The room in City Hall was smaller than Deputy Chief Mariana Solano had expected: eleven chairs around a conference table, two rows of public seating behind them, a wall-mounted screen showing a body-worn camera (BWC) footage freeze-frame from an incident the board had been reviewing for six weeks. Five members of the civilian oversight board sat across from her. Two were retired educators. One was a civil rights attorney. One ran a neighborhood mediation program. One was a former police dispatcher. None of them were hostile. All of them had read the Electronic Frontier Foundation (EFF) report on AI police reports, and all of them had questions she had not been asked before.

Transparency as Legitimacy: The Framework That Changes the Room

Deputy Chief Solano had walked into previous oversight board meetings in a defensive posture: prepared to explain decisions, to justify operations, to manage the gap between what the board suspected and what the agency was actually doing. This meeting was different, and she had prepared for it differently. She had prepared to be transparent, not because transparency was a legal obligation in this particular forum, but because she had come to understand a structural truth about civilian oversight: the agency that shares information proactively earns legitimacy; the agency that shares information only when compelled destroys it.

Legitimacy in this context is not a soft concept. It is operational. An agency that the community trusts can deploy new tools, expand programs, and respond to incidents in a climate of at least minimal good faith. An agency that the community does not trust faces a headwind on every decision: the new technology becomes a surveillance program, the efficiency gain becomes a cover for cutting accountability, and the oversight board that might have been a partner becomes an adversary. The cost of that headwind is real. It shows up in hiring, in community cooperation with investigations, in city council budget votes, and in the political cover the chief needs to run a program that is actually working.

The EFF has raised specific, substantive concerns about AI police reports. Those concerns are worth stating directly, because an agency that cannot articulate the legitimate case against its own program cannot responsibly run it. The EFF has questioned whether agencies are being sufficiently transparent with prosecutors, defense attorneys, and the public about the role AI plays in generating evidence. It has raised concerns about the auditability of AI-drafted reports, about whether the public can access information about how the tool works and what its error rate is, and about whether oversight bodies have been given meaningful access to AI program documentation. These are not frivolous objections. They are the same objections that led the King County, Washington prosecutor's office to bar AI-written reports from cases it would charge. They deserve a direct, substantive answer, not a dismissal.

Transparency is not a concession to oversight. It is the mechanism by which a law enforcement agency demonstrates that its AI program is working the way the policy says it is, and earns the legitimacy to keep running it.

Deputy Chief Solano's answer to the EFF's concerns was not a defense of the technology. It was a description of the governance: here is the use policy, here is the verification standard, here is the disclosure statement that goes on every AI-assisted report, here is the correction history that the prosecutor's office can access, here is the audit trail. She then paused and said: "I am here to answer your questions and to hear your concerns. If there are gaps in what I have described, I want to know about them." That posture, genuine openness to the board's findings rather than a performance of openness, is what changed the room.

What Civilian Oversight Boards Actually Want

It is worth naming what civilian oversight boards are and what they are not, because misunderstanding their function leads to mismanaging the relationship. A civilian oversight board is not a jury. It is not a disciplinary body with unilateral authority to override operational decisions (in most jurisdictions; some boards do have formal authority, and the agency's legal advisor should know the specific authority structure in their jurisdiction). It is a mechanism for community accountability: a standing body charged with reviewing police operations, receiving complaints, and making recommendations that the elected government acts on or does not act on.

What oversight boards actually want, in most cases, is not to stop the AI program. It is to understand it well enough to represent it honestly to the community. When a community member comes to the oversight board and says "I heard the police are using AI to write reports, is that true, and should I be worried?", the board member needs an answer. If the only answer they have is "I've asked the department and they haven't told me anything," the board member's credibility with the community is damaged. If the answer is "Yes, and here is how it works, here is how the officer verifies it, and here is how you can find out if it was used in a case involving you," the board member becomes an informed participant in governance rather than an uninformed audience to it.

The specific concerns civilian oversight boards raise about AI police reports fall into predictable categories, and agencies should be prepared to address each of them:

Accuracy and Error: Can the AI Get It Wrong in Ways That Hurt People?

