← seanbair.ai
Field Guide

AI in Public Safety
Field Guide

A practical reference for agency leaders deploying AI across public safety. 7 domains, 35 use cases, and the guardrails that make them safe to deploy. Organized around capability categories, not vendor products, so it stays useful as the technology moves.
Created by Sean Bair
Author of AI in Policing (BYU Academic Publishing, 2025) and Tactical Crime Analysis (CRC Press, 2009). 30+ years building AI for public safety. Founder of BAIR Analytics (acquired by LexisNexis) and CEO of SWORN.ai.
Questions? Get in touch
7 Domains
35 Use Cases
150+ Guardrails
6 Ethical Principles
Technique
Readiness
Risk Level

Start Here: 5 Quick Wins

New to AI in public safety? These five use cases have the best ratio of impact to risk, proven, low-risk, and ready to deploy today.
Dynamic Resource Positioning
Mature Risk: Low Emergency Response
Optimize ambulance and patrol staging to cut response times 15–30% without adding units. The math is proven and the risk is minimal.
Case Document Processing
Established Risk: Low Courts
Automate extraction from legal filings. Court clerks spend 60–70% of their time on document processing, AI slashes that dramatically.
Judicial Workload Forecasting
Mature Risk: Low Courts
Predict caseload volume to allocate judges and courtrooms efficiently. Some courts report 20–30% improvement in calendar efficiency.
Victim & Witness Support Automation
Mature Risk: Low Courts
Automate victim notifications and resource matching. Meets VOCA requirements more consistently than manual processes.
Environmental Design Analytics
Established Risk: Low Community Safety
Identify environmental features linked to crime. Greening vacant lots in Philadelphia reduced gun violence by 29%.

Law Enforcement & Policing

AI capabilities for patrol, investigation, analysis, and community engagement.
How It Works
Machine learning models trained on 3–5 years of calls-for-service data, weighted by time of day, day of week, weather, events, and spatial clustering. Outputs heat maps and staffing recommendations refreshed every shift.
Real-World Impact
Agencies see 10–20% reductions in response times and more equitable distribution of patrol presence. Shifts from reactive “drive around” policing to data-informed deployment.
Ethical Considerations
  • Models trained on biased arrest data can perpetuate over-policing in minority communities
  • Must use calls-for-service data (community-generated) not arrest data (officer-generated)
  • Recommendations must be advisory, officers retain discretion on where to patrol
  • Regular bias audits required to detect disparate impact by geography and demographics
Required Guardrails
  • Use calls-for-service data, not arrest or stop data, as the primary input
  • Conduct quarterly disparate impact analysis by census tract demographics
  • Publish methodology and aggregate results to the community
  • Human supervisor must approve resource allocation, AI recommends, humans decide
  • Establish a civilian oversight mechanism for algorithmic policing tools
From the Field
At Tempe PD, a college bar called Seymour Dogs gamed the crime stats by telling victims to call police from the Taco Bell next door. The bar looked clean while Taco Bell appeared to have an epidemic. We caught it through NLP analysis of dispatch narrative comments. The lesson I teach every class: garbage in, garbage out, just with more complex mathematical formulas attached. (AI in Policing, Ch. 3)
Key question for leaders: Are you using calls-for-service data or arrest data? The answer determines whether this tool reduces bias or amplifies it.
How It Works
Entity extraction from reports (NLP) builds a knowledge graph of people, places, vehicles, and events. Graph algorithms (community detection, shortest path, centrality) identify non-obvious relationships. Analysts query the graph through natural language interfaces.
Real-World Impact
Reduces time to identify suspect connections from days to hours. Surfaces cross-jurisdictional links that siloed databases miss. Clearance rates improve when analysts spend time on leads instead of data entry.
Ethical Considerations
  • Risk of guilt by association, graph proximity does not imply criminal involvement
  • Data retention policies must limit how long non-suspect contact information persists
  • Facial recognition integration raises significant civil liberties concerns
  • Must distinguish between intelligence (leads) and evidence (court-admissible)
Required Guardrails
  • Leads are investigative starting points, never probable cause
  • Document the analytical path from data to lead for court defensibility
  • Purge non-suspect data per agency retention schedule (28 CFR Part 23)
  • Require supervisory approval before acting on AI-generated leads
  • Annual audits for racial and demographic bias in lead generation patterns
From the Field
Officer Mike Kerney in Daytona Beach noticed that window screen cuts on burglary scenes curved in a “C” shape, indicating a left-handed offender. That single inference narrowed the suspect pool from thousands to 7% of the population. He made an arrest within two weeks connecting four cases. I tested the same evidence on ChatGPT and Claude, both independently reached the same conclusion. I stared at my screen in disbelief. The art of seeing what isn’t there is no longer uniquely human. (AI in Policing, Ch. 5)
Key question for leaders: Can your analysts explain the analytical path from data to lead in court? If not, the tool is a liability, not an asset.
How It Works
Speech-to-text converts body-camera audio. LLMs synthesize dispatch logs, CAD data, and transcribed audio into a structured narrative following agency report templates. Officers review, edit, and sign the final report.
Real-World Impact
Officers spend 30–45 minutes per report on average. AI-assisted drafting cuts this to 5–10 minutes of review. More time on patrol, more detailed reports, fewer errors from fatigue-driven writing.
Ethical Considerations
  • AI-drafted reports must be clearly marked as AI-assisted in the record
  • Officers must genuinely review and take ownership, rubber-stamping is dangerous
  • LLM hallucination risk means fabricated details could appear in official records
  • Body-camera audio capture must comply with state wiretapping and consent laws
Required Guardrails
  • Officers must review every sentence and certify accuracy under their name
  • AI-generated text must be flagged in the report management system metadata
  • Supervisors spot-check AI-assisted reports at a defined audit rate (e.g., 10%)
  • Agency policy must prohibit submitting AI drafts without substantive review
  • Retain original audio/data sources alongside the final report
From the Field
When I built AutoLog at Tempe PD, we discovered that standard radio dispatch captured only about 30% of what officers actually did during a shift. The other 70%, community contacts, problem-solving, self-initiated activity, disappeared. If your report writing system only captures what’s “important enough” to document, you’re losing the data that would have saved you. (AI in Policing, Ch. 8)
Key question for leaders: What is your audit rate for AI-assisted reports? If the answer is zero, your officers are rubber-stamping, and that will surface in court.
How It Works
Computer vision models detect anomalies in camera feeds (fights, abandoned objects, crowd surges). Acoustic sensors classify gunshot events. LPR systems flag wanted vehicles. All feeds aggregate into a GIS-based dashboard with automated alerting and correlation.
Real-World Impact
Reduces the gap between incident and response. Enables proactive deployment to developing situations. Provides commanders with situational awareness during critical incidents.
Ethical Considerations
  • Persistent surveillance infrastructure raises Fourth Amendment concerns
  • Camera placement patterns can reflect and reinforce racial profiling
  • Social media monitoring of protest activity threatens First Amendment rights
  • Data retention from mass surveillance creates privacy risks even for non-suspects
Required Guardrails
  • Publish camera placement maps and policies publicly
  • Prohibit facial recognition on live feeds without explicit policy and legal authorization
  • Social media monitoring restricted to specific, documented criminal investigations
  • Implement data minimization, auto-purge non-evidentiary footage per retention schedule
  • Annual civil liberties impact assessment with public reporting
From the Field
I worked the Super Bowl XXX Emergency Operations Center in Tempe. We coordinated thousands of officers with static maps and radio. Now imagine that same operation with Bayesian inference systems updating threat probabilities in real time, sentiment analysis on social media feeds, and AI-optimized resource allocation. The gap between what we had and what’s possible today is staggering, but so is the ethical responsibility that comes with it. (AI in Policing, Ch. 4)
Key question for leaders: Do your residents know where your cameras are and what they capture? Transparency is cheaper than litigation.
How It Works
NLP models analyze body-camera transcripts for escalation patterns in language (volume, word choice, command frequency). Workload algorithms track consecutive high-stress calls, overtime hours, and critical incident exposure. Outputs wellness risk scores and training recommendations to supervisors.
Real-World Impact
Early identification of officers at risk for burnout, excessive force, or mental health crises. Shifts culture from reactive discipline to proactive support. Reduces use-of-force incidents when combined with targeted training.
Ethical Considerations
  • Officer privacy, monitoring communication patterns can feel like surveillance
  • Union concerns about data being used punitively rather than supportively
  • Risk of penalizing officers for natural emotional responses to traumatic calls
  • Must be part of a supportive culture, not a gotcha system
Required Guardrails
  • Data used for wellness support only, explicitly prohibited from disciplinary proceedings
  • Officers informed about what is monitored and how data is used
  • Results go to peer support and wellness units, not internal affairs
  • Collective bargaining agreement provisions for AI monitoring
  • Voluntary participation where legally required; mandatory only with union agreement
From the Field
I built SWORN to solve this. One mid-sized department’s 18-month pilot showed a 22% reduction in sick leave, 31% decrease in workers’ comp claims, 18% improvement in retention, and $450,000+ saved in year one. A sergeant with 22 years on the job told me: “For the first time, I feel like the department actually gives a damn about me as a person, not just as a badge number.” One-size-fits-all wellness is about as effective as department-issue uniforms, technically they fit everyone, but they look good on no one. (AI in Policing, Ch. 8 & 12)
Key question for leaders: Is your department culture ready for this? If officers see it as surveillance, it will backfire. The tool works only inside a culture of support.

