The Global Technology Stack Behind Predictive Law Enforcement

Full report

On this page

Executive Summary

The integration of advanced computational models into global law enforcement and state security apparatuses represents a fundamental epistemological shift in the anticipation, classification, and mitigation of risk. Across jurisdictions, agencies have transitioned from reactive investigative postures to proactive, algorithmically driven methodologies aimed at forecasting criminality, terrorism, social disorder, recidivism, and border violations. This report provides an exhaustive deconstruction of the global technology stack underpinning predictive law enforcement, mapping the complete data lifecycle, structural architectures, and mathematical paradigms that define modern state surveillance and intervention. By analyzing the underlying mechanics of 25 distinct systems deployed across more than 15 countries, this research delineates the boundaries between statistically valid forecasting, probabilistic inference, and pseudoscientific overreach. The analysis reveals a deeply interconnected ecosystem where military-grade intelligence architectures, academic criminological theories, and commercial machine-learning platforms converge. As foundation models and generative artificial intelligence are ingested into this stack, the velocity of intelligence production is accelerating, simultaneously introducing novel vectors for adversarial manipulation, systemic bias, and catastrophic epistemological failure.

Global Reference Architecture and the Technology Lifecycle

The architecture of predictive law enforcement is rarely monolithic; it is a composite of disparate data ingestion pipelines, processing engines, and operational deployment mechanisms. To accurately assess the efficacy and legality of these systems, the complete technology lifecycle must be mapped across nineteen discrete phases. The lifecycle initiates with the data collection phase, involving the mass harvesting of structured and unstructured signals from terrestrial sensors, digital surveillance, and administrative databases. This raw material undergoes ingestion into centralized data lakes or federated nodes via application programming interfaces (APIs), batch processing, or real-time streaming architectures. Because raw data is inherently noisy, the system executes algorithmic cleaning to remove erroneous, corrupted, or null data points, establishing a baseline of computational viability. Subsequently, normalization translates disparate data formats into a standardized schema, such as converting varying timestamp formats or geocoding addresses to standard spatial coordinates. To build comprehensive profiles, the architecture relies on identity matching, a process that links distinct data points to a singular entity using deterministic Boolean logic or probabilistic machine learning. This unified profile is subjected to enrichment, where internal police records are augmented with third-party datasets, such as commercial data broker files or telecommunications metadata. The analytical engine then performs feature construction, the mathematical transformation of raw data into predictive variables—for example, calculating the spatial distance to the nearest prior burglary or deriving behavioral metrics from network centrality. With the feature space defined, model training applies algorithms to historical data to identify correlative patterns, optimize parameters, and minimize error functions. Before live application, rigorous testing evaluates model accuracy, precision, and recall using hold-out datasets, backtesting, or spatial cross-validation. Once validated, deployment migrates the trained model into a production environment capable of ingesting live telemetry. The operational core is real-time scoring, the continuous generation of risk probabilities, threat classifications, or spatial hot-spot coordinates based on incoming data. These outputs demand analyst review, embedding a human-in-the-loop to validate algorithmic determinations, a process increasingly augmented by automated summarization. Analysts authorize an operational action, resulting in physical or digital intervention by state actors, such as dispatching patrols or revoking a visa. Following the intervention, feedback collection records the outcome—verifying, for instance, whether a crime was actually intercepted at a predicted coordinate1. This feedback loop dictates model updating, the continuous or periodic retraining of the algorithm to account for new baseline realities and to mitigate statistical drift. Simultaneously, the architecture must manage the physical data through retention policies, storing historical predictions and inputs in accordance with legal mandates. When errors are identified, correction mechanisms allow for the manual or automated rectification of false positives and corrupted records. Ultimately, lifecycle governance requires the deletion of obsolete, unlawful, or time-expired data, culminating in system retirement when legacy algorithms are decommissioned and intelligence is migrated to next-generation architectures.

