Sergeant Diana Okafor has been reviewing AI-assisted call-triage output for three months when she notices the pattern. Calls originating from two zip codes on the city's east side are consistently flagged by the system as lower priority than calls from other parts of the city with similar stated circumstances. A disturbance call from a business on Lakewood Avenue gets a Priority 2 flag. An almost identically described disturbance call from a commercial block in the northeast quadrant gets a Priority 1. The telecommunicators have been accepting the AI recommendations at high rates because the system was introduced as a time-saving tool and because disagreeing with it feels like extra work on already-full shifts. Okafor pulls the data. Over ninety days, the east-side discrepancy holds. She does not know whether the disparity originates in the model's training data, in the call history the agency fed the system during setup, or in some interaction between the two. But she knows the disparity is there. And she knows it has been shaping how fast units have been responding to her neighbors.

Where Bias Enters a Public Safety AI System

Bias in AI systems is not a single thing. It is not a bug that can be patched in one update, and it is not a result of bad intentions on the part of the people who built the system. It is a structural feature of how machine learning models work: they learn patterns from data, and if the data encodes disparities, the model encodes and in some cases amplifies those disparities. Understanding where bias enters is the prerequisite for detecting and responding to it.

Training Data Bias

Training data is the raw material from which a model learns its patterns. For a police report drafting tool like Axon Draft One, the training data includes a large corpus of police reports. Those reports were written by human officers under a wide range of conditions, with a wide range of writing conventions, and in a wide range of operational contexts. They also encode historical policing patterns.

If the historical data reflects that certain neighborhoods were policed more intensively, that certain subject demographics appeared more frequently in use-of-force reports, or that certain call types were more often coded as high-priority in certain areas, the model learns those patterns. It does not learn them as deliberate choices. It learns them as statistical regularities. When it encounters a new report or a new call, it draws on those regularities to complete its output. The result can be a model that, for example, uses more aggressive descriptive language in its report drafts when the incident involves subject demographics that appeared more frequently in high-intensity incident reports in the training data.

This is not a hypothetical concern. Research on predictive policing systems, risk assessment tools, and natural language processing (NLP) systems used in criminal justice settings has consistently found disparate patterns along racial, socioeconomic, and geographic lines. The specific mechanisms vary by system, but the structural root cause is consistent: historical data that reflects historical disparities, fed into a model that learns from data.

Feedback Loop Bias

A second entry point for bias is the feedback loop. AI systems used in public safety often improve through use: the more the system is used, the more data it has to learn from, and the more the system's outputs are accepted or corrected by operators, the more those acceptances and corrections shape future outputs. This is a useful property when the feedback is high quality. It is a dangerous property when the feedback reinforces an existing bias.

If a call-triage AI consistently flags east-side calls lower and telecommunicators accept those recommendations (because they trust the system, because disagreeing feels burdensome, or because they have internalized the priority pattern as normal), the system receives signal that those recommendations were correct. It learns to reproduce them more confidently. The bias strengthens over time, not because anyone intended it to, but because the feedback loop closed in the wrong direction.

This mechanism is sometimes called automation bias in combination with a feedback loop: the human accepts the machine's recommendation, the machine learns the recommendation was right, and the bias compounds. The solution is not to eliminate AI assistance. The solution is to build override tracking, outlier monitoring, and regular disparity audits into the operational workflow, which this lesson will address in its practical section below.

Reporting and Documentation Bias

A third, more specific entry point is relevant to AI report-writing tools. The model that drafts police narratives has learned from existing police reports. Existing police reports vary in how they describe similar situations depending on the demographics of the subjects involved. Research in criminal justice data has found that language used to describe subject behavior, demeanor, and actions differs systematically by subject race in police documentation, even when controlling for incident type. If a model trained on that corpus learns to produce descriptions that systematically differ by perceived subject demographics, it reproduces and potentially amplifies that variation in every AI-assisted report it drafts.

This is a direct evidentiary concern, not only an equity concern. A report that overcharacterizes the threat posed by a subject from one demographic group and undercharacterizes the same threat from another, when the underlying footage is identical or comparable, is not an accurate report. Under Brady v. Maryland (1963), the prosecution has an obligation to disclose material evidence. An AI-drafted report that systematically mischaracterizes subject behavior along demographic lines is not just a fairness problem: it is a documentation accuracy problem that can affect charging decisions, plea negotiations, and trial outcomes.

The Specific Systems and What the Research Shows

The field of AI bias research in criminal justice has examined several categories of systems. This lesson covers the three categories most directly relevant to patrol officers, dispatchers, and records staff: predictive systems, risk scoring tools, and generative report drafting tools.

Predictive and Classification Systems

Predictive policing systems (which use historical crime data to direct patrol resources) and call-triage classification systems (which use incoming call data to suggest priority levels) have been the subject of substantial research. The consistent finding is that systems trained on historical arrest and crime data tend to direct resources to neighborhoods that have historically been heavily policed, which generates more arrests and reports in those neighborhoods, which feeds back into the model as confirmation that high deployment is warranted. The Electronic Frontier Foundation (EFF) has documented this cycle in its analyses of predictive policing tools, arguing that the systems do not predict crime so much as they predict where police historically went, which disproportionately affects lower-income and minority communities.