Questions to Ask Your Vendor: Law Enforcement & Policing

  1. Can you show me a bias audit on your system’s outputs, broken down by race and geography?
  2. What data does your system use as inputs, calls-for-service or arrest data? (If arrest data, walk away.)
  3. What happens to our data if we cancel the contract? Do you retain it?
  4. Has this been validated on an agency our size, in a jurisdiction like ours?
  5. Can a patrol officer explain your system’s recommendation to a community member in plain language?
  6. What is your false positive rate, and how do you measure it?
  7. Do you have an independent third-party audit of your algorithm? Can I see it?

Emergency Response & Dispatch

Intelligent dispatch, resource optimization, and real-time situational awareness for fire, EMS, and 911 centers.
How It Works
Speech recognition processes caller audio while NLP classifies the incident type. Audio analysis detects distress indicators (breathing patterns, background sounds). Models recommend priority level and unit type (patrol, fire, EMS, mental health) based on classification.
Real-World Impact
Faster triage during peak call volume. Reduced over-dispatch (sending 4 units when 2 will do). Better identification of medical emergencies masked as other call types.
Ethical Considerations
  • Misclassification can delay response to life-threatening emergencies
  • Accent and language bias in speech recognition may disadvantage non-English callers
  • Over-reliance on AI triage could reduce dispatcher critical thinking
Required Guardrails
  • AI recommendation displayed alongside, never replaces, dispatcher judgment
  • Multilingual capability mandatory; flag low-confidence language detection for human review
  • Fail-safe: any call AI cannot classify defaults to highest priority
  • Monthly accuracy audits comparing AI triage to actual incident outcomes
  • Dispatcher override rate tracked, if too low, investigate rubber-stamping
From the Field
Technology without strategy is just expensive junk collecting dust. I’ve seen tech rollouts tank because systems were designed for IT, not for the people who actually use them. If it doesn’t work for the dispatcher behind the console at 3 AM on a Saturday, it’s dead on arrival. (AI in Policing, Ch. 9)
Key question for leaders: What happens when the AI gets it wrong? If the answer isn’t “the dispatcher catches it,” your fail-safe is missing.
How It Works
Demand forecasting models predict call volume by geography and time. Optimization algorithms calculate ideal staging positions given current unit availability, traffic, and weather. Recomputed every 15–30 minutes or when unit availability changes.
Real-World Impact
15–30% improvement in average response times without adding units. More equitable coverage, eliminates coverage gaps that disproportionately affect underserved neighborhoods.
Ethical Considerations
  • Optimization targets (minimize average vs. minimize worst-case) embed value judgments
  • Historical demand patterns may under-serve communities that historically under-report
  • Union and personnel concerns about AI-directed posting locations
Required Guardrails
  • Optimize for equity (worst-case response time) not just efficiency (average response time)
  • Supplement historical demand data with demographic and access-to-care data
  • Dispatchers can override positioning recommendations for operational reasons
  • Track response time equity metrics by neighborhood demographics quarterly
From the Field
At Tempe PD, officers were assigned to specific geographic areas and acted like a mini Chief of Police for their community. That model created accountability and local knowledge. AI-driven positioning should enhance that model, not replace it, because an algorithm can optimize a staging location, but it can’t build the trust that comes from an officer who knows every business owner on the block. (AI in Policing, Ch. 14)
Key question for leaders: Are you optimizing for average response time or worst-case response time? The first looks good on paper; the second saves lives in underserved areas.
How It Works
Trained on historical incident data linking initial dispatch codes to actual outcomes. Features include call type, location, time, weather, repeat caller history, and concurrent incident load. Outputs probability of escalation, likely patient count, and recommended resource staging.
Real-World Impact
Paramedics arrive better prepared. Incident commanders stage resources proactively instead of reactively. Reduces the “surprise factor” that costs lives in dynamic incidents.
Ethical Considerations
  • Predictions based on location risk profiling neighborhoods by socioeconomic status
  • False escalation predictions waste resources; false de-escalation predictions cost lives
  • Prediction confidence must be communicated, not just the prediction itself
Required Guardrails
  • Display confidence intervals, not just point predictions
  • Never reduce response below standard protocol based on a prediction alone
  • Track prediction accuracy by neighborhood to detect geographic bias
  • Incident commanders make staging decisions, AI informs, not directs
From the Field
Use a Traffic Light System for AI recommendations: Green means high confidence, low risk, follow the guidance. Yellow means mixed confidence, moderate risk, verify first. Red means low confidence, high risk, human judgment dominant. Simple enough for any officer to understand in 30 seconds. (AI in Policing, Ch. 4)
Key question for leaders: Can your commanders see the confidence level behind each prediction, or just the prediction? Certainty without context is dangerous.
How It Works
NLP processes radio transcripts and CAD logs to reconstruct incident timelines. Process mining algorithms compare actual response patterns to standard operating procedures. Statistical analysis identifies recurring failure modes across hundreds of incidents.
Real-World Impact
Agencies that do thorough after-action reviews learn faster. Most agencies only review critical incidents. AI enables review of every incident, surfacing patterns invisible in one-off reviews.
Ethical Considerations
  • Individual performance data in reviews can feel punitive
  • Automated review may miss contextual factors that explain deviations from SOP
  • Data from reviews must not be used in litigation without appropriate safeguards
Required Guardrails
  • Aggregate findings for systemic improvement, do not single out individuals without context
  • Human reviewers validate AI-identified SOP deviations before any action
  • Reports are quality improvement documents, protected from discovery where law permits
  • Include frontline input in interpreting AI-generated findings
From the Field
Leadership in AI-enabled policing is really leadership in continuous improvement. Unlike traditional technology purchases, AI requires continuous attention, updating, and refinement. Algorithms must be monitored for bias and accuracy. Training data must be kept current. The real value comes from iteration, not perfection out of the gate. (AI in Policing, Ch. 16)
Key question for leaders: How many after-action reviews did your agency complete last year? If the answer is “only the big ones,” you’re missing the patterns hiding in the routine calls.
How It Works
NLP analyzes caller language, CAD notes, and address history for mental health indicators. Classification models flag calls as probable mental health crisis. Dispatch systems automatically recommend co-response teams, mobile crisis units, or peer counselors based on availability.
Real-World Impact
Right responder for the right call. Reduces unnecessary law enforcement contact with people in crisis. Diverts individuals to treatment instead of jail. Reduces officer and subject injury rates.
Ethical Considerations
  • Labeling addresses or individuals as “mental health” creates stigma and profiling risk
  • Misclassification can send clinicians to dangerous scenes or officers to crisis scenes
  • Data sharing between law enforcement and health systems raises HIPAA/privacy issues
Required Guardrails
  • Address-based flagging reviewed and purged annually
  • Co-response teams dispatched in addition to, not instead of, safety resources when risk is uncertain
  • Health information firewalled, dispatch classification does not enter law enforcement records
  • Community mental health stakeholders involved in model design and oversight
  • Track outcomes: hospitalizations, arrests, injuries by response type
From the Field
During one crisis week at Tempe, two missing kids, a string of sexual assaults, record heat, I averaged 2 hours of sleep on caffeine and adrenaline. My body was screaming that I was running on empty, but I couldn’t see it. I was too close to the problem. That’s the data we weren’t capturing and the reason the right responder matters. Worn-down officers responding to mental health calls is a recipe for tragedy. (AI in Policing, Ch. 12)
Key question for leaders: Are you sending the right responder to the right call, or are you sending a badge and a gun to every 911 call because that’s all you have?