Data-Flow Diagram: General Predictive Architecture

Epistemological Distinctions in Intelligence Analytics

The discourse surrounding predictive policing frequently conflates distinct analytical operations. Establishing a precise taxonomy of these operations is critical for auditing system claims and distinguishing between retroactive data retrieval and genuine foresight. The most foundational operation is record retrieval, which relies on deterministic database queries to extract stored data matching exact parameters, such as pulling a vehicle's registration based on a specific license plate. Moving up the complexity gradient, data matching involves linking records across different databases based on identical shared attributes, such as cross-referencing a social security number across tax and criminal justice repositories. When shared attributes are incomplete or inconsistent, systems employ identity resolution, a highly probabilistic process that determines whether disparate records—such as varying transliterations of an Arabic name, a partial phone number, and a social media handle—refer to the same real-world entity3. Analytical models inherently rely on correlation, the statistical observation that two variables move in tandem (e.g., higher ambient temperatures and increased emergency calls), without establishing a causal mechanism. Classification takes these correlated features to assign an entity or event into predefined categories, such as an acoustic array classifying a sharp waveform as a gunshot rather than a firecracker5. The concept most associated with "predictive policing" is forecasting, which involves the estimation of future events in aggregate over a spatiotemporal domain, such as calculating an elevated probability of property crime in a specific postal code over the next 72 hours. This must be strictly differentiated from causal prediction, the assertion that specific variables directly and invariably cause an outcome, enabling precise counterfactual analysis. True causal prediction is extraordinarily rare and scientifically contested within criminological machine learning6. When analyzing human actors, systems perform threat assessment, evaluating an individual's capability, intent, and proximity to executing a specific hostile act, often utilizing behavioral indicators and rule-based psychological rubrics8. To operationalize these assessments across large populations, algorithms engage in prioritization, ranking entities, cases, or geographic zones to optimize the allocation of finite state resources based on aggregate risk scores11. The ultimate, and most legally precarious, operational threshold is automated intervention, where the execution of a defensive or offensive action occurs without human intermediation, such as an algorithmic flag automatically revoking a traveler's digital visa or authorizing a drone intercept.

Data Ingestion Vectors and the Surveillance Backcloth

The efficacy of predictive models is strictly bound by the breadth, dimensionality, and integrity of their training data. Worldwide, agencies leverage a vast spectrum of data types, inherently transforming public and private life into a machine-readable surveillance backcloth. Criminal justice data forms the foundational training set for most models, encompassing reported crimes, arrests, convictions and acquittals, police contacts (such as stop-and-frisk field interview cards), and probation and parole records. However, this data is chronically afflicted by reporting and enforcement biases; it measures law enforcement activity and deployment density rather than the true underlying prevalence of crime13. To augment this, authorities increasingly ingest administrative and state data. This includes border and travel information and visa and immigration records to track transnational movement. More invasively, models are beginning to ingest education records, health and mental-health information, and welfare and housing data alongside generic public records. The ingestion of welfare and housing data marks a profound shift toward policing the socioeconomic margins, where algorithms treat poverty indicators as proxy variables for risk13. The physical world is digitized through an expansive array of surveillance sensors. CCTV and body-camera footage provide continuous visual monitoring, while aerial imagery enables wide-area spatial analysis. Vehicle data is captured via automated license plate readers, tracking the movement vectors of millions of citizens daily. Correspondingly, human physicality is encoded via biometric data, incorporating facial images, fingerprints, DNA, voiceprints, and increasingly, gait and behavioral biometrics derived from video analytics18. Digital and commercial data provides a window into intent and routine activity. Agencies routinely ingest financial activity, telecommunications metadata (call detail records), internet and social-media activity, and historical location history acquired via mobile applications or cell-site simulators. Because law enforcement is restricted by domestic surveillance laws, they frequently bypass warrants by utilizing commercial databases maintained by third-party data brokers. Finally, to understand criminal syndicates, agencies rely on relational data, combining raw intelligence reports and structured relationship networks (co-arrests, familial ties, financial linkages) to map systemic threat architectures20.

Technical Taxonomy of Predictive and Analytical Models

Law enforcement agencies deploy a vast array of mathematical architectures, each suited to specific operational vectors ranging from spatial forecasting to semantic intelligence extraction.

Model-Family Comparison Table

System-by-System Technical Inventory

To map the operational reality of predictive policing, the following section reconstructs the architectures of 25 intelligence and analytical systems deployed across 16 countries.