This does not mean these systems have no legitimate utility. It means the utility must be evaluated against independently derived ground truth, not against the historical deployment patterns the model was trained on. An agency using a call-triage AI should be able to answer: "What is our disparity rate by zip code or neighborhood? How does AI-assisted triage compare to human-only triage on disparity measures? Has the disparity increased or decreased since we deployed the system?" If those questions cannot be answered because the data is not tracked, that is itself a governance failure.

Risk Assessment Tools

Risk assessment tools used in pretrial detention decisions, parole recommendations, and similar criminal justice contexts have received extensive public scrutiny since the ProPublica investigation into the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) system in 2016. That analysis found that COMPAS assigned higher risk scores to Black defendants at almost twice the rate it incorrectly assigned lower risk scores, even controlling for reoffense rates. The COMPAS developers disputed the methodology, and the technical debate about how to measure fairness in risk assessment is genuinely complex: different mathematical definitions of fairness turn out to be mutually incompatible in certain conditions.

For officers and agencies, the practical takeaway is not that risk assessment tools are uniformly biased or uniformly reliable. It is that these tools should never be used as the sole or determinative factor in a consequential decision, that the tool's methodology should be subject to external audit, and that systematic disparities in outcomes should trigger review of the tool's configuration and training data, not just acceptance of its outputs.

Generative Report Drafting Tools

Generative AI report drafting tools are the newest category and the one with the least independent research as of 2026. What is known from NLP research more broadly is that large language models trained on human-generated text reproduce the linguistic patterns of that text, including patterns of differential description. Models trained on corpora that describe certain demographic groups differently, in terms of threat level, aggression, or agency, tend to reproduce those differential descriptions in their outputs.

For agencies using tools like Axon Draft One, the key question is whether the model has been evaluated for differential output patterns, and if so, what that evaluation found. Vendors have an interest in demonstrating their tools are accurate and fair, but the definition of "fair" and the methods used to measure it are not standardized across the industry. Agencies should ask: "Has this system been tested for differential language outputs across subject demographics? What was the methodology? What were the findings? What corrective measures were taken?" Those questions should be in any procurement contract, and the answers should be in the agency's AI use policy.

The approximately 45 million dollar, multi-year, sole-vendor contracts that some agencies have signed for bundled body-camera, cloud, and AI packages make these questions especially urgent. When an agency has committed to a vendor for ten years and those commitments include the AI tools used for report drafting, triage, and evidence management, the agency has very limited leverage to demand tool changes or data access after the contract is signed. The procurement stage is the highest-leverage point for asking bias and fairness questions. After signing, the agency is largely dependent on the vendor's voluntary cooperation.

What Evenhanded Looks Like: The Real Opportunity Alongside the Real Risk

This lesson is about bias and fairness, and it does not flinch from the research showing where AI systems can encode and amplify disparity. But evenhandedness requires stating the other side with equal clarity: AI tools, when designed well and used responsibly, can also reduce human bias in public safety documentation.

Human-generated police reports are also subject to the biases of their authors. Research has found that human officers, under stress, with limited attention, and shaped by their own training and experiences, produce reports that can vary significantly in language and characterization based on subject demographics. An AI tool constrained by a well-designed persona prompt, trained to describe only what the footage shows, and evaluated regularly for differential output patterns, can in principle produce more consistent, less variable documentation than a tired officer at the end of a long shift working from memory.

The goal is not to eliminate AI from police documentation. The goal is to use AI in a way that reduces the total amount of bias in the system rather than substituting AI bias for human bias or combining them. That goal is achievable. It requires rigorous procurement questions, regular disparity audits, override tracking, and persona engineering that explicitly prohibits differential language. None of those practices are technically difficult. They require institutional commitment and ongoing supervision, which is the harder part.

The 82 percent decrease in report-writing time that officers reported during Axon Draft One testing is a real and significant benefit. Officers who spend 30 to 40 percent of their shift on paperwork can recover substantial time for patrol, community engagement, and investigation when AI handles the drafting. That time recovery is worth pursuing. The pursuit should not require accepting uncritical AI outputs. It should be structured so the time savings and the accuracy and fairness requirements are both met.

Practical Steps for Officers and Supervisors

This section translates the theoretical framework above into concrete practices for officers, supervisors, and agency policy staff. The goal is not compliance theater. The goal is a set of habits that actually catch bias when it is present and build an institutional record that demonstrates responsible AI use.

The Individual Officer Level

Read your own AI drafts for differential language. When you are running the footage-grounded verification pass on an AI draft, include a check for language consistency: does the draft describe this subject's behavior in terms you would use for any subject in the same circumstances, or does it use language that seems calibrated to the subject's apparent demographics? If you notice language that seems differentially charged, correct it to match what the footage shows. Document the correction.