Questions to Ask Your Vendor: Emergency Response & Dispatch

  1. What languages does your system support, and what’s the accuracy rate for non-English speakers?
  2. What’s the fail-safe when your system can’t classify a call? (Only acceptable answer: defaults to highest priority.)
  3. Can dispatchers override every recommendation, and is the override logged for quality improvement?
  4. What’s the latency? In a 911 environment, 2 seconds matters.
  5. How does your system handle simultaneous multi-casualty events?
  6. What training do you provide, for IT staff or for the dispatchers who actually use it?

Courts & Judicial System

AI tools for case management, risk assessment, document processing, and court operations.
How It Works
Models trained on historical court data predict failure-to-appear and pretrial rearrest probabilities. Inputs include charge severity, criminal history, employment stability, community ties, and substance use history. Outputs a risk score on a structured scale.
Real-World Impact
Reduces reliance on cash bail (which penalizes poverty). Provides judges with consistent, evidence-based information. When implemented well, reduces pretrial detention without increasing failures to appear.
Ethical Considerations
  • Criminal history data encodes decades of racially disparate enforcement
  • Proxy variables (zip code, employment) can replicate racial and economic bias
  • Risk scores can be misused as a “black box” justification for detention
  • Due process requires defendants be able to understand and challenge the basis for decisions
Required Guardrails
  • Risk scores are advisory only, judges must articulate independent reasoning for detention decisions
  • Validate tool for racial and ethnic bias before deployment; revalidate annually
  • Defendants and defense counsel must have access to their score, the factors, and the methodology
  • Prohibit using zip code, employment status, or education as direct model inputs
  • Independent validation study required before adoption; results published publicly
From the Field
ProPublica’s 2016 investigation found risk assessment algorithms flagged Black defendants as future criminals nearly twice as often as white defendants. These are not just technical issues, they translate directly into real-world harms that affect people’s lives and liberty. The badge, the gun, and now the algorithm, all require the same careful judgment and accountability. (AI in Policing, Ch. 3)
Key question for leaders: Can a public defender explain your risk tool to their client? If not, you have a due process problem.
How It Works
OCR and NLP pipelines extract entities (names, dates, charges, statutes), classify document types, identify key clauses, and summarize lengthy filings. Legal-domain LLMs provide more accurate extraction for jurisdiction-specific terminology.
Real-World Impact
Court clerks spend 60–70% of their time on document processing. AI reduces this dramatically. Faster case processing reduces backlog and time-to-trial for defendants.
Ethical Considerations
  • Extraction errors in legal documents can have serious due process implications
  • LLM summaries of legal filings may omit or mischaracterize critical arguments
  • Dependency on AI processing can mask staffing needs and erode institutional knowledge
Required Guardrails
  • All AI-extracted data subject to human verification before use in proceedings
  • Error rates tracked and reported; retraining triggered when accuracy drops below threshold
  • AI summaries flagged as summaries, never substituted for the original filing
  • Staff trained on AI limitations, not just AI capabilities
From the Field
My analyst Janet Smith manually entered every auto theft report for an entire year. She knew each case by heart. When we ran her data through ATAC Workstation, the system found a pattern in the first hour she had completely missed, thefts every Tuesday and Thursday, 2–4 PM, specific vehicle types. Human cognition has hard limits at scale. That’s not a criticism of Janet, it’s physics. AI does the scale work; humans provide the judgment. (AI in Policing, Ch. 7)
Key question for leaders: What is your error rate on AI-extracted document data, and who is tracking it? If nobody is, you’re trusting the machine more than you should.
How It Works
Historical case filing data combined with external factors (legislative changes, economic indicators, arrest trends) feed time-series models. Outputs forecast filings by case type, estimated hearing time, and resource allocation recommendations.
Real-World Impact
Reduces hearing delays, optimizes courtroom scheduling, and provides data-driven justification for budget and staffing requests. Some courts report 20–30% improvement in calendar efficiency.
Ethical Considerations
  • Efficiency optimization must not compromise defendants’ right to adequate hearing time
  • Forecasting arrest trends can inadvertently normalize enforcement patterns
Required Guardrails
  • Minimum hearing times per case type never reduced below constitutional standards
  • Forecasts used for resource planning, not to pressure faster case disposition
  • Judges retain full authority over individual case scheduling
From the Field
A rural sheriff’s office I worked with used AI workload analysis to identify and eliminate 4.3 hours of unnecessary admin tasks per deputy per week, then used that data to secure 4 new deputy positions from the county. The sheriff told me: “We’ve always known our deputies were overworked, but now we can prove it with data.” That’s the power of forecasting, it turns hunches into budget justifications. (AI in Policing, Ch. 8)
Key question for leaders: Are you using workload data to get more judges, or to push existing judges to move faster? One is good management; the other is a justice risk.
How It Works
Statistical models analyze sentencing outcomes controlling for legally relevant factors. Disparity dashboards highlight statistically significant differences by judge, demographics, defense counsel type, and geography.
Real-World Impact
Brings transparency to a process that has historically been opaque. Enables chief judges to address unwarranted disparity through training and policy. Provides data for sentencing reform advocacy.
Ethical Considerations
  • Publishing judge-level disparity data can be politically sensitive and affect judicial independence
  • Controlling for “legally relevant factors” itself involves judgment calls about what is relevant
  • Disparity does not always equal discrimination, context matters
Required Guardrails
  • Analysis controlled for legally relevant factors before any disparity finding
  • Results shared with judicial education programs, not media, as the first channel
  • Individual judge data anonymized in public reports; named data for internal improvement only
  • Methodology reviewed by independent statisticians and legal scholars
From the Field
One department I worked with published their predictive methodology publicly, created a citizen-facing dashboard, and established a citizen advisory committee with data access. The result? Enhanced scrutiny actually improved the system, the community identified bias issues that technical staff had missed. Transparency isn’t a liability; it’s your best quality control. (AI in Policing, Ch. 3)
Key question for leaders: Do your judges know their own sentencing patterns compared to their peers? The data often surprises them, and surprise is the first step to change.
How It Works
Automated systems track case milestones and send personalized notifications via preferred channels. NLP chatbots answer common procedural questions. Resource matching connects victims with appropriate advocacy services based on case type.
Real-World Impact
Victim and witness participation rates increase when they are informed and supported. Reduces failure-to-appear for witnesses. Meets VOCA notification requirements more consistently than manual processes.
Ethical Considerations
  • Automated notifications about sensitive matters require trauma-informed design
  • Contact information security is paramount, disclosure could endanger victims
  • Chatbot responses must not be confused with legal advice
Required Guardrails
  • Victim advocate review of all notification templates before deployment
  • Secure channels with identity verification for all victim communications
  • Chatbots clearly disclaim that responses are informational, not legal advice
  • Opt-out mechanism for victims who prefer human contact exclusively
From the Field
Sometimes the most powerful innovations come not from solving complex technical problems, but from asking simple human questions: “What are people really doing all day, and how is it affecting them?” That question drove AutoLog, SWORN, and every system I’ve built. Apply the same thinking to victims and witnesses, what do they actually need from the court system, and where is the system failing them? (AI in Policing, Ch. 8)
Key question for leaders: How many victim notification requirements did your court miss last quarter? If you don’t know the number, you need this system.