Advanced Mathematical Mechanics in Criminological Modeling

The efficacy of predictive policing is heavily reliant on the specific mathematical architecture applied to the spatial and temporal distribution of crime. Two paradigms dominate the academic and operational space: Self-Exciting Point-Process Models and Risk-Terrain Modeling. Self-Exciting Point-Process Models: Rooted in seismology, models like the Epidemic-Type Aftershock Sequence (ETAS) treat crimes not as independent, identically distributed variables, but as contagious events. Developed extensively by Mohler, Brantingham, and others, the conditional intensity function, [Figure omitted from source export], represents the expected rate of events given the historical accumulation of points [Figure omitted from source export]. The basic formulation is [Figure omitted from source export], where [Figure omitted from source export] is the stationary background crime rate, and [Figure omitted from source export] is the triggering kernel dictating how a single crime temporarily elevates the probability of subsequent crimes in its immediate spatio-temporal vicinity48. While mathematically elegant and highly effective for near-repeat phenomena like residential burglary or retaliatory gang violence, these models often fail when applied to crimes driven by broader socioeconomic factors rather than peer contagion. Furthermore, the estimation of parameters in continuous time can be numerically unstable, requiring advanced Monte Carlo simulations to resolve missing data50. Risk-Terrain Modeling (RTM): Rather than retroactively plotting where past crimes occurred, RTM analyzes the environmental backcloth to identify vulnerabilities. It utilizes spatial logistic regression and kernel density estimation to evaluate geographic features—such as proximity to liquor stores, subway stations, or abandoned buildings—to determine criminogenic vulnerability6. This is formalized by dividing a geography into grid cells and establishing the presence, absence, or density of risk factors, culminating in a composite risk map. While lauded for avoiding a pure reliance on historical arrest data, RTM inadvertently codifies socio-economic disparities. Because the "attractors" of crime are frequently the structural features of impoverished or systematically marginalized neighborhoods, the mathematical output naturally directs police back to those exact demographic centers43.

Institutional Paradigms and the Vendor-Market Map

The predictive policing market is highly fragmented, categorized by four distinct institutional paradigms, each offering distinct advantages regarding technical capability and profound disadvantages regarding transparency.

1. Proprietary Commercial Systems (The Dominant Paradigm): Companies such as Palantir (Gotham, AIP), SoundThinking (ShotSpotter, CrimeTracer), and Cellebrite (Pathfinder) provide end-to-end, black-box solutions. These vendors benefit from massive capital investment, cloud-native scalability, and rapid deployment capabilities. However, they introduce severe vendor lock-in, exorbitant licensing costs, and a near-total lack of algorithmic transparency32. Law enforcement agencies rarely have access to the underlying weights of the models dictating their patrol routes or suspect link-analysis.

2. Government-Built Systems: Architectures constructed entirely in-house by state security apparatuses, such as the Dutch CAS (Crime Anticipation System) or the Chinese IJOP. These systems are deeply integrated with municipal databases and sovereign data lakes, avoiding private-sector dependency13. However, they are heavily reliant on internal technical talent, can suffer from bureaucratic stagnation, and are almost never subjected to independent, adversarial public auditing.

3. Academic-State Partnerships: Systems developed by university researchers in direct collaboration with local constabularies, such as the Kent ETAS (UCLA/Mohler) or the Durham HART algorithm (Cambridge). These projects generally feature high scientific rigor, peer-reviewed methodology, and greater transparency1. Despite this, they often struggle to scale commercially or maintain long-term funding beyond the initial research grant.

4. Open-Source Applications: Platforms like Rio's CrimeRadar or basic Risk Terrain Modeling scripts available on GitHub. While highly transparent, democratized, and cost-effective, they lack the secure, hardened infrastructure required for handling classified state intelligence or processing massive volumes of highly sensitive criminal justice telemetry41.

Foundation Models and the Generative AI Shift

The intelligence landscape is currently undergoing a violent paradigm shift driven by Large Language Models (LLMs) and Generative AI. Historically, querying intelligence databases required analysts trained in SQL, Boolean logic, or specialized graph-query languages. Platforms such as Palantir’s Artificial Intelligence Platform (AIP) now leverage LLMs to establish a natural language interface over an organization's proprietary data ontology30. This architecture allows field officers or tactical commanders to input conversational prompts—for example, "Show me all known associates of Subject X who have crossed the border in the last 72 hours, mapped against recent acoustic gunshot alerts." This generative shift empowers agencies to execute rapid unstructured data synthesis, allowing models to ingest thousands of pages of interrogation transcripts, intercepted communications, and disjointed patrol narratives, transforming them into structured intelligence briefs and actionable insights. Furthermore, the technology is moving toward automated interventions. In advanced configurations, LLM-based agents no longer just respond to chat interfaces; they execute workflows, automatically drafting subpoenas, generating comprehensive suspect dossiers, or triggering secondary surveillance requests without human prompting32. However, the ingestion of LLMs introduces catastrophic epistemological vulnerabilities. Because LLMs are probabilistic text generators rather than deterministic fact engines, they are highly prone to hallucinations. Without strict ontological grounding and "reflection" techniques—forcing the LLM to verify its output against a hardened database—the model may fabricate intelligence, leading directly to unlawful arrests or compromised investigations31.