Yes. This is the answer, stated plainly. Axon Draft One, which drafts patrol narratives from body-worn camera audio, can produce what the AI field calls hallucinations: confident, fluent statements of fact that are not grounded in the audio. In the report-writing context, these take three specific forms: the gap-fill (inventing a detail to complete a narrative where the audio had a gap), the softened fact (underrepresenting a use of force or a subject's behavior), and the invented quote (attributing words to a subject or witness that the audio does not support). Each of these errors is a problem in a sworn report, because a sworn report is evidence, and evidence must be accurate.

The way the agency addresses this is the footage-grounded verification pass: the officer checks every factual claim in the draft against the body-worn camera recording, the computer-aided dispatch (CAD) entry, and their field notes before adopting the report. CAD is the timestamped record from the dispatch center. The records management system (RMS) is the database where adopted reports are stored. The verification pass is not optional. It is the standard, and it is logged.

The oversight board's legitimate follow-up question is: how does the agency know the verification pass is actually being completed? The answer must be honest. Verification completion is logged in the audit trail, supervisors review reports flagged for AI assistance, and the quality-metrics program tracks error rates across AI-assisted and hand-typed reports to detect patterns. That is a complete answer. An agency that does not yet have all of those mechanisms in place should say so, and say what the timeline is for putting them in place. Aspirational governance that the board discovers has not been implemented is worse than honest acknowledgment of a gap under active remediation.

Bias: Does the AI Treat People Differently Based on Who They Are?

This is one of the most important concerns civilian oversight boards raise, and it deserves a careful answer rather than a defensive one. AI models are trained on existing data. If the existing data reflects historical patterns of differential treatment, the model may replicate those patterns in its outputs. In the report-writing context, this concern is most acute in the descriptions of subjects' behavior and demeanor. If the model has learned from training data that certain call types or locations are associated with certain subject behaviors, it may apply those associations in ways that are not grounded in what the specific footage shows.

The oversight board's concern here is legitimate. The honest answer is that this risk exists, that it is addressed partially but not completely by the verification standard (because a verification pass that checks individual facts may not surface a subtle pattern of language applied differently across demographic groups), and that the agency's quality-metrics program should include a regular review of AI-assisted report language across cases to detect such patterns. This review is not the same thing as the per-report verification pass. It is an aggregate review that surfaces systemic problems rather than individual errors.

An agency that has not conducted this aggregate review should acknowledge that to the oversight board and commit to a timeline. An agency that has conducted it and found no concerning patterns should share those findings with the board. Either answer builds more trust than a blanket assurance that bias is not a problem.

Surveillance Creep: Is This the Beginning of Something Larger?

Civilian oversight boards that ask about AI-assisted report writing are often asking a broader question: is this the first step toward a more comprehensive AI surveillance capability that the community has not consented to? That concern is not paranoid. Bundled vendor contracts in 2026 commonly package body-worn cameras, license plate readers, drone programs, cloud evidence storage, and AI analytics into a single multi-year procurement. Contracts of this scale, reaching approximately $45 million over ten years for mid-size agencies, can lock an agency into a vendor ecosystem for a decade, and the AI capabilities within that ecosystem may expand over the contract period without explicit re-authorization from the city council or the oversight board.

The honest answer to the surveillance-creep question acknowledges this dynamic. The agency's current AI use policy covers the specific tools currently authorized. Any expansion of AI capability, including new uses of data already collected through the existing platform, requires a policy review and, for significant expansions, a return to the oversight board or the city council. The answer is not "don't worry, we're not doing that." The answer is "here is the governance structure that ensures you would know about it and have a say in it before it happens."

This answer requires the governance structure to actually exist. If the agency's AI use policy does not include a process for authorizing new AI tools or new uses of existing tools, the oversight board's concern about surveillance creep is well-founded. The policy should be revised before the next oversight board meeting to include that process, and the board should be told that the revision was made in response to their concern.

The Community Engagement Model: Before, During, and After

A single oversight board meeting is not a community engagement program. Deputy Chief Solano understood this. The oversight board is one channel; it is not the community. The community includes people who will never attend a board meeting but who interact with officers every day, who have family members who have been arrested or cited or cleared, and who read the local news coverage of police AI stories with a combination of skepticism and genuine concern.