Questions to Ask Your Vendor: Courts & Judicial System

  1. Can a public defender access the factors and methodology behind a risk score?
  2. Has your tool been validated for racial and ethnic bias on a population similar to ours?
  3. What’s your error rate on document extraction, and who tracks it?
  4. If a defendant challenges the algorithm in court, can you provide an expert witness?
  5. Do you publish the methodology publicly, or is it proprietary? (Black boxes have no place in courtrooms.)
  6. What happens when legislation changes, how fast can your system adapt?

Corrections & Reentry

Facility management, classification, threat assessment, and programs that reduce recidivism.
How It Works
Models incorporate criminal history, institutional behavior, gang indicators, mental health screening, and vulnerability factors to recommend custody level and housing. Separation constraints (co-defendants, rival groups, protective custody) are enforced algorithmically.
Real-World Impact
Reduces inmate-on-inmate assaults through better separation. Prevents over-classification (housing minimum-risk people in maximum security), which is costly and harmful. Identifies programming needs at intake.
Ethical Considerations
  • Classification tools can embed racial bias from criminal history inputs
  • Gang classification databases are notoriously inaccurate and disproportionately affect minorities
  • Over-classification restricts privileges, program access, and can extend time served
Required Guardrails
  • Validated classification instruments only, no locally developed ad hoc scoring
  • Override tracking: when staff override AI, document the reason
  • Reclassification at defined intervals (90–180 days), not just at intake
  • Gang designations independently verified, not inherited from prior booking data
  • Disparate impact analysis by race and ethnicity annually
From the Field
Officer Kevin Johnson conducted a behavioral experiment at home, had his wife simulate rummaging through dresser drawers like a burglar. He discovered that inexperienced offenders open top-to-bottom (only bottom drawer left open), while experienced ones go bottom-to-top (all drawers open). That insight was confirmed through arrests. The same principle applies to classification: behavior patterns reveal more than criminal history alone. (AI in Policing, Ch. 22)
Key question for leaders: What percentage of your population is over-classified? Every over-classified inmate costs you money and increases your liability.
How It Works
Computer vision analyzes mail imaging and body scanner outputs. NLP monitors approved communication channels for threat indicators and contraband negotiation. Behavioral analytics track movement patterns for anomalies suggesting preparation for violence.
Real-World Impact
Proactive threat detection prevents assaults, drug overdoses, and escapes. Reduces reliance on intrusive physical searches, which are degrading and resource-intensive.
Ethical Considerations
  • Monitoring inmate communications raises attorney-client privilege concerns
  • False positives result in punitive responses for innocent behavior
  • Pervasive surveillance in custody compounds existing power imbalances
Required Guardrails
  • Attorney-client communications categorically excluded from monitoring
  • AI flags reviewed by trained staff before any adverse action
  • False positive rates tracked; system recalibrated when rate exceeds threshold
  • Inmate notification that non-privileged communications are subject to monitoring
From the Field
Behind a grocery store on late-night patrol, I encountered a known offender named Alan. While talking to him, a clear voice entered my mind: “Turn around.” I did, and found his brother sneaking toward me with a 2x4 in a swinging position. I drew my Glock and ended the threat. Was it intuition, training, or something else? Whatever it was, AI doesn’t have it. Humans have extraordinary powers that machines do not. That’s why every AI threat detection system needs a human making the final call. (AI in Policing, Ch. 22)
Key question for leaders: Are attorney-client communications categorically excluded from your monitoring system? If the answer isn’t an absolute yes, you have a constitutional problem.
How It Works
Risk-needs-responsivity (RNR) models assess criminogenic needs. Recommendation algorithms match individuals to available community programs based on risk/need profile, geographic constraints, and program capacity.
Real-World Impact
Structured reentry planning reduces recidivism by 10–20% compared to no planning. Connecting individuals to services before release dramatically improves program engagement rates.
Ethical Considerations
  • Risk scores can create self-fulfilling prophecies if used to deny resources rather than target them
  • Algorithmic matching may not account for individual motivation and readiness for change
  • Over-reliance on risk tools can replace the human relationships that drive desistance
Required Guardrails
  • Risk tools used to increase services for high-risk individuals, never to deny them
  • Individual input and preferences incorporated into reentry plans alongside AI recommendations
  • Case managers maintain primary relationship, AI augments, not replaces, human judgment
  • Outcome tracking: what percentage of recommended programs are completed? Do completers recidivate less?
  • Revalidate risk tools on the facility’s own population every 2–3 years
From the Field
SWORN’s approach to officer wellness applies directly to reentry: personalized baselines, not one-size-fits-all. Officer Garcia responds to physical activity. Officer Williams responds to guided meditation. The system learns which interventions work for which people. Apply the same logic to reentry, match the person to the program, not the program to the policy. (AI in Policing, Ch. 12)
Key question for leaders: Are you using risk scores to target services at the people who need them most, or to justify denying them? The same tool can do either.
How It Works
NLP analyzes intake questionnaires, prior medical records, and behavioral observations to flag health needs. Classification models prioritize referrals by acuity. Predictive models identify individuals at elevated risk for self-harm.
Real-World Impact
Timely identification of medical and mental health needs reduces in-custody deaths, suicide, and chronic disease complications. Proper treatment in custody is both a constitutional obligation and a reentry success factor.
Ethical Considerations
  • Health data in correctional settings requires strict confidentiality protections
  • Suicide risk prediction errors have life-or-death consequences in both directions
  • AI screening must not replace qualified medical and mental health professionals
Required Guardrails
  • AI screening is a triage tool, all positive screens require clinical evaluation
  • Health data compartmentalized from security/classification systems
  • Suicide risk flags trigger immediate human welfare check, not just a database entry
  • HIPAA-compliant data handling even in correctional settings
  • Clinical override authority preserved for all AI recommendations
From the Field
Dr. Amy Arnsten at Yale showed that stress shifts brain resources from the prefrontal cortex to fight-or-flight centers. Dr. Judith Andersen at University of Toronto confirmed this directly impacts officer decision-making. The same neuroscience applies to inmates under facility stress. MIT is now developing smart fabrics with embedded sensors that track vitals through uniform material, no wearables needed. Context-aware systems that know the difference between elevated heart rate during a yard fight and elevated heart rate during a medical emergency. (AI in Policing, Ch. 12)
Key question for leaders: How quickly does a suicide risk flag in your system result in a face-to-face welfare check? If the answer is “next business day,” your system isn’t fast enough.
How It Works
Time-series models track leading indicators: grievance filing rates, disciplinary spikes, commissary pattern changes, unusual movement clusters, and social network disruption. Anomaly detection flags deviations from baseline.
Real-World Impact
Enables preventive interventions (targeted programming, staffing adjustments, preemptive de-escalation) before violence erupts. Proactive approach vs. reactive lockdowns.
Ethical Considerations
  • Collective profiling raises fairness concerns
  • Preemptive lockdowns based on predictions restrict the rights of all inmates
  • Sensitivity of data inputs (e.g., religious or ethnic group tensions)
Required Guardrails
  • Predictions trigger investigation and de-escalation, not preemptive punishment
  • Facility-level predictions, not individual blame
  • Interventions reviewed by multidisciplinary team (security, programming, mental health)
  • Track prediction accuracy; discontinue if false alarm rate undermines trust
From the Field
The IZE Method, Categorize, Generalize, Organize, Minimize, Maximize, was built to overcome linkage blindness, when related incidents remain unconnected. In 2003, seventeen seemingly random residential burglaries turned out to be one offender targeting homes with dog doors. Nobody connected them until the IZE framework was applied. The same methodology works for facility incident prediction: you’re looking for the pattern hiding in the noise. (AI in Policing, Ch. 6)
Key question for leaders: When your grievance rate spikes, do you see it as a warning signal or an annoyance? This tool works only if leadership treats early signals as valuable intelligence.