Cybersecurity Threat Model for Law Enforcement AI

Artificial intelligence introduces entirely novel attack surfaces that standard cybersecurity protocols (such as firewalls and endpoint detection) are ill-equipped to defend against. Relying on the MITRE ATLAS (Adversarial Threat Landscape for Artificial-Intelligence Systems) framework, the cybersecurity threat model for predictive policing systems includes several highly specific vectors52:

1. Data Poisoning (T0051): Adversaries manipulate the training data to fundamentally alter the model's behavior. In a law enforcement context, organized crime syndicates could systematically generate false emergency calls, spoof GPS locations, or deploy automated bots to manipulate social media sentiment. This poisons the spatial hot-spot algorithms, effectively clearing a physical or digital corridor for illicit activity55.

2. Model Evasion (T0024): Attackers craft specific inputs (adversarial examples) designed to bypass algorithmic classification. Examples include specialized clothing or infrared LEDs that break computer vision facial recognition, or broadcasting adversarial acoustic noise that prevents sensors from classifying a weapon discharge56.

3. Model Extraction and Inversion: State-sponsored actors query the law enforcement model repeatedly to reverse-engineer its parameters. Once the model is extracted, adversaries can potentially reconstruct the sensitive training data (e.g., classified intelligence files) or locate precise operational blind spots in border security algorithms53.

4. Supply Chain Compromise: Relying on open-source machine learning libraries or pre-trained foundation models (such as those hosted on Hugging Face) exposes law enforcement architectures to embedded backdoors. A malicious payload triggered during model serialization can grant an attacker a reverse shell into a highly classified police network54.

5. Prompt Injection: With the rise of LLMs, adversaries can embed hidden text strings within data the police might scrape (e.g., a hidden prompt on a suspect's public web page instructing the police LLM to classify the suspect as "low-risk" or to execute a malicious command upon ingestion)56.

Technical Risks, Algorithmic Degradation, and Systemic Failure

Beyond explicit cyberattacks, the deployment of mathematical models in the socio-legal sphere introduces compounding technical risks. Because criminality is a deeply sociological construct, the data measuring it is inherently flawed, leading to systemic downstream failures. Proxy Variables and Target Leakage: Algorithms are frequently legally barred from using protected variables, such as race or religion. However, machine learning models optimize by identifying proxy variables—such as ZIP code, education level, or housing data—that heavily correlate with race, thereby recreating discriminatory patterns under a veneer of mathematical objectivity43. Target leakage occurs when the model is trained on variables that intrinsically contain the outcome it is trying to predict. For example, using "number of prior arrests" to predict "future likelihood of arrest" is tautological; it merely predicts police presence and enforcement patterns, not necessarily underlying civilian criminality. Circular Intelligence and Feedback Loops: This is the most profound failure mode in geospatial predictive policing. If a model predicts high crime in Area A, commanders dispatch more officers to Area A. These officers subsequently make more arrests (often for discretionary, low-level offenses like loitering or drug possession) simply due to their increased presence. These new arrests are fed back into the system, validating the model's prediction and causing it to continually designate Area A as high-risk. The system ceases to predict crime and instead begins to predict—and dictate—police behavior, resulting in severe over-policing of minority populations7. Model Drift and Concept Drift: As the behavioral patterns of adversaries evolve, or as macro-socioeconomic conditions shift (e.g., a pandemic, economic collapse), the static historical data on which the model was trained becomes obsolete. Model drift occurs when the statistical properties of the target variable change, leading to a silent, steady degradation in predictive accuracy that necessitates continuous undocumented model updates59. Concept drift occurs when the very definition of the crime changes, altering the mathematical baseline. Imbalanced Classes and Low Base Rates: Genuinely catastrophic events, such as terrorism or mass shootings, have exceedingly low base rates. Training a machine learning classifier on extremely rare events invariably leads to either massive overfitting (the model simply memorizes the statistical noise of the few past events) or unacceptably high false-positive rates60. Poor calibration further exacerbates this; an algorithm might output an 80% risk score, but in reality, the event only occurs 10% of the time, leading to unwarranted automated interventions. Inaccurate or Stale Data and Entity Failures: In federated intelligence systems, identity resolution algorithms attempt to merge records. Variations in transliteration (e.g., Arabic to Latin script), data entry errors, or duplicate identities lead to catastrophic entity merging failures. A false positive in entity resolution can legally taint an innocent individual with the criminal history of a syndicate member. Furthermore, a failure to delete obsolete records ensures that individuals are perpetually scored based on stale data, violating basic data protection frameworks. Finally, poorly configured cloud storage of these massive data lakes frequently results in severe data leakage and unauthorized access via insider abuse57.

