AI in Policing: What's Already Here, What's Coming Next

AI
Written By Don Bradford
AI in Policing: What's Already Here, What's Coming Next

AI in Policing: What's Already Here, What's Coming Next

LodiEye — July 2026

Overview

Artificial intelligence already reads license plates, drafts police reports, extracts and analyzes seized phones, mines social media, and flags people as future suspects in California departments, including in San Joaquin County. This report organizes the technology into four categories: surveillance and identification, predictive and risk analysis, records and evidence processing, and the real-time crime centers now fusing all three into a single dashboard.

  • Surveillance tools like license-plate readers and social-media mining platforms are the most mature and the most widely deployed today.
  • Predictive and risk-scoring tools are moving from pilots into active use, despite a documented history of flawed data driving them.
  • AI report-writing tools now draft first-version reports from body-camera audio in minutes, with officers reviewing before signing.
  • Real-time crime centers are quietly combining all of the above into single dashboards, a convergence point civil liberties groups now consider the biggest emerging risk.
  • Lodi Police already operate license-plate cameras, a multi-drone program dating to 2018, and a Flock Safety gunshot-detection system; a dedicated section covers what is in use locally and where the department has signaled it is headed.

Police AI Tools by Category and Deployment Status, 2026

Source: LodiEye analysis of vendor documentation, court filings, and legislative records

Category 1: Surveillance and Identification

This is the most mature category of police AI, and the most widely deployed. These tools watch, listen, and identify, feeding everything they collect into searchable databases.

Deployed today

License-Plate Reader Networks

Flock Safety and similar vendors sell camera networks that photograph every passing vehicle, convert the plate to text, and check it against hotlists for stolen vehicles and Amber Alerts. A single mid-sized city's network can log hundreds of thousands of vehicle detections in a month, the overwhelming majority unconnected to any crime. Flock has since expanded beyond plate-reading: EFF's 2025 investigations found the company building tools that flag vehicles based on driving patterns alone, not just plate matches, and is developing audio-detection features for its camera network.

No statewide law governs how these cameras share data between agencies. Some cities restrict sharing to in-state law enforcement; others participate in national lookup networks that can extend access to federal agencies. Several California cities discovered gaps between their stated sharing policy and their actual data flows only after commissioning an independent audit.

Deployed today

Social Media and Open-Web Analysis

Some California sheriff's offices, including San Joaquin County's, use commercial platforms that search the open web, the deep web, and the dark web for information tied to investigations. San Joaquin's contract with Cobwebs Technologies for a platform called "Tangles," which includes facial-recognition capability, is currently the subject of litigation: the Electronic Frontier Foundation is trying to unseal the contract terms, and the vendor itself sued to keep them secret as trade secrets.

Deployed today

Facial Recognition

California banned pairing facial recognition with body cameras in 2020. That ban expired in January 2023 and was never renewed. Today, individual cities set their own policies. San Francisco, Berkeley, and Oakland ban facial recognition outright. Most California cities have no local policy on it in either direction, meaning a department could adopt it tomorrow without a public vote.

Deployed today Emerging features

Device and Communications Analysis

When police seize a phone or computer under warrant, AI now reads it far faster than any human analyst could. Cellebrite's "Guardian Investigate" platform interrogates call records, text messages, images, and video pulled from a device, and in one Texas sheriff's pilot summarized a phone holding 200,000 text messages in a fraction of the time a manual review would take. It maps a device's location over time by combining cell-site data with mapping software and builds timelines that link messages, social posts, and video into a single sequence. A related tool, Cellebrite Genesis, surfaced evidence of a planned attack in Australia in three minutes, work analysts estimated would have taken two to three weeks.

The newer, emerging features push further. AI agents now try to identify a device's owner by analyzing the email addresses used to log into its apps and cross-referencing them against open-source web data, and flag behavioral anomalies, such as two people in regular contact who suddenly stop, or a phone abruptly switched to airplane mode. One distinction matters throughout this category: nearly all of it is forensic analysis of devices already seized under warrant, which is legally separate from real-time interception of live messages. Parallel compliance platforms like Smarsh capture and index communications across voice, email, chat, and mobile in regulated industries, using language models to detect intent rather than just keywords, the same underlying technology aimed at a different target.

California Statewide Rules by Police AI Tool, 2026

Source: California legislative records; coverage scored 0 (none) to 3 (full)

Category 2: Predictive and Risk Analysis

Where surveillance tools record what already happened, this category tries to forecast what happens next, either by location or by person. It is also the category with the most documented failures.

Deployed today

Place-Based Prediction

Software maps historical crime data to identify "hot spots" by time and location, directing patrols toward areas the model flags as high-risk. Named commercial tools in this space have included PredPol and ShotSpotter Missions (formerly HunchLab), which combine statistical modeling with real-time crime feeds. Related "risk terrain modeling" tools map environmental factors, like liquor stores or vacant buildings, to crime probability, treating geography rather than people as the input.