Questions to Ask Your Vendor: Corrections & Reentry

  1. Are attorney-client communications categorically excluded from your monitoring system?
  2. How often does your classification tool get revalidated on our facility’s population?
  3. What is your false positive rate on threat detection, and what happens to inmates falsely flagged?
  4. Can your system integrate with community service providers for reentry planning?
  5. Does your risk tool increase services for high-risk individuals or deny them? Show me the data.
  6. How do you handle HIPAA compliance for health screening data in a correctional setting?

Homeland Security & Intelligence

Threat detection, infrastructure protection, intelligence fusion, and counter-terrorism applications.
How It Works
NLP processes unstructured intelligence reports across agencies. Entity resolution links references to the same people, organizations, and locations across databases. Knowledge graphs map relationships and detect patterns spanning multiple intelligence streams.
Real-World Impact
Breaks down information silos that contributed to pre-9/11 intelligence failures. Reduces time from collection to actionable intelligence. Enables threat detection at the intersection of datasets no single analyst could hold in their head.
Ethical Considerations
  • Mass data aggregation creates comprehensive profiles of individuals, including innocent people
  • Intelligence products can justify surveillance of constitutionally protected activity
  • Cross-agency data sharing multiplies the impact of errors in any single source
  • Minority communities disproportionately represented in intelligence databases
Required Guardrails
  • Strict need-to-know access controls on fused intelligence products
  • Predication requirements: documented reasonable suspicion before querying individuals
  • First Amendment-protected activity excluded from pattern analysis
  • Inspector General oversight with access to query logs and audit trails
  • Data provenance tracking, every fact traces back to its source and reliability rating
From the Field
ATAC Workstation’s first public demonstration linked four commercial burglaries that human detectives had filed as unrelated. The system even predicted a fifth target location, where the suspect was arrested. A thirty-year detective who had dared me to teach him something new never ran another briefing without the software. That’s what intelligence fusion does: it sees the connections hiding in plain sight across siloed databases. (AI in Policing, Ch. 7)
Key question for leaders: Can you trace every fact in an intelligence product back to its original source and reliability rating? If not, you’re building analysis on unverified foundations.
How It Works
Baseline models of normal infrastructure behavior detect deviations in SCADA/ICS telemetry, network traffic, and sensor readings. Multi-modal fusion correlates anomalies across systems, a simultaneous power fluctuation and network intrusion is more concerning than either alone.
Real-World Impact
Detects sophisticated attacks that evade single-sensor monitoring. Reduces mean time to detect from weeks to minutes for advanced persistent threats targeting infrastructure.
Ethical Considerations
  • Infrastructure monitoring data can reveal patterns of private behavior
  • False positives can trigger costly emergency responses and public alarm
  • Centralized monitoring creates a single point of failure and an attractive attack target
Required Guardrails
  • Aggregated anomaly detection only, no individual behavior tracking from infrastructure data
  • Multi-factor confirmation required before emergency response escalation
  • Redundant monitoring, AI supplements, does not replace, human operators
  • Regular red-team exercises to test detection capabilities and false positive rates
From the Field
Quantum computing will transform how law enforcement approaches complex analytical challenges. Qubits exist in multiple states simultaneously, enabling pattern recognition across seemingly unrelated activities, evaluating millions of relationships simultaneously. But it can also break sophisticated encryption in minutes. The infrastructure protection game is about to change fundamentally. (AI in Policing, Ch. 25)
Key question for leaders: Is your AI monitoring system itself hardened as critical infrastructure? The system protecting your grid is only as secure as its own defenses.
How It Works
Models analyze advance passenger information, travel itineraries, historical crossing patterns, and watch lists. Cargo screening models assess manifests, shipper history, and routing anomalies. Officers receive risk scores with explanatory factors to guide secondary screening.
Real-World Impact
Focuses limited inspection resources where most needed. Reduces wait times for low-risk travelers. Improves interdiction rates for contraband and security threats.
Ethical Considerations
  • Profiling by nationality, ethnicity, or religion in screening algorithms violates civil rights
  • Opaque risk scores deny travelers the ability to challenge screening decisions
  • Automation bias, officers may defer to AI instead of exercising trained judgment
  • Refugee and asylum seekers may be disadvantaged by travel pattern models
Required Guardrails
  • Protected characteristics (race, religion, national origin) prohibited as model inputs
  • Travelers flagged for secondary screening told the general category of concern
  • Officer discretion to override AI in both directions
  • Independent civil liberties review before deployment and annually
  • Asylum and refugee processing pathways exempt from standard risk scoring
From the Field
The MIT Gender Shades study found facial recognition error rates of 34.7% for darker-skinned women vs. 0.8% for lighter-skinned men. When you deploy screening technology at scale, those error rates translate directly into real-world harms. The question isn’t whether your data has bias, it does. The question is whether you’re measuring and mitigating it. (AI in Policing, Ch. 3)
Key question for leaders: If a civil rights organization filed a FOIA on your screening algorithm, what would they find? Answer that question before they ask it.
How It Works
Distributed sensor networks monitor for anomalous chemical, biological, and radiological signatures. Detection triggers atmospheric dispersion models that predict exposure zones based on weather, terrain, and release characteristics.
Real-World Impact
Minutes matter in CBRN events. AI-accelerated detection and modeling gives emergency managers actionable guidance before human analysts could complete the same calculations.
Ethical Considerations
  • False alarms can cause panic, economic disruption, and erode public trust
  • Dispersion model errors can lead to evacuation of wrong areas
  • Sensor network gaps may leave disadvantaged communities unmonitored
Required Guardrails
  • Multi-sensor confirmation required before public alerting
  • Model uncertainty communicated, dispersion predictions include confidence zones
  • Sensor coverage equity analysis: no community underserved by detection network
  • Regular calibration and testing of sensor networks and dispersion models
From the Field
When you activate emergency AI tools, surveillance, social media monitoring, facial recognition, remember the PATRIOT Act. FISA was passed in 1978 with careful safeguards. After 9/11, temporary emergency powers expanded dramatically and never contracted. Establish pre-event ethical boundaries. Strict sunset provisions. Independent oversight. Mandatory data purging. Temporary powers have a way of becoming permanent. (AI in Policing, Ch. 4)
Key question for leaders: Where are the gaps in your sensor coverage, and which communities are in them? If you don’t know, you haven’t looked.
How It Works
NLP models analyze public-facing content for radicalization indicators. Behavioral models identify patterns associated with mobilization (operational security changes, network isolation, target research). Multi-factor models combine online and offline indicators.
Real-World Impact
Shifts from reactive (responding to attacks) to proactive (identifying pre-attack indicators). Enables diversion to countering violent extremism programs before violence occurs.
Ethical Considerations
  • Monitoring speech and ideology is a direct tension with the First Amendment
  • Models disproportionately target specific communities depending on design
  • The line between protected speech and actionable threat is legally complex
  • Intervention programs can stigmatize communities and individuals
  • Predictive models for rare events produce extremely high false positive rates
Required Guardrails
  • Focus on behavioral indicators of violence mobilization, not ideological beliefs
  • Constitutional review by legal counsel before any monitoring program
  • Community-led intervention programs, not solely law enforcement
  • Strict prohibition on targeting based on religion, political ideology, or ethnicity
  • Sunset provisions: programs reviewed and reauthorized at defined intervals
From the Field
Establish ethical frameworks in peacetime, not during a crisis. When time is limited and stakes involve immediate public safety, how do we ensure ethical principles aren’t sacrificed for operational expediency? The answer: you don’t figure it out in the moment. You build the guardrails before you need them, with redundant communication systems, clear override channels, and pre-established rules of engagement that AI cannot modify. (AI in Policing, Ch. 4)
Key question for leaders: Are you targeting behaviors or beliefs? If you can’t clearly articulate the difference, you’re not ready to deploy this tool.