Auditing, Governance, and Accountability Frameworks

To mitigate the catastrophic technical and societal risks inherent in algorithmic policing, rigorous governance frameworks must be instituted across all jurisdictions prior to deployment.

Model-Audit Checklist

• \[ \] Base Rate Verification: Does the model account for the empirical base rate of the target event, or is the training data artificially balanced to inflate accuracy metrics?
• \[ \] Proxy Variable Audit: Have features been rigorously tested for multi-collinearity with protected classes (race, religion, socioeconomic status)?
• \[ \] Target Leakage Check: Are all predictive variables causally independent of the target outcome (i.e., avoiding police-action data to predict civilian criminality)?
• \[ \] Calibration Testing: Does a predicted 80% risk score actually correspond to an event occurring 8 out of 10 times in empirical reality?
• \[ \] Data Provenance: Are the legal origins, consent mechanisms, and chain of custody for all training data fully documented and verifiable?

Minimum Documentation Standard for Law Enforcement AI

Agencies must mandate a "Model Card" or equivalent cryptographic ledger detailing the operational boundaries of the system. This must include the Optimization Function, explicitly defining the mathematical objective the model is programmed to achieve (e.g., minimizing false negatives versus minimizing false positives). It must provide a transparent listing of Feature Weights/Importance, indicating which data points exert the highest influence on the model (e.g., the exact coefficient of "prior arrests" versus "age")29. Furthermore, the documentation must explicitly state the Training Data Topography (the temporal bounds, geographic limits, and exact datasets used) and outline Known Failure Modes, including demographic blind spots and confidence intervals.

Proposed Logging and Traceability Requirements

To establish legal accountability, systems must enforce immutable audit trails. All outputs that result in an operational action (e.g., arrests, dispatch, asset seizure) must be logged to a cryptographic, append-only ledger to prevent post-hoc alteration by internal actors53. Furthermore, systems must capture feature snapshots; the precise values of the variables that led to an alert must be frozen in time, allowing independent auditors or defense attorneys to recreate the exact mathematical state of the model at the moment a decision was made.

Recommended Red-Team and Field-Testing Methodology

Prior to operational deployment, systems must undergo rigorous, independent evaluation.

1. Silent Backtesting: The model must be run against historical data it has never seen, evaluating its hypothetical predictions against known reality.

2. Shadow Deployment (Silent Testing): The system generates live predictions that are explicitly withheld from field officers. Auditors then measure if the predicted events actually occurred without police interference, establishing a true baseline2.

3. Randomized Controlled Trials (RCT): Following shadow deployment, the system must be subjected to an RCT (e.g., the Brantingham/Mohler methodology utilized in Kent and Los Angeles). Hot-spots are randomly assigned to be patrolled based on algorithm predictions versus traditional human analyst predictions, measuring statistically significant divergences in crime reduction without introducing runaway feedback loops1.

4. Adversarial Penetration Testing: Specialized red teams must attempt to poison the model's training data, extract its architecture via APIs, and execute prompt-injections to evaluate its cryptographic and logical resilience53.

