal activity is afoot. That standard is lower than probable cause, which is required for a full arrest or search, but it is still an individualized standard: the officer must be able to point to facts about this person, not statistical generalizations about this neighborhood.
Predictive policing creates a Fourth Amendment problem precisely at that intersection. If an officer stops an individual solely because an algorithm has identified the surrounding area as a predicted crime zone, the stop lacks the individualized suspicion that Terry requires. The algorithm’s output is a statistical inference about location and time — not an observation about the individual being stopped. Courts have not definitively resolved how much weight a predictive policing alert can contribute to a reasonable suspicion analysis, whether combined with other observations or on its own, and appellate case law directly addressing the question remains sparse. That unresolved status means the suppression argument is preserved and available.
The Carpenter reasoning from Chapter 24 adds a second layer for ShotSpotter specifically. Carpenter v. United States, 585 U.S. 296 (2018), held that the government’s aggregation of digital records over time can transform individually innocuous data points into a constitutionally significant intrusion requiring a warrant. A sensor network that continuously monitors sound across entire city neighborhoods, classifies those sounds through machine learning, and generates enforcement alerts affecting residents of targeted communities operates at a scale that raises the same aggregation concern Carpenter identified in the location data context. Whether that argument will prevail in courts remains to be litigated, but the doctrinal foundation is available.
IV. The EU AI Act: a different structural answer
EU AI Act, Regulation (EU) 2024/1689, Article 5(1)(d), Annex III EU Charter of Fundamental Rights, Article 48 — Presumption of Innocence
As established in Chapter 24, the EU AI Act, Regulation (EU) 2024/1689, Article 5(1)(d), prohibits AI systems used to assess or predict the risk of a natural person committing a criminal offence based solely on profiling or on personality traits and characteristics, unless the system supports a human assessment already grounded in objective and verifiable facts directly linked to criminal activity. That prohibition applies to person-based predictive systems.
Place-based systems — those that analyze geographic and temporal crime patterns without prediling individuals — present a more nuanced picture under the AI Act. They may not fall within Article 5(1)(d) if their output is genuinely limited to spatial and temporal predictions. But where a system effectively identifies individuals for heightened attention — through persistent targeting of specific neighborhoods, through integration with person-based data, or through use in individual stop-and-frisk decisions — the Article 5(1)(d) analysis becomes more demanding.
In either case, AI systems used in law enforcement for individual risk assessment and profiling fall within the high-risk category under Annex III. That classification triggers the full compliance architecture: risk management under Article 9, data governance under Article 10, transparency to deployers under Article 13, and effective human oversight under Article 14. It also places the burden of demonstrating compliance before deployment — not after harm is documented — on the provider and the deployer.
The EU Charter of Fundamental Rights, Article 48, which guarantees the presumption of innocence, provides additional constitutional grounding for the European regulatory response. A system that generates a probabilistic prediction that a person will commit a crime — and that influences police action toward that person before any conduct occurs — is in tension with a presumption of innocence that is not merely a procedural rule but a fundamental right. That tension is precisely what the Article 5(1)(d) prohibition addresses.
V. Practical litigation strategy
U.S. Constitution, Amendment IV — Terry v. Ohio, 392 U.S. 1 (1968) Mapp v. Ohio, 367 U.S. 643 (1961) — exclusionary rule Wong Sun v. United States, 371 U.S. 471 (1963) — fruit of the poisonous tree
When predictive policing has contributed to an enforcement action — a stop, a search, an arrest — the defense lawyer’s immediate task is to establish the causal chain from algorithmic output to officer conduct. That requires discovery, and the discovery request should be specific: what system was in use at the time; what output the system generated; what that output told the officer; and what the officer did in response to it.
The Fourth Amendment motion to suppress is the primary vehicle. The argument is that a stop or search initiated on the basis of an algorithmic prediction about location — without additional individualized, articulable facts about the specific person stopped — lacks the reasonable suspicion Terry requires. If the stop is unlawful, evidence obtained in the course of the stop is suppressible under the exclusionary rule, with downstream evidence excluded under the fruit of the poisonous tree doctrine.
The discovery challenge is real. Predictive policing vendors have claimed trade secret protection for their algorithms, resisting disclosure of training data, model architecture, and performance records. Where proprietary protection is asserted, the defense should argue that the state cannot both rely on a proprietary tool to justify police action and deny the defense access to information necessary to test that tool’s reliability. The confrontation and due process principles established in the COMPAS context in Chapter 25 apply here with equal force.
Expert testimony from data scientists or algorithmic auditors can establish, where the record permits, the feedback loop dynamic, the demographics of enforcement targeting, and the divergence between the system’s claimed accuracy and its documented operational performance. That testimony supports both the Fourth Amendment suppression argument and the equal protection record.
VI. The open constitutional question
Predictive policing has generated sustained controversy in academic literature, civil liberties advocacy, and municipal policy debates. What it has not yet generated is a definitive federal appellate ruling on the Fourth Amendment status of an algorithm-initiated stop.
That gap is a litigation opportunity. Every criminal case in which a predictive policing alert contributed to an enforcement action is an opportunity to build the record, preserve the constitutional argument, and move the developing case law toward a resolution that takes the Terry individualization requirement seriously in the algorithmic context. The argument is not that predictive policing is categorically unlawful. It is that the state must be held to the same constitutional standard in an algorithmic context that Terry has always required in a traditional one: specific and articulable facts about this person, not statistical generalizations about this neighborhood.
Next: Chapter 29 — AI evidence in court: the Daubert standard and the trade secret problem.
Leave a Reply