Artificial intelligence tools originally developed for predictive policing have found new life in retail and hospitality loss prevention. Predictive policing systems are traditionally used to forecast crimes across geographic areas. Now, retailers have begun to deploy computer-vision systems, facial recognition, and behavioral analytics to identify which shoppers or guests pose elevated theft risks. This technological evolution represents both a significant opportunity and a burgeoning legal challenge for the retail industry.

Traditional loss prevention strategies for brick-and-mortar retailers relied heavily on physical deterrents: security personnel, alarm systems, and surveillance cameras. However, these passive methods have proven increasingly inadequate in an era of expanded self-service options. According to 2024 a National Retail Federation survey of retail executives at 164 retail brands,  shoplifting incidents rose 93% from 2019 to 2023, and dollar losses increased by 90% over the same period.

Retail predictive loss prevention systems mirror the architecture of predictive policing tools. Early predictive policing platforms analyzed historical incident data to generate "hot spots" (locations and times with elevated crime risk) and sometimes created person-based risk scores using various profiling criteria. The retail adaptation follows a similar model within the closed environment of a physical store.

These systems typically employ ceiling-mounted cameras and shelf sensors that feed video data into sophisticated algorithms capable of multiple functions: (1) identifying individuals through facial recognition or appearance tracking; (2) identifying products via barcode scanning or visual product detection; and (3) interpreting customer actions such as item pickup, concealment, scanning, or store exit without payment. The system then processes these inputs through artificial intelligence models to generate real-time risk scores and alerts. Common outputs may include notifications such as "this checkout lane likely missed a scan," "this customer matches an existing trespass notice," or "this shopping basket's contents don't reconcile with scanned items."

The practical significance of these systems lies in their influence on real-time decisions throughout the retail environment. AI-generated alerts can determine whether a store employee approaches a customer, whether a transaction requires re-scanning at self-checkout, or whether management detains someone for shoplifting.

However, the relative rarity of actual theft events caught by AI-generated systems compared to overall customer transactions creates a fundamental challenge. Even systems with strong reported accuracy rates will inevitably produce false positives. Without robust review protocols, this dynamic can lead to unnecessary and potentially discriminatory customer confrontations.

More concerning is the potential for these systems to perpetuate and amplify historical prejudices embedded in their training data. AI models trained on past apprehension records, trespass notices, or staff incident reports will naturally learn and replicate the patterns present in that historical data, including any embedded biases. For example, if store employees historically scrutinized and disproportionately detained Black male customers, an AI system trained on those records may continue or even accelerate this discriminatory pattern. Unless retailers actively measure and correct for such disparities, disparate impact becomes not only a hypothetical risk, but also the system's default operational mode.

These operational realities create immediate and significant legal exposure across multiple areas of law. Federal and state civil rights statutes can be triggered when AI systems produce error rates or intervention rates that disproportionately affect protected classes, regardless of discriminatory intent. The disparate impact doctrine makes clear that neutral policies producing discriminatory outcomes can violate civil rights laws.[1] Identity-matching features raise complex compliance obligations under state privacy laws and biometric statutes. Jurisdictions like Illinois, Texas, Washington and New York City impose requirements for notice, consent, and limitations on biometric data retention and sharing. Perhaps most immediately dangerous are false arrest and false imprisonment claims that arise when AI alerts are treated as sufficient grounds for customer detention.

Given these substantial legal exposures, retailers implementing predictive loss prevention systems must develop comprehensive risk management strategies that extend beyond the technology itself. Retailers must work to ensure their insurance coverage adequately addresses AI-related risks. Traditional general liability policies may exclude discrimination claims, privacy violations, or technology errors and omissions. Insurance coverage must be paired with robust operational protocols with clear escalation procedures that require human review before any customer detention. Regular bias testing and algorithmic auditing should also be implemented to test how the model is functioning over time. Retailers who are not creating their artificial intelligence models and systems in-house must also carefully evaluate AI system vendors' compliance programs, data security practices, and willingness to provide indemnification for system failures, harm to third-parties, or regulatory violations.

Predictive loss prevention technology offers retailers powerful tools to address the growing challenge of inventory loss. However, the legal risks are substantial and evolving. Success in this area requires not only technological sophistication, but also comprehensive risk management that includes appropriate insurance coverage, operational oversight, and proactive review of relevant law. Retailers who treat these systems as purely technological solutions, rather than complex risk management challenges, do so at their own peril.

[1] See Nathan Freed Wessler, Police Say a Simple Warning Will Prevent Face Recognition Wrongful Arrests. That's Just Not True, ACLU (Apr. 30, 2024), link; Jannice Cebreros, Facial Recognition Technology and Wrongful Arrests in the Digital Policing Era, 100 WASH. L. REV. ONLINE 33, 36-41 (2025).