Questions to Ask Your Vendor: Homeland Security & Intelligence

  1. How do you ensure First Amendment-protected activity is excluded from pattern analysis?
  2. What predication requirements exist before an analyst can query an individual?
  3. Can every fact in an intelligence product trace back to its source and reliability rating?
  4. What are the sunset provisions? Under what conditions does data get purged?
  5. Has your system undergone an independent civil liberties review?
  6. How do you prevent mission creep? The PATRIOT Act taught us that temporary powers become permanent.

Community Safety & Prevention

Upstream prevention, early intervention, community engagement, and violence reduction programs.
How It Works
Social network analysis maps co-offending networks and identifies individuals at elevated violence risk positions. Models combine network position with individual risk factors (recent incarceration release, prior victimization). Outreach workers, not police, use the information to engage individuals.
Real-World Impact
Programs like Chicago CRED and Richmond ONS demonstrate 30–60% reductions in gun violence in target areas. AI sharpens targeting so limited outreach resources reach the right people.
Ethical Considerations
  • Network-based targeting can profile people based on associations, not actions
  • Data sharing between law enforcement and community organizations can compromise trust
  • Individuals labeled “high risk” may face collateral consequences
  • Community organizations must be primary actors, not extensions of law enforcement
Required Guardrails
  • Community organizations control outreach decisions, law enforcement provides data, not direction
  • Strict data use agreements: information shared for outreach cannot be used for prosecution
  • Individuals must be able to decline engagement without consequences
  • Program evaluations measure violence reduction and community trust, not arrests
  • Community advisory board with direct influence over program design and data use
From the Field
Social network analysis used to be restricted to federal intelligence agencies. Now it’s available as affordable SaaS applications. But the technology isn’t the hard part, the governance is. One department I worked with published their methodology publicly, created a citizen-facing dashboard, and let a citizen advisory committee see the data. Enhanced scrutiny improved the system. The community identified bias issues the technical staff had missed. (AI in Policing, Ch. 3 & 11)
Key question for leaders: Who controls the outreach decisions, community organizations or law enforcement? The answer determines whether your community will trust this program.
How It Works
Integrated data systems combine records across housing, health, education, and social services (with consent and legal authority). Predictive models identify early warning indicators of crisis. Service navigators receive referral recommendations.
Real-World Impact
Shifts from crisis-driven to prevention-driven social services. Reduces emergency department visits, foster care placements, and shelter stays. Every dollar spent on prevention saves $7–10 in crisis response.
Ethical Considerations
  • Integrated data systems create comprehensive profiles that can be used for surveillance
  • Risk labels can lead to intrusive government attention on vulnerable families
  • Marginalized communities have justified distrust of government data systems
  • Consent and data sharing authority must be clear and respected
Required Guardrails
  • Opt-in or informed consent wherever legally feasible
  • Data used exclusively for service connection, never for enforcement or investigation
  • Community members represented in data governance structures
  • Right to access: individuals can see what data is held about them
  • Strict purpose limitation enforced technically and legally
From the Field
Data, data, data. Here’s the brutal truth: data will make or break AI. Period. The systems that get access to the best, most comprehensive real-world data will dominate. The ones that don’t will become irrelevant footnotes. But “best” doesn’t mean “most.” It means most representative, most consented, and most governed. Social service data without governance is surveillance. (AI in Policing, Ch. 8)
Key question for leaders: Would the families you’re trying to help be comfortable knowing what data you hold about them? If not, you have a consent problem.
How It Works
Spatial models correlate crime patterns with environmental features: lighting, sight lines, land use, vacancy rates, infrastructure condition. Computer vision assesses conditions from street-level imagery at scale. Outputs prioritized intervention lists with predicted impact.
Real-World Impact
Environmental interventions are among the most cost-effective crime reduction strategies. Greening vacant lots in Philadelphia reduced gun violence by 29% in surrounding blocks. AI enables citywide assessment.
Ethical Considerations
  • Neighborhood-level analysis can stigmatize communities
  • Infrastructure investment decisions should be community-driven
  • Gentrification risk: environmental improvements can drive displacement
Required Guardrails
  • Community input required before implementing interventions
  • Anti-displacement protections paired with environmental improvements
  • Data shared openly with community for collaborative planning
  • Interventions evaluated on community wellbeing metrics, not just crime statistics
From the Field
In the early days, “high tech” meant acetate pin maps and a 486 desktop that took 10 minutes to boot. We tracked crime data with colored pins, static, unable to answer new questions without rebuilding from scratch. Kim Rossmo was doing geographic profiling on mainframes. Now computer vision can assess environmental conditions from street-level imagery across an entire city in hours. The evolution is staggering, but the principle hasn’t changed: space shapes behavior. (AI in Policing, Ch. 1 & 6)
Key question for leaders: Are you investing in the environment to reduce crime, or to raise property values? Without anti-displacement protections, you may be doing both, and that’s not always a win for the people who live there.
How It Works
Risk screening at school and community contact points identifies youth who may benefit from intervention. Recommendation algorithms match youth to programs based on risk factors, age, interests, and program effectiveness data. Case managers coordinate referrals.
Real-World Impact
Youth who are diverted are 50% less likely to reoffend than youth who are formally processed. Keeping youth out of the justice system is the single highest-return investment in public safety.
Ethical Considerations
  • Labeling children as “at risk” can stigmatize and create self-fulfilling prophecies
  • School-based data collection raises FERPA and privacy concerns
  • Diversion systems can widen the net of social control over youth
  • Racial disparities in school discipline mean inputs are already biased
Required Guardrails
  • Family consent required for any youth screening or data sharing
  • Strengths-based language, “opportunities for support” not “risk factors”
  • Diversion resources must actually exist, don’t identify needs you can’t meet
  • Demographic parity tracking: are opportunities distributed equitably by race?
  • Data destroyed when youth ages out of eligibility, no permanent record
From the Field
AI in community policing starts with real-time community intelligence. From paper surveys to digital understanding, AI tools for sentiment analysis and multi-language translation break down barriers between agencies and the communities they serve. But building trust through authentic engagement requires more than technology, it requires officers who know their communities. AI can bridge languages; only people can bridge trust. (AI in Policing, Ch. 15)
Key question for leaders: Are diversion opportunities distributed equitably by race, or are some kids getting second chances while others get processed? Check the numbers.
How It Works
Real-time feeds from emergency departments, EMS, poison control, and wastewater analysis detect anomalous patterns. Epidemiological models classify emerging clusters and predict spread. Automated alerting to public health officials and first responders.
Real-World Impact
Rapid detection of fentanyl-laced drug supply changes saves lives through targeted naloxone distribution. Early disease outbreak identification enables containment before exponential spread.
Ethical Considerations
  • Health surveillance data can be used for enforcement (criminalizing substance use)
  • Community-level data publication can stigmatize neighborhoods
  • Individual health data requires robust privacy protections
Required Guardrails
  • Public health data used for health response only, firewalled from law enforcement
  • Individual data de-identified before analysis; re-identification prohibited
  • Community notification designed to reduce harm, not stigmatize
  • Data governance board including public health, community advocates, and privacy experts
From the Field
Advanced wearable pilot data showed officers on night shifts had significant sleep disruption even on their days off. Six months into my career, I transitioned to nights, zombie state within three weeks, nearly rear-ended a stopped car on patrol. The same principle applies to community health surveillance: the data you need most is the data nobody is collecting, and the patterns most dangerous are the ones that look normal until they don’t. (AI in Policing, Ch. 12)
Key question for leaders: Is your public health data firewalled from law enforcement? If not, the people most at risk won’t seek help, because help looks like a trap.