Final Assessment of Predictive Capabilities

A rigorous analysis of the global technology stack reveals a stark dichotomy in the credibility of predictive law enforcement systems. The scientific validity of these tools is highly dependent on the nature of the crime being predicted and the underlying architecture deployed. The Most Technically Credible Capabilities: The application of spatial analytics (such as Risk Terrain Modeling) and self-exciting point processes (Hawkes/ETAS) to highly specific, near-repeat property crimes (e.g., commercial burglary, motor vehicle theft, and localized gang retaliation) demonstrates genuine, statistically measurable predictive validity2. Because property crimes rely heavily on geographic opportunity rather than complex psychological intent, the mathematical modeling of their spatial dispersion holds empirical weight. Furthermore, semantic link analysis, entity resolution, and Natural Language Processing (as utilized by platforms like Palantir and Cellebrite) are exceptionally powerful at establishing post-hoc relationships and mapping complex, disorganized criminal networks across vast, unstructured data lakes4. The Least Credible Capabilities: Conversely, systems designed for individual, causal behavioral forecasting—attempting to predict complex human violence, ideological radicalization, or localized homicides based on broad historical variables—border on mathematical pseudoscience16. The use of artificial intelligence to infer intent, emotion, or prospective threat from biometrics or generalized social media scraping entirely lacks empirical grounding and introduces severe civil liberties violations. Furthermore, geospatial predictive policing directed at generalized "street crime" or narcotics offenses serves primarily as an algorithmic laundering mechanism. It uses the veneer of objective machine learning to justify the continuous deployment of patrol resources into historically over-policed, socioeconomically disadvantaged communities, generating perpetual feedback loops rather than authentic foresight. The future of intelligence analytics lies not in the omniscient prediction of human behavior, but in the rapid, AI-augmented synthesis of the past and the present. The profound risk is that state security apparatuses will continue to procure the former by abusing the accelerating capabilities of the latter.

Works cited

1. (PDF) Randomized Controlled Field Trials of Predictive Policing \- ResearchGate, https://www.researchgate.net/publication/282772661\_Randomized\_Controlled\_Field\_Trials\_of\_Predictive\_Policing

2. Randomized Controlled Field Trials of Predictive Policing \- AWS, https://pjb-serve-pubs.s3.us-west-1.amazonaws.com/MohlerEtAl-JASA-2015.pdf

3. Top 7 AI-Driven Forensic Investigator Tools in 2026 | Energent.ai, https://www.energent.ai/use-cases/en/compare/ai-driven-forensic-investigator

4. D7.5 DATA VISUALIZATION V2, ROXANNE PLATFORM V2, https://www.roxanne-euproject.org/results/files/d7-5-data-visualization-v2-roxanne-platform-v2-v2-0.pdf

5. Gunshot Detection Systems in Civilian Law Enforcement | Request PDF \- ResearchGate, https://www.researchgate.net/publication/275247958\_Gunshot\_Detection\_Systems\_in\_Civilian\_Law\_Enforcement

6. Epistemologies of predictive policing: Mathematical social science, social physics and machine learning \- Uni Freiburg, https://freidok.uni-freiburg.de/files/194343/1vouFZ0hy5JvBaMP/20539517211003118.pdf

7. COUNTERFACTUAL FAIRNESS WITH HUMAN IN ... \- OpenReview, https://openreview.net/pdf?id=I09JonzQJV

8. Artificial intelligence in public security and criminal justice systems in Latin America \- Fair Trials, https://www.fairtrials.org/app/uploads/2024/08/Artificial-intelligence-in-public-security-and-criminal-justice-systems-in-Latin-America.pdf

9. Germany \- Automating Society Report 2020 \- Algorithm Watch, https://automatingsociety.algorithmwatch.org/report2020/germany/

10. 'PREDICTIVE POLICING', 'PREDICTIVE JUSTICE', AND THE USE OF 'ARTIFICIAL INTELLIGENCE' IN THE ADMINISTRATION OF CRIMI \- AIDP, https://www.penal.org/wp-content/uploads/2025/09/A-02-23.pdf

11. Predictively policed: the Dutch CAS case and its forerunners, https://repository.ubn.ru.nl/bitstream/handle/2066/289967/289967.pdf?sequence=1\&isAllowed=y

12. Assessing the viability and utility of predictive policing in Australia \- Australian Institute of Criminology, https://www.aic.gov.au/sites/default/files/2023-04/crg\_3216-17\_assessing\_the\_viability\_of\_predictive\_policing\_v4\_-\_april.pdf

13. The Politics and Biases of the “Crime Anticipation System” of the Dutch Police \- CEUR-WS.org, https://ceur-ws.org/Vol-2103/paper\_6.pdf

14. Predictive policing: The impact of crime forecasting technology on the performance of the Dutch police force \- Essay \- UT Student Theses, https://essay.utwente.nl/93249/1/Westenberg\_MA\_BMS.pdf

15. Predictive Policing Using AI & ML for Domestic Law Enforcement: Critical Analysis & Framework Development in EU, https://dspace.cuni.cz/bitstream/handle/20.500.11956/187387/120460620.pdf?sequence=1\&isAllowed=y

