AI-Driven Predictive Policing: From Statistical Scores to Real-World Arrests

AI-based policing tools are deployed in dozens of U.S. cities, yet no public registry documents their full footprint. These systems ingest historical crime data and assign risk scores to neighborhoods, directing officers toward algorithm-generated “hot spots.” The rapid expansion of such technology outpaces public oversight, leaving the scale of its use largely invisible to citizens and policymakers.

The Conversation

Predictive models do not deliver binary answers; they output probabilities and confidence levels. Engineers set a confidence threshold—often a hidden “control knob”—that determines when an alert triggers police deployment. A 95 % confidence setting signals a high likelihood of threat, while lower thresholds generate more alerts but increase false alarms. Vendors and agencies typically configure these thresholds without public disclosure, converting statistical outputs into operational decisions.

The Conversation

The choice of threshold creates a trade-off akin to medical diagnostics. A lower threshold may enable earlier intervention but raises the risk of mistaken identifications, exemplified by cases such as Angela Lipps and the escalated encounter involving Taki Allen. Conversely, a higher threshold reduces wrongful alerts but can miss genuine threats. Researchers suggest calibration techniques—receiver operating characteristic (ROC) curve analysis and precision-recall analysis—to quantify how threshold adjustments affect true-positive and false-positive rates, though these methods cannot resolve the societal question of acceptable algorithmic uncertainty.

The Conversation

Courts apply established standards of proof—probable cause, preponderance of the evidence, beyond a reasonable doubt—to ensure that evidence meets a defined certainty before legal authority is exercised. AI models, by contrast, rarely express uncertainty; they present a confidence score even when the answer is incorrect. This mismatch raises concerns as AI outputs begin to inform law-enforcement actions, courtroom arguments, and public-sector decisions without the evidentiary safeguards that traditional legal processes require.

The Conversation

Scholars caution that treating probabilistic predictions as definitive can standardize police actions absent further verification. While some researchers argue that predictive policing does not increase arrests of racial minorities relative to traditional methods, the broader issue is the institutionalization of algorithmic suspicion. Studies by the University of Virginia’s Digital Technology for Democracy Lab reveal that some police departments operate under strict policies governing technology use, whereas others lack any formal guidance, amplifying the risk of wrongful arrests and convictions.

The Conversation

• “We are researchers who study the intersection of technology, law and public administration.” — Lead author, The Conversation article

The Conversation

• “cities, although no public registry tracks the full footprint.” — Research team

The Conversation

• “When generative AI models such as ChatGPT or Claude respond to human requests, they are not searching a database and pulling out facts.” — Study authors

The Conversation

• “The danger arises when people believe that the model is retrieving truth rather than generating likelihoods.” — Researchers

The Conversation

• “It is important to realize that these thresholds are not neutral features of the technology; they are choices embedded by the creators in the model’s code.” — Authors

The Conversation

Future efforts must prioritize transparent registries of deployed AI tools, standardized policy frameworks for threshold setting, and mandatory training that emphasizes algorithmic uncertainty. Ongoing calibration research and independent audits could help align AI-generated risk scores with legal standards, reducing the likelihood of false arrests and wrongful convictions.

The Conversation