Community engagement on AI programs should happen at three points: before deployment, during operation, and after significant incidents.

Before Deployment: The Public Notice and Comment Window

Before deploying an AI-assisted report-writing tool at scale, the agency should provide public notice and an opportunity for community comment. This is not a legal requirement in most jurisdictions (some have passed algorithmic accountability ordinances that do require it, and the agency's legal advisor should know whether theirs is one of them). It is a best-practice standard for agencies that want to build trust rather than manage a disclosure crisis after the fact.

The notice should explain, in plain language: what tool is being deployed, what it does, what the officer's role is in reviewing and adopting the output, what the verification standard is, what is prohibited, and how the community can learn whether AI was used in a case involving them. The comment period should be genuine: the agency should read the comments, respond to specific concerns in writing, and, where the comments surface a legitimate concern the agency has not already addressed, revise the use policy or the deployment plan.

An agency that holds a comment period and then proceeds with a deployment unchanged, without responding to any of the comments, has achieved the worst of both worlds: it has signaled that community input is a box to be checked, not a conversation to be had. That signal travels faster than any press release about the program's benefits.

During Operation: Transparency Reporting

Once the program is operating, the agency should publish regular transparency reports, at minimum annually, that cover: the number of AI-assisted reports submitted in the reporting period, the error rate detected during verification (how many corrections were made per report, and of what type), the disclosure compliance rate (what percentage of AI-assisted reports included the standard disclosure statement), any incidents in which an AI-assisted report required supplemental documentation in prosecution, and the results of the bias review across AI-assisted report language.

These reports should be written for a general audience, not a technical one. They should be available on the agency's website without a public-records request. They should be presented to the oversight board at the annual review meeting. And they should be honest: if the error rate in the first quarter was higher than expected, say so, say what was done about it, and report the second-quarter rate alongside it. An agency that publishes an honest report of a problem it has addressed is doing governance. An agency that publishes only numbers that look good is doing public relations, and the oversight board will notice the difference.

After Incidents: The Case-Specific Response

When an AI-assisted report becomes the subject of a legal challenge, a disciplinary review, or a community complaint, the agency's response should be prompt, specific, and, to the extent possible, public. "We take this seriously and we are reviewing it" is not a response. A response describes what happened, what the verification standard required, whether it was followed, and what the agency is doing to address the gap if it was not followed.

The oversight board should be briefed on significant AI-related incidents before they read about them in the news. Briefing the board first, when the agency has enough information to say something substantive, is the structural expression of the transparency principle. It acknowledges that the board's role is to provide oversight, and that oversight requires information. An agency that briefs the board first and the press second builds a working relationship with the board. An agency that briefs the press first, or lets the board read about it in the paper, signals that the board is not actually part of the governance structure.

Giving the Evenhanded Account: Opportunity and Concern Together

Deputy Chief Solano made one decision in preparing for the oversight board meeting that she later identified as the single most important thing she did: she did not prepare a defense of the AI program. She prepared an honest account of it. The account included the efficiency gains (officers reporting an 82 percent decrease in report-writing time, returning hours to patrol and community engagement), the verification standard and why it is necessary, the disclosure policy, the known limitations and error modes of the tool, the EFF's concerns and how the agency's governance addresses them, and the questions that are still open.

The board member who was a civil rights attorney asked about Brady v. Maryland and Giglio v. United States specifically. Brady requires disclosure of exculpatory evidence; Giglio requires disclosure of impeachment material about government witnesses, including the method by which a report was generated if that method is material to its reliability. Deputy Chief Solano answered directly: the disclosure statement and the correction history are the agency's Brady and Giglio answer. They document the AI's role, the officer's review, and the corrections made, so that the defense and the court have everything they need to evaluate the report. That answer satisfied the attorney on the board. It would not have satisfied her if it had been delivered as an assurance rather than a description of a specific, implemented practice.

The board member who ran the neighborhood mediation program asked whether community members could find out if an AI tool was used in a case involving them. The answer should be yes: the disclosure statement is part of the case record, and community members with the right to access their own case file can see it. The agency should also establish a specific contact for community members who want to request information about AI use in a case involving them, one that does not require a formal public-records request as the first step.