Questions to Ask Your Vendor: Community Safety & Prevention

  1. Who controls outreach decisions, community organizations or law enforcement?
  2. Are there strict data use agreements preventing information shared for outreach from being used for prosecution?
  3. Can individuals decline engagement without any consequences?
  4. How do you measure success, arrests, or violence reduction and community trust?
  5. What anti-displacement protections are built into environmental improvement recommendations?
  6. Is family consent required before any youth screening or data sharing?

Digital Forensics & Cybercrime

Digital evidence processing, network forensics, dark web intelligence, and cyber threat response.
How It Works
AI classifies files by type and relevance. Image classifiers identify contraband content. NLP extracts key communications from messaging apps, email, and social media. Priority scoring surfaces the most relevant evidence for examiner review first.
Real-World Impact
Digital forensic labs face 6–18 month backlogs. AI triage reduces examiner time per device from days to hours. Gets evidence to prosecutors faster, reducing pretrial detention periods.
Ethical Considerations
  • Automated scanning of personal devices raises Fourth Amendment concerns
  • AI classification errors can flag innocent content or miss relevant evidence
  • Examiner exposure to disturbing content is a significant wellness concern AI can mitigate
Required Guardrails
  • AI triage operates within scope of search warrant, no evidence outside warrant scope
  • All AI-identified evidence verified by certified forensic examiner before use
  • Chain of custody documentation includes AI processing steps
  • Regular accuracy testing against known test datasets
  • AI-assisted review reduces examiner exposure to harmful content, a genuine welfare benefit
From the Field
Janet Smith knew every auto theft case by heart after manually entering a year’s worth of reports. ATAC found the pattern in the first hour. Human cognition struggles with large datasets despite expertise, that’s not a criticism, it’s physics. The same principle applies to digital forensics: a 6-month backlog isn’t a staffing problem you can hire your way out of. It’s a scale problem that requires AI. (AI in Policing, Ch. 7)
Key question for leaders: How long is your forensic backlog, and how many cases are delayed because of it? AI triage doesn’t replace examiners, it multiplies their capacity.
How It Works
Baseline models of normal network behavior detect anomalies in traffic volume, port usage, authentication patterns, and data flows. Behavioral analytics identify patterns consistent with MITRE ATT&CK techniques. Correlation engines connect anomalies across segments.
Real-World Impact
Government networks are targeted constantly. AI detection reduces mean time to detect from 207 days (industry average) to hours or minutes.
Ethical Considerations
  • Network monitoring can capture employee personal communications on government systems
  • False positives can disrupt legitimate operations
  • AI-detected threats may be escalated with reduced oversight
Required Guardrails
  • Clear acceptable use policies for government networks inform monitoring scope
  • False positive rate tracked; alert fatigue mitigation implemented
  • Escalation protocols with defined thresholds for each response level
  • Regular penetration testing to validate detection capabilities
From the Field
It’s like watching someone buy a Ferrari without knowing how to drive: impressive at first glance, but ultimately headed for disaster. That’s what happens when agencies deploy sophisticated AI monitoring without the governance to match. Intrusion detection is only as good as the response protocols behind it. (AI in Policing, Ch. 9)
Key question for leaders: What is your mean time to detect a network intrusion? If you don’t know the number, you don’t have the monitoring you need.
How It Works
Blockchain analysis clusters wallet addresses by entity using transaction patterns. Graph algorithms map fund flows through mixers, exchanges, and DeFi protocols. Attribution models link addresses to known entities. Visualization tools present complex flows for investigators and juries.
Real-World Impact
Cryptocurrency is no longer anonymous when sophisticated analysis is applied. Major ransomware payments, darknet operations, and money laundering schemes have been disrupted through blockchain analysis.
Ethical Considerations
  • Financial surveillance of cryptocurrency users not suspected of crimes
  • Attribution errors can lead to investigation of innocent wallet holders
  • Privacy coins and mixing create an arms race
Required Guardrails
  • Investigation predicated on specific criminal activity, not general financial surveillance
  • Attribution confidence documented, provisional links distinguished from confirmed
  • Legal process required before exchange-level de-anonymization
  • Expert witness qualification standards for blockchain analysis testimony
From the Field
Blockchain creates an immutable, transparent, distributed verification system. Every digital file, body camera footage, forensic reports, can be secured through blockchain with every access, transfer, and modification permanently recorded in a tamper-proof digital ledger. The same technology criminals use for anonymity can become an unbreakable chain of custody. (AI in Policing, Ch. 25)
Key question for leaders: Do your investigators understand the difference between a confirmed wallet attribution and a probabilistic one? The distinction matters in court.
How It Works
Perceptual hashing (PhotoDNA and similar) identifies known CSAM regardless of format changes. AI classifiers detect previously unknown material. Victim identification models cross-reference environmental details across images to identify filming locations.
Real-World Impact
The volume of CSAM online is beyond human review capacity. AI enables detection at platform scale. NCMEC processes millions of reports annually. Victim identification has directly led to child rescues.
Ethical Considerations
  • Investigator wellness: repeated exposure causes documented psychological harm
  • False positives can subject innocent users to investigation for the most stigmatizing crime
  • Technology used for detection can also inform evasion techniques
Required Guardrails
  • AI reduces human exposure to harmful content, this is a primary design goal
  • Two-stage review: AI flags, trained human confirms before any legal action
  • High-confidence threshold required before platform-level reporting to NCMEC
  • Investigator wellness programs mandatory for units using these tools
  • Detection methodology protected as law enforcement sensitive
From the Field
The measure of success isn’t technical complexity but human benefit. If we can make people’s lives easier, more productive, and less stressful, that’s technology worth building. For CSAM investigators, AI isn’t a nice-to-have, it’s a wellness imperative. Every image an AI system triages is one a human investigator doesn’t have to see. (AI in Policing, Ch. 8)
Key question for leaders: What is your investigator wellness program for the people reviewing this material? If you don’t have one, you’re burning out your best people.
How It Works
Automated systems monitor accessible dark web platforms. NLP classifies content by threat type. Entity resolution links pseudonymous accounts across platforms. Alert systems notify investigators of emerging threats, breaches, or specific threat indicators.
Real-World Impact
Proactive identification of stolen credentials enables notification of affected residents. Detection of weapons and drug sales enables interdiction. Monitoring threat actor communications provides early warning.
Ethical Considerations
  • Undercover operations raise entrapment and authorization issues
  • Collection may capture constitutionally protected anonymous speech
  • Intelligence from dark web sources is difficult to verify and may be fabricated
Required Guardrails
  • Collection limited to publicly accessible content unless authorized by legal process
  • Automated collection must be supervised, not autonomous investigation
  • Source reliability assessment for all intelligence before action
  • Legal review of collection methods before deployment
  • Intelligence products clearly marked with confidence and reliability ratings
From the Field
No-code AI agent development platforms now let domain experts, detectives and analysts, not programmers, build autonomous workflows. Detective Sergeant Ramirez built a “Serial Case Detective” agent that continuously scans incoming reports across jurisdictions for behavioral signatures. He told me: “I thought they were trying to replace detectives with robots. But this is completely different, it’s like having a tireless partner who handles the mundane database searches while I focus on the aspects that require human judgment.” (AI in Policing, Ch. 11)
Key question for leaders: Can you distinguish between credible dark web threats and noise? Without confidence ratings on your intelligence, you’re chasing shadows.