Deployed today

Person-Based Risk Scoring

Other tools score individuals as likely future offenders or victims using arrest history and social-network analysis, attempting to map associations between people the way an org chart maps a company. The Los Angeles Police Department's own Inspector General audited its person-based predictive program and found "significant inconsistencies" in the underlying data: half the people flagged had few or no ties to the crimes the system targeted. LAPD scrapped the program in response. Santa Cruz banned predictive policing outright, becoming the first U.S. city to do so.

Deployed today Emerging

Social Network Analysis and Proactive Social-Media Monitoring

Person-based prediction increasingly runs on social-media data. Social network analysis maps the connections between people in gangs or co-offending groups and ranks who is most likely to commit, or fall victim to, violence next. Big-data platforms feed crime reports, arrest records, and public social-media activity into a single analytical stream, and counterterrorism systems monitor communication and travel patterns to try to flag attack planning before it happens.

This is the most forward-leaning and most contested application in the category, because it shifts policing from responding to a reported crime toward acting on a prediction about a crime that has not occurred. The same LAPD data-quality problems that sank its person-based program apply here with higher stakes: a flawed inference drawn from who someone follows or messages can flag a person with no involvement in any offense.

Emerging

Deception and Emotion Detection

A newer and more speculative application attempts to use AI to detect deception or emotional state during interviews or interrogations. The underlying science is unproven, and no California department has publicly disclosed operational use, but vendors are actively marketing the capability to law enforcement buyers.

LAPD Predictive Policing Audit: Flagged Individuals by Documented Ties to Crime

Source: Los Angeles Police Department Office of the Inspector General

No state law currently prevents a new department from adopting a predictive or risk-scoring tool. LAPD's and Santa Cruz's experiences did not become binding limits on other cities, they became cautionary case studies that other departments are free to ignore.

Category 3: Records, Reports, and Evidence Processing

This category turns raw material, audio, text, and case files, into structured records officers and prosecutors can use. It is the only category with a dedicated California statute.

Deployed today

AI-Written Police Reports

Software like Axon's "Draft One" turns body-camera audio into a first-draft police report in minutes. Officers review and edit before signing. SB 524, now Penal Code Section 13663, took effect January 1, 2026, and requires a disclosure statement on the report, permanent retention of the AI's original draft, and an audit trail showing who ran the tool and what footage fed it. Except for the officer's final signed version, the law says an AI draft "shall not constitute an officer's statement."

The gap between the law and the software is already visible. Draft One was not originally built to record which words came from the AI and which came from the officer. San Jose's police department ordered its officers, in April 2026, to stop using AI to write reports until the software could meet the new law.

Deployed today

Evidence Management and Case Analysis

Departments increasingly use AI to organize and search across large volumes of case evidence, body-camera footage, and digital records, work that used to require staff manually reviewing files. Natural-language processing tools extract patterns from police reports, tip lines, and social-media posts, surfacing connections a human reviewer might miss, or might take weeks to find manually.

Where It All Converges: Real-Time Crime Centers

Convergence point

One Dashboard, Every Feed

Real-time crime centers are centralized hubs that fuse camera feeds, license-plate data, 911 records, and facial recognition into a single dashboard, monitored live by staff who can direct officers in the field. Vendors including Axon, Motorola, Flock, and Genetec now market these as integrated suites rather than standalone products. Flock's newer "Nova" platform reportedly merges plate data with 911 dispatch records, criminal databases, and third-party data sources into one searchable system.

This convergence is why civil liberties groups increasingly treat real-time crime centers, not any single tool, as the central risk. A license-plate camera alone raises one set of questions. The same camera feeding a system that cross-references it against social-media activity, arrest history, and a predictive risk score raises a different one, and almost no jurisdiction has written rules for the combined system rather than its individual parts.

The Feeds Flowing Into a Crime Center

A real-time crime center is only as powerful as the streams it pulls together. Each feed below is a standalone surveillance tool in its own right; the crime center's value, and its risk, comes from cross-referencing them against one another in real time. The chart shows how continuously each feed updates, from constant camera streams to periodic database pulls.