The former dispatcher on the board asked a question Deputy Chief Solano had not anticipated: "What happens to the AI draft before the officer adopts it? Where does it go, and who can see it?" The answer required knowing the vendor contract's data handling provisions: the pre-adoption draft is stored on the vendor's cloud platform, subject to the CJIS (Criminal Justice Information Services) security agreement, retained for a specified period, and accessible only to authorized personnel within the agency's evidence platform. The dispatcher nodded and wrote something down. She had understood, from her own operational experience, that custody of information before it is formalized is as important as custody after. That question, and the agency's ability to answer it specifically, built more trust than any prepared statement about the program's benefits.

What Trust Looks Like, and How You Lose It

Trust between a law enforcement agency and its oversight board, and between the agency and the community the board represents, is built slowly and lost quickly. In the context of AI programs, the speed of loss is particularly high because the public narrative around police AI is already skeptical. Every story about a wrongful outcome in which AI played an undisclosed role becomes, in the public's understanding, a story about the agency's program, not just about the specific case. The agency that has built a transparent, documented, governed AI program is insulated from that narrative. The agency that has not built it is exposed to it.

Trust is lost in predictable ways. It is lost when the board learns that the agency deployed a tool without telling them. It is lost when a community member discovers through a news story that an AI tool was used in their case and the agency had not established a way for them to find that out. It is lost when the error-rate data the agency reports does not match what defense attorneys are finding in case files. It is lost when the agency's response to a legitimate concern is a defensive performance of openness rather than actual openness.

Trust is built in equally predictable ways. It is built when the agency brings the board a policy before the tool is deployed, not after. It is built when the transparency report includes numbers that are not all good, alongside an honest account of what was done about them. It is built when the agency's legal advisor and the board's legal resource can look at the same documents and reach the same conclusions about what the governance structure requires. And it is built when, as happened in Deputy Chief Solano's meeting, the chief officer in the room listens to a question from a civilian board member about pre-adoption data custody and says, "That is a good question, and here is the complete answer."

The accountability move and the efficiency move are the same move. An agency that runs its AI program transparently, with a documented verification standard and a genuine engagement with oversight, is not sacrificing efficiency for accountability. It is building the political and legal foundation on which the efficiency gains can be sustained. An oversight board that receives a genuine transparency report is not going to recommend a moratorium on a program that is clearly working and clearly governed. An oversight board that is kept in the dark, and then reads about an AI-related case problem in the newspaper, will.

Key Takeaways

• Transparency is not a concession to oversight. It is the mechanism by which the agency demonstrates that its AI program is working as described and earns the legitimacy to continue running it. The agency that shares information proactively builds legitimacy; the agency that shares information only when compelled destroys it.
• The EFF's concerns about AI police reports, auditability, transparency with prosecutors and the public, and meaningful access for oversight bodies, are legitimate and deserve direct, substantive answers, not dismissals. The governance structure is the answer.
• Civilian oversight boards are not adversaries of well-governed AI programs. They are the mechanism by which community trust is built and maintained. Their questions about accuracy, bias, and surveillance creep are predictable, and agencies should prepare substantive answers to each of them before the first meeting.
• The AI use policy must include a process for authorizing new AI tools or new uses of existing tools, with the oversight board and city council in the approval chain for significant expansions. Without that process, the oversight board's concern about surveillance creep is well-founded.
• Community engagement should happen at three points: before deployment (public notice and a genuine comment period), during operation (annual transparency reports available without a public-records request), and after significant incidents (briefing the board before the press).
• Brady v. Maryland and Giglio v. United States make AI-assisted report writing a constitutional disclosure question. The disclosure statement and the correction history are the agency's Brady and Giglio answer, and they must be described as implemented practices, not assurances.
• The agency that publishes an honest transparency report, including error-rate data that is not uniformly positive, alongside an account of what was done to address it, is doing governance. An agency that publishes only favorable numbers is doing public relations, and the oversight board will notice the difference.
• The accountability move and the efficiency move are the same move. A transparent, governed AI program with a documented verification standard and genuine oversight engagement is the foundation on which the 82 percent report-time reduction becomes a sustainable operational gain, not a political liability.