Questions to Ask Your Vendor: Digital Forensics & Cybercrime

  1. Does your triage system operate strictly within the scope of the search warrant?
  2. What is your classification accuracy rate, and how often is it tested against known datasets?
  3. Does your chain of custody documentation include AI processing steps?
  4. What investigator wellness support do you provide for units reviewing disturbing content?
  5. What’s the confidence rating on your blockchain attribution, and can it withstand Daubert scrutiny?
  6. How do you handle evidence from dark web sources that may be fabricated or planted?

Ethical Principles for AI in Public Safety

Six core principles. Each includes a practical test, if you can’t pass the test, you’re not ready to deploy.

Transparency

If we can’t explain to a community member how these systems work, we shouldn’t be using them to make decisions that affect their lives. Transparency must be woven into every aspect of AI implementation, not as documentation, but as plain-language explanations accessible to officers and the public alike.
Practical test: Can your agency answer a FOIA request about this system? Would a journalist understand the response?

Accountability

AI makes recommendations. Humans make decisions. When something goes wrong, a human is responsible, and that must be defined in writing before deployment. The badge, the gun, and now the algorithm all require the same careful judgment and accountability.
Practical test: Is there a named person accountable for this system’s outcomes, not a department, a person?

Fairness & Non-Discrimination

AI systems trained on biased data will produce biased results. The question isn’t whether your data has bias, it does. The question is whether you’re measuring and mitigating it. Apply the Four-Step Bias Mitigation Framework: (1) Data Audits & Cleansing, (2) Algorithmic Transparency & Explainability, (3) Continuous Validation & Testing, (4) Human Oversight & Intervention.
Practical test: Have you run a disparate impact analysis on this system’s outputs, broken down by race and geography?

Privacy & Data Minimization

Collect only what you need. Retain only what you must. The most secure data is data you never collected. Remember the PATRIOT Act: FISA was passed in 1978 with careful safeguards. After 9/11, temporary emergency powers expanded dramatically and never contracted. Sunset provisions aren’t optional, they’re essential.
Practical test: Could you justify every data element you collect if challenged in court? Do you have sunset provisions?

Human Oversight & Control

Technology amplifies judgment; it never replaces it. The best AI systems need human context, just like the finest service weapon is useless without training and discipline. Use the Traffic Light System: Green (high confidence, follow guidance), Yellow (mixed confidence, verify first), Red (low confidence, human judgment dominant). Assign a “Challenge Round”, an officer whose job is to question the AI’s recommendations.
Practical test: Can a frontline operator override this system, and do they know how? Is there a designated challenger?

Proportionality

The invasiveness of the tool must be proportional to the severity of the problem it addresses. Deploying mass surveillance to prevent shoplifting fails this test. Every AI tool is just another tool in our belt, powerful when used right, dangerous when misunderstood.
Practical test: Would a reasonable citizen accept this trade-off between safety and liberty? Would you accept it for your own family?

The SPACE Model: Implementation Framework

A five-phase framework for deploying AI in public safety, from AI in Policing (BYU Academic Publishing, 2025).
S

Strategic Foundation

Define clear objectives before evaluating AI solutions, not the other way around. Identify what chiefs, officers, and the community expect. Align technology with departmental goals. Conduct a data reality check: what do you have, what’s missing, what’s biased? Establish baseline metrics before deployment. A few hours of alignment can save months of friction.
“Implementations fail most often when departments jump into tech acquisition before getting real clarity on what problems they’re solving.”
P

Planned Architecture

Design for the people who will use it, not the IT department that will install it. Involve officers and analysts in workflow decisions. Conduct a privacy impact assessment and civil liberties review. Engage the community before procurement, not after deployment. Require vendor transparency: no black-box algorithms for public safety decisions.
“If it doesn’t work for the person on patrol or behind the desk, it’s dead on arrival.”
A

Active Deployment

Pilot in a limited scope with defined success criteria before full deployment. Shadow systems and peer-led training often make a bigger difference than top-down mandates. Train every user on what the AI does, what it doesn’t do, and how to override it. Create psychological safety for learning, dialogue, and failure.
“Deployments that skip testing and overestimate readiness end up with angry users and lost momentum.”
C

Continuous Integration

Establish ongoing monitoring: accuracy, bias, false positive/negative rates. Refine based on usage data, not gut feelings. Officers and analysts will tell you what works and what doesn’t, if you listen. Celebrate early wins to build momentum. The real value comes from iteration, not perfection out of the gate.
“Unlike traditional technology purchases, AI requires continuous attention, updating, and refinement.”
E

Evolutionary Growth

Create a sunset provision: under what conditions would you shut this system down? Publish an annual transparency report on performance and impact. Promote ongoing training and strategic capability expansion. If the system can’t grow with the department, you’re back at square one in a few years.
“Technology itself doesn’t create lasting change; people do. By thoughtfully integrating technology with organizational dynamics, agencies can achieve meaningful transformation.”

Should We Use AI for This?

A decision framework for agency leaders. If you can’t get past all six questions, you’re not ready.

Q1Is there a clear, specific problem this AI would solve?

Yes
Continue to the next question.
No
Stop. You’re looking for a problem to match a solution. That’s backwards. Technology without strategy is just expensive junk collecting dust.

Q2Does the problem involve decisions that affect people’s liberty, safety, or rights?

Yes
Highest-bar deployment. All guardrails required. Apply the Four-Step Bias Mitigation Framework. Continue.
No
Standard deployment protocols apply. Continue.

Q3Do you have sufficient, representative data to train or validate the AI?

Yes
Continue to the next question.
No
Stop. AI without good data is expensive guessing. Garbage in, garbage out, just with more complex mathematical formulas attached.

Q4Can you explain the AI’s outputs to the people it affects?

Yes
Continue to the next question.
No
Stop. If we can’t explain to a community member how these systems work, we shouldn’t be using them to make decisions that affect their lives.

Q5Do you have the staffing and budget for ongoing monitoring, not just deployment?

Yes
Continue to the next question.
No
Stop. It’s like buying a Ferrari without knowing how to drive. AI requires continuous attention, updating, and refinement, not a one-time purchase.

Q6Has the community been engaged in the decision to deploy this tool?

Yes
Proceed with deployment following the SPACE Model above.
No
Do this first. Trust is a prerequisite, not an afterthought. One department published their methodology publicly and let a citizen advisory committee see the data. Enhanced scrutiny improved the system.

No use cases match your search. Try a different keyword.

Copied!
More from Sean
AI Policy Library AI Prompt Cookbook AI Model Cost Matrix Artistic Style Library First Responder Wellness Build Your Own JARVIS Book Resources All Tools Books