16. AUTOMATING INJUSTICE: \- Police and Human Rights Resources, https://policehumanrightsresources.org/content/uploads/2021/09/Automating\_Injustice.pdf?x19059

17. Predictive Policing and Crime Control in The United States of America and Europe: Trends in a Decade of Research and the Future of Predictive Policing \- MDPI, https://www.mdpi.com/2076-0760/10/6/234

18. Artificial Intelligence and Countering Violent Extremism: A Primer, https://gnet-research.org/wp-content/uploads/2020/10/GNET-Report-Artificial-Intelligence-and-Countering-Violent-Extremism-A-Primer\_V2.pdf

19. BALANCING SURVEILLANCE BETWEEN NEEDS OF PRIVACY AND SECURITY:, https://cale.law.nagoya-u.ac.jp/en/wp-content/uploads/2023/03/CALE20DP20No.206-2011.08.22.pdf

20. Predictive Policing in Practice: A Case Study of Chicago's Strategic Subject List \- Scholarship @ Claremont, https://scholarship.claremont.edu/cgi/viewcontent.cgi?article=5092\&context=cmc\_theses

21. The Rise and Fall of a Predictive Policing Pioneer \- AlgorithmWatch, https://algorithmwatch.org/en/predictive-policing-pioneer-keycrime/

22. ITALY \- AlgorithmWatch, https://algorithmwatch.org/en/automating-society-2019/italy/

23. CRIME prevention and predictive analysis: the italian case \- Agenfor International, https://www.agenformedia.com/wp-content/uploads/2020/10/Technology1\_VCinelli.pdf

24. CHINA'S TECHNOGENOCIDE \- Campaign For Uyghurs, https://campaignforuyghurs.org/wp-content/uploads/2026/04/Capstone-1-1.pdf

25. China: Big Data Fuels Crackdown in Minority Region \- Human Rights Watch, https://www.hrw.org/news/2018/02/27/china-big-data-fuels-crackdown-minority-region

26. Japan to Launch AI System to Predict Crime | GlobalSpec, https://insights.globalspec.com/article/8105/japan-to-launch-ai-system-to-predict-crime

27. Achieving Equity with Predictive Policing Algorithms: A Social Safety Net Perspective \- OUCI, https://ouci.dntb.gov.ua/en/works/40LgqGy4/

28. Predicting Violent Crime Hotspots \- RIT Digital Institutional Repository, https://repository.rit.edu/cgi/viewcontent.cgi?article=13548\&context=theses

29. Constraining Big Brother: The Legal Deficiencies Surrounding Chicago's Use of the Strategic Subject List, https://legal-forum.uchicago.edu/print-archive/constraining-big-brother-legal-deficiencies-surrounding-chicagos-use-strategic

30. Full article: Data integration and analysis platforms as digital platforms: a conceptual proposal \- Taylor & Francis, https://www.tandfonline.com/doi/full/10.1080/1369118X.2024.2442394

31. What is Brief History of Palantir Technologies Company?, https://businessmodelcanvastemplate.com/blogs/brief-history/palantir-technologies-brief-history

32. Palantir Technologies: A Strategic Analysis | PDF | Artificial Intelligence \- Scribd, https://www.scribd.com/document/961079617/Palantir-technologies

33. Precision and accuracy of acoustic gunshot location in an urban environment \- SoundThinking, https://www.soundthinking.com/wp-content/uploads/2021/08/TN-098-Accuracy-of-Acoustic-Gunshot-Location.pdf

34. GUEST EDITOR Ivana Luknar AI AND POLICY Ivana Damnjanović, Marko Pejković, Serdar Selim, İsmail Yılmaz, Pınar Kınıklı, V, https://www.ips.ac.rs/assets/pdf/SPM%206%20sa%20koricom.pdf

35. Body camera stories at Techdirt., https://www.techdirt.com/tag/body-camera/

36. Policing Agency Data Trusts \- Scholarly Commons, https://scholarlycommons.law.northwestern.edu/cgi/viewcontent.cgi?article=1633\&context=nulr

37. Argentina's AI Crime Prediction Initiative Raises Human Rights, https://oecd.ai/en/incidents/2024-08-02-00e4

38. AI in law enforcement and disaster risk management: Governing with Artificial Intelligence | OECD, https://www.oecd.org/en/publications/2025/06/governing-with-artificial-intelligence\_398fa287/full-report/ai-in-law-enforcement-and-disaster-risk-management\_99fc1804.html