  • License-plate camera feeds: a constant stream of vehicle detections, updating the moment any camera in the network reads a plate, often hundreds of thousands of reads per day across a city.
  • Live video and CCTV: continuous footage from fixed cameras, business partnerships, and sometimes officer body cameras, increasingly paired with object and facial recognition.
  • 911 and computer-aided dispatch: incoming calls and dispatch records that place an incident on the map the instant it is reported, driving where analysts point the other feeds.
  • Facial recognition matches: identity checks run on demand against photo databases when an image is captured or uploaded.
  • Criminal and records databases: arrest histories, warrants, and case files pulled in periodically to attach context to a person or vehicle.
  • Seized-device extractions: messages, call logs, images, and location history pulled from phones and computers under warrant, then analyzed by AI, added to the picture when a device is in custody rather than continuously.
  • Social-media and network analysis: public posts and mapped connections between people, used both to add context to an individual and, more controversially, to flag potential future activity.
  • Third-party and web data: commercial data brokers and web-mining results, the slowest-updating and least standardized feed, but often the most personal.

Data Feeds a Real-Time Crime Center Combines, by Update Frequency

Source: LodiEye analysis of Axon, Motorola, Flock, and Genetec product documentation

The Constitutional Backdrop

The U.S. Supreme Court's 2018 ruling in Carpenter v. United States found that police need a warrant to gather "deeply revealing" location data, since it offers "an intimate window into a person's life." That ruling was written for cell-site records, not license-plate networks or predictive-policing databases, and no California court has yet applied its reasoning directly to either. Legal scholars and advocacy groups expect that test to come, particularly as camera networks and predictive tools generate the same kind of comprehensive location history the Court found constitutionally significant in 2018.

A parallel effort is underway at the federal level. A proposed Federal Rule of Evidence 707 would subject AI-generated evidence to the same reliability standards used for expert testimony, a response to the difficulty of cross-examining a report no single person fully wrote.

Quick Reference: Key Terms

ALPR (Automated License Plate Reader)
A camera that photographs license plates, converts the image to text, and checks it against watchlists automatically.
Hotlist
A list of plates tied to stolen vehicles, missing persons, or Amber Alerts, used to trigger an alert on a match.
Real-time crime center
A centralized hub that fuses camera feeds, plate data, 911 records, and other databases into one live dashboard.
Predictive policing
Software that analyzes past crime data to forecast where crimes will happen or who might commit them, either by place or by person.
Risk terrain modeling
A place-based predictive method that scores locations by environmental risk factors rather than by individual history.
Social network analysis (SNA)
A method that maps connections between people to identify who is most central to a group and most likely to be involved in future activity.
Forensic device extraction
Pulling and analyzing the contents of a seized phone or computer under warrant, distinct from intercepting live communications.
Hallucination
A factual error an AI system invents that was not present in its source material.
Carpenter v. United States (2018)
Supreme Court ruling that police generally need a warrant to obtain data revealing a person's location history.
National Lookup
A Flock feature letting one agency search plate data collected by other agencies' cameras, sometimes nationwide, depending on each agency's settings.

Close to Home: What Lodi Police Have Deployed

Lodi Police Department already runs several of the tools described above, and its own public records document what is in use today and where it has said it wants to go.

Deployed today

License-Plate Reader Network

As of April 2024, Lodi PD operated 14 Flock Safety automated license-plate readers and stored the data for 30 days, according to the Atlas of Surveillance; the department's own transparency figures have since shown a larger camera count. Lodi publishes an ALPR usage and privacy policy and a Flock transparency portal on the city website. The cameras capture plates and vehicle images but, per the department's stated configuration, do not run facial recognition.

Deployed today

Drone Program (UAS)

Lodi has run a drone program since 2018, when the city council approved it after a department presentation. The team has grown to eight FAA-licensed sworn pilots, used to clear rooftops and yards for suspects, document crime and accident scenes, search for missing persons, and support the fire department with infrared "hot spot" detection during structure fires. In July 2026, the department publicly notified residents it would be flying drones over the city for a stretch of enforcement operations.

Confirmed Drone Models

The city's official Drone Team page names one model directly and confirms two additional aircraft without identifying them:

  • DJI Phantom 4 — the department's first drone, donated by the Lodi Police Foundation (which provided $2,500 for the purchase) and approved by the City Council in 2018. It stays airborne about 30 minutes per flight.
  • Two additional drones — the city states the department "has added two more drones" since the Phantom 4, but does not publish their make or model. Their infrared "hot spot" capability, cited by the department, indicates at least one carries a thermal camera, a feature common to public-safety models such as DJI's Matrice and Mavic Enterprise thermal lines, though Lodi has not confirmed which.

Accuracy note: Only the DJI Phantom 4 is confirmed by name in Lodi's own records. The makes and models of the other two drones are not disclosed publicly as of this report; any specific model beyond the Phantom 4 would need to be verified through a public-records request to the department.

Deployed today

Gunshot Detection — Flock Safety Raven

Lodi PD uses Flock Safety's "Raven" gunshot-detection system, according to the Atlas of Surveillance. Raven works through acoustic sensors mounted on street lights or the sides of buildings that passively listen for gunfire and alert officers to shots that often go unreported to 911. The choice keeps Lodi's detection and its license-plate cameras under a single vendor: a gunshot alert from a Raven sensor can be paired with vehicle data from the department's Flock ALPR network, letting officers connect the sound of gunfire to vehicles moving nearby.