Do not accept demographic characterizations you cannot source to the footage. If the draft describes a subject as "aggressive," "combative," "threatening," or "erratic," and you cannot identify the specific footage element that supports that characterization, treat it the same as any other unsupported claim: correct or remove it. These words carry substantial weight in prosecutorial decisions, bail hearings, and plea negotiations. If the footage supports the characterization, it should be in the report. If the footage does not support it, it should not.

Flag unusual language patterns. If you notice that the AI tool consistently produces differently toned descriptions for similar incidents involving different demographics, flag that observation to your supervisor. One observation is a data point. A pattern across multiple officers is a systemic finding that should be escalated to the agency's AI oversight function.

The Supervisor Level

Build disparity monitoring into regular report review. If the agency is using AI-assisted report writing, the supervisor's review of reports should include a periodic check for differential language patterns. This does not require statistical analysis on every report. It requires the kind of pattern awareness that supervisors develop when they review a large volume of reports: a sense of whether the language used to describe subjects varies in ways that correlate with demographic factors rather than incident facts.

Track override rates for triage and classification systems. For dispatch and call-triage AI, the rate at which telecommunicators override the AI recommendation is a key quality metric. Low override rates in a system that produces biased recommendations means the bias is propagating without correction. Supervisors should know the override rate, should know whether it differs by call type, neighborhood, or demographic, and should investigate anomalies. The King County, Washington, prosecutors' decision to bar AI-written police reports was in part motivated by concerns about accountability and transparency in exactly this kind of oversight gap.

Conduct regular disparity audits. At least annually, and ideally quarterly, the agency should compare AI-assisted outputs against human-generated outputs on disparity measures: language patterns, priority classifications, outcome correlates by neighborhood and demographic. This does not require an external consultant for every audit. It requires a defined methodology, consistent data collection, and honest review of findings. When a disparity is found, it requires a documented response: is the disparity in the data the system was trained on? In the system prompt? In how officers are using the tool? The response depends on the diagnosis.

The Procurement and Policy Level

Ask bias and fairness questions before signing. The procurement stage is when the agency has maximum leverage. Contract language should require the vendor to provide bias evaluation methodology, findings, and corrective actions taken. It should require the vendor to notify the agency of model updates that could affect output patterns. And it should specify data access rights: the agency should be able to audit the system's outputs independently, not only rely on the vendor's self-reported metrics.

Document the CJIS compliance posture. The CJIS Security Policy places data handling obligations on the agency. When an AI tool processes criminal justice information, the agency is responsible for ensuring the tool meets CJIS requirements regardless of vendor assurances. That responsibility includes ensuring that the data fed to the AI tool does not create unacceptable data security or bias risks. The CJIS obligation does not transfer to the vendor on signing.

Build the review cycle into policy. An AI use policy that does not include a defined review cycle for bias and fairness is a policy that will drift. Technology changes. Model updates change behavior. Operational context changes. A policy that was accurate at deployment may be inaccurate two years later if it has not been reviewed. Build the review cycle in at the policy level, assign ownership to a named role, and make the review documented and public-facing where appropriate.

Key Takeaways

• Bias in public safety AI enters through training data that encodes historical disparities, feedback loops that strengthen biased recommendations when they are accepted without correction, and language patterns in generative tools that reproduce differential descriptions learned from historical reports.
• The EFF and criminal justice researchers have documented disparity patterns in predictive policing systems, call-triage tools, and risk assessment instruments. The COMPAS research is the most extensively analyzed case, but the structural mechanism, historical data encoding historical disparities, applies broadly.
• Generative report drafting tools trained on historical police reports may reproduce differential language patterns by subject demographics. This is both an equity concern and a direct evidentiary accuracy concern under Brady v. Maryland and the obligation to produce accurate sworn accounts.
• AI tools can also reduce human bias in documentation when designed well: a model constrained to describe only what the footage shows, evaluated for differential output, and used with a verification pass can produce more consistent documentation than a fatigued officer working from memory. The goal is a net reduction in total system bias, not the substitution of AI bias for human bias.
• Individual officers should check AI drafts for differential language during the verification pass, decline to adopt characterizations not sourced to the footage, and flag unusual patterns to supervisors. These are not burdens added to an already-full job. They are part of the verification discipline this program teaches throughout.
• Supervisors should track override rates for AI classification systems, conduct periodic pattern reviews of AI-assisted reports, and escalate systematic disparity findings to the agency's AI oversight function.
• Procurement is the highest-leverage point for bias and fairness requirements. Multi-year, sole-vendor contracts on the order of 45 million dollars and ten years create long-term obligations; the bias and fairness questions must be answered in the contract, not after signing.
• CJIS data handling obligations and report accuracy obligations stay with the agency, not the vendor. Building disparity audits, override tracking, and review cycles into policy is how an agency meets those obligations in an AI-assisted environment.