39. Why does data governance matter for smart cities? \- OECD, https://www.oecd.org/en/publications/smart-city-data-governance\_e57ce301-en/full-report/component-4.html

40. PREDICTIVE POLICING AND CONSTITUTIONAL ... \- HeinOnline, https://heinonline.org/hol-cgi-bin/get\_pdf.cgi?handle=hein.journals/lwfyrinl3§ion=79

41. Urban Future With a Purpose \- Living in EU, https://living-in.eu/sites/default/files/files/deloitte-urban-future-with-a-purpose-study-set2021.pdf

42. Simultaneous Causality and the Spatial Dynamics of Violent Crimes as a Factor in and Response to Police Patrolling \- MDPI, https://www.mdpi.com/2413-8851/8/3/132

43. (PDF) Predicting Public Violent Crime Using Register and OpenStreetMap Data: A Risk Terrain Modeling Approach Across Three Cities of Varying Size \- ResearchGate, https://www.researchgate.net/publication/385424213\_Predicting\_Public\_Violent\_Crime\_Using\_Register\_and\_OpenStreetMap\_Data\_A\_Risk\_Terrain\_Modeling\_Approach\_Across\_Three\_Cities\_of\_Varying\_Size

44. Predicting Public Violent Crime Using Register and OpenStreetMap Data, https://d-nb.info/1353907147/34

45. Self-exciting point process modelling of crimes on linear networks \- IRIS UniPA, https://iris.unipa.it/retrieve/e3ad8928-0794-da0e-e053-3705fe0a2b96/hawkes\_processes\_on\_networks\_for\_crime\_data%20%283%29.pdf

46. Cellebrite 20-F 2023, https://investors.cellebrite.com/static-files/c69d1e3a-363f-479c-af47-b3437a990ec5

47. “A DIGITAL PRISON” \- Amnesty International, https://www.amnesty.org/ar/wp-content/uploads/2024/12/EUR7088132024ENGLISH.pdf

48. A temporal Hawkes process model for shooting occurrences in Sweden \- Lund University Publications, https://lup.lub.lu.se/student-papers/record/9145774/file/9145776.pdf

49. Self-exciting point process modeling of crime \- UCLA Statistics & Data Science, http://www.stat.ucla.edu/\~frederic/papers/crime1.pdf

50. Bayesian Modeling of Self-Exciting Point Processes with Missing Temporal Histories \- OSTI.GOV, https://www.osti.gov/servlets/purl/1429783

51. Requirements for AI in Production in Insurance Underwriting \- Palantir Blog, https://blog.palantir.com/requirements-for-ai-in-production-in-insurance-underwriting-04f7c1eed13d

52. Emerging threats in AI: a detailed review of misuses and risks across modern AI technologies \- Frontiers, https://www.frontiersin.org/journals/communications-and-networks/articles/10.3389/frcmn.2025.1727425/full

53. From Theory to Practice: Implementing MITRE ATLAS Defenses \- Medium, https://medium.com/@michael.hannecke/from-theory-to-practice-implementing-mitre-atlas-defenses-77dde19f9769

54. Model Threat Landscape \- AI Secured by Design, https://aisecuredbydesign.io/building/model-selection/threat-landscape/

55. Strengthening AI Critical Infrastructure Security with the MIT AI Risk Repository and MITRE ATLAS Frameworks \- Academic Conferences & Publishing International, https://papers.academic-conferences.org/index.php/eccws/article/download/3713/3265/13146

56. 2\. Input threats | AI Exchange, https://owaspai.org/docs/2\_threats\_through\_use/

57. Case Studies | MITRE ATLAS™ :: Reader View, https://cdn.lawreportgroup.com/acuris/files/ACR-New/AI%20Incidents%20MITRE%20Case%20Studies.pdf

58. AUTOMATING INJUSTICE: \- Fair Trials, https://www.fairtrials.org/app/uploads/2021/11/Automating\_Injustice.pdf

59. Self-Exciting Point Process Modeling of Crime \- ResearchGate, https://www.researchgate.net/publication/227369342\_Self-Exciting\_Point\_Process\_Modeling\_of\_Crime

60. Keeping Score: Predictive Analytics in Policing \- ProHIC, https://prohic.nl/wp-content/uploads/2020/11/5-23maart2020-PredictivePolicingAnnualreviewCriminology.pdf

61. A Systematic Review of Multi-Scale Spatio-Temporal Crime Prediction Methods \- MDPI, https://www.mdpi.com/2220-9964/12/6/209