Local reporting at the start of 2026, when Lodi recorded its lowest homicide rate in eight years, described the city as running gunshot detection alongside its Flock cameras, and the department has credited its combined technology tools as a factor in recent enforcement results.

Documented direction

What Lodi Has Signaled Next

Lodi has not published a formal plan to adopt AI report-writing, predictive policing, or a drone-as-first-responder program. Its documented trajectory so far is incremental: more license-plate cameras than it started with, a drone fleet it has repeatedly said it intends to grow with more pilots and more capable aircraft, and a stated interest in higher-end drone features like zoom and thermal imaging. The larger moves happening elsewhere in the region, such as Stockton's 2026 approval of a multi-million-dollar drone-as-first-responder expansion, show the direction a department of Lodi's type could follow, but Lodi has not announced anything comparable as of this report.

Where This Is Heading

Four shifts define the near-term direction of AI in policing, based on what vendors are building and what departments are buying right now:

  • From standalone tools to fused platforms. The market is moving away from single-purpose products toward integrated suites that combine plate reading, facial recognition, records, and prediction in one system.
  • From reactive to predictive. Tools that document what happened are giving way to tools that forecast what might, both by location and by person, even where the underlying data has proven unreliable.
  • From ground cameras to autonomous response. AI-directed drones dispatched to 911 calls ahead of officers are expanding from pilots into standing programs, with Stockton's 2026 expansion as an early California example.
  • From human-written to machine-drafted records. AI report-writing is spreading quickly, and the products are racing to add the audit trails and disclosure features that some departments now require before they will deploy them.

LodiEye is the original civic-reporting and analysis arm of Lodi411.com, a citizen-run civic data and transparency platform serving Lodi, California and San Joaquin County. LodiEye gathers information of public interest, applies editorial judgment to public records, meetings, and data, and publishes original explanatory reporting for its readers — the work of a newsroom, and a representative of the news media as that term is defined under federal law. Our reporting emphasizes primary sources, public data, and full source transparency so readers can check every claim. LodiEye complements, and does not replace, the other outlets covering this region; for additional reporting on Lodi, San Joaquin County, and the broader region, we also encourage readers to consult the Lodi News-Sentinel, Stocktonia, The Sacramento Bee, CalMatters, and other established news organizations. Our full editorial standards and news-media-status statement is published at lodi411.com/editorial-standards.

This LodiEye report was produced using artificial intelligence tools under the direction and review of the founder. Lodi411 uses multiple AI platforms in its research and publication workflow, including Anthropic's Claude (primarily Opus and Sonnet models) and Perplexity AI across a variety of large language models offered by each. These tools were used in the following capacities:

Source Discovery: AI-assisted search identified more than 25 primary sources, including California legislative records, court filings in Pen-Link v. County of San Joaquin, EFF's Flock Safety investigations, and reporting from NPR, KQED, ABC7, the Marshall Project, and Stocktonia. Perplexity AI handled initial source discovery and real-time data retrieval; Claude conducted deeper analysis of the identified sources.

Credibility Validation: AI cross-referenced claims across independent sources, prioritizing primary government and court records first, followed by nonprofit legal-advocacy reporting (EFF, ACLU, Brennan Center) and established news organizations. Multiple AI models independently verified key figures and flagged inconsistencies for review.

Analysis and Synthesis: Claude Opus and Sonnet assisted in organizing police AI tools into four categories, surveillance and identification, predictive and risk analysis, records and evidence processing, and real-time crime centers, and in mapping which categories carry statewide regulation versus none.

Presentation: Claude assisted in drafting and structuring the report, including the category-overview, regulatory-coverage, LAPD-audit, and data-feed charts, all built with Kendo UI, each sourced directly to California legislative records or the news reports cited below.

Final Review: Multiple AI models reviewed the completed draft for factual consistency, source attribution accuracy, logical coherence, and balanced presentation. Throughout the process, the editor set the report's goals, scope, and tone; shaped and revised the draft; integrated independent fact checks; and reviewed the AI cross-checks and validations. Multi-tool cross-checking across independent models and sources is the primary error-reduction mechanism.

Lodi411/LodiEye believes that transparency about how our research is produced — including our use of AI under human direction — strengthens trust with readers and the broader information ecosystem. Readers who spot an error are encouraged to write editor@lodi411.com so we can correct it.

References

Questions or corrections: editor@lodi411.com

Don Bradford
Previous

The Fertilizer Emergency: What Trump's Phosphate Declaration Means for U.S. Crops
Next

The Paradox of Plenty: U.S. Dairy and San Joaquin County