POST and the Council on Criminal Justice Just Set the Bar for AI Police Reports. Both Landed on the Body Camera.

By Chris Ryan, Chief Product Officer and former Lieutenant

Two of the most serious bodies looking at AI in policing have now put their standards on paper. The California Commission on Peace Officer Standards and Training released its 2026 Artificial Intelligence Guidelines, statewide operational guidance for agencies using AI, including report writing under Penal Code 13663 (California POST, 2026). Then the Council on Criminal Justice Task Force on Artificial Intelligence published a case study applying its national User Decision Framework specifically to AI-generated police report writing tools (Council on Criminal Justice, 2026).

One is a state regulator and the other is a national, nonpartisan policy group, and they came at the question from different directions. But read them side by side and they keep circling back to the same place: what happens when the report starts with the body camera instead of the officer. That is the question I have been writing about since these tools first hit the market, and it is the reason we built Policereports.ai the way we did.

Where POST and CCJ agree

Neither document says agencies should stop using AI to write reports. POST is explicit that it is not trying to discourage innovation, and CCJ classifies the whole category as both substantial risk and substantial opportunity, which under its framework means careful implementation rather than avoidance. Both treat the AI draft as exactly that, a draft. Both put the signature, and the accountability, squarely on the officer. And both warn that a polished narrative is dangerous precisely because it looks finished. POST tells officers not to rely on a draft just because it reads as complete and professional, and CCJ spends real time on automation bias, the tendency to accept machine output without enough scrutiny.

Where they get specific, though, is the input. POST names two kinds of covered systems under PC 13663: tools that draft narratives from in-car or body-worn camera audio and video, and tools that take an officer's dictation and enhance it with generative AI. CCJ notes that products differ in whether they draft exclusively from camera audio or rely on officer dictation, video, or a mix, and says those design differences determine what records exist, what can be audited, and how reliable the output is. In other words, the architecture is not a technical footnote. It is the whole ballgame.

The audio problem

Anyone who has worn a body camera knows that most of what it records has nothing to do with the incident. Radio traffic, a TV in the living room, a bystander talking over everyone, your partner on the phone. Feed all of that into a large language model and ask for a report, and you are asking the AI to decide what mattered. CCJ cites the two examples everyone in this space now knows by heart: the Heber City, Utah report that had an officer turning into a frog because a movie was playing in the background, and the King County, Washington errors, including an officer placed at a scene when he was only on the radio, that led the prosecutor's office to stop accepting AI-assisted narratives altogether (Brewster, 2026; Council on Criminal Justice, 2026). Those are not edge cases. They are what happens when the model is handed raw audio and left to sort it out.

Audio also misses everything nobody said out loud. Anchorage walked away from its audio-based tool because officers were stuck hand-entering every visual detail they never narrated on scene (Brewster, 2026). And the time savings never showed up. CCJ points to the randomized controlled trial by Adams and colleagues, which found no improvement in report writing speed, and says that finding contradicts vendor time-savings claims (Adams et al., 2024; Council on Criminal Justice, 2026). As I have written before, that study measured elapsed time rather than writing time, but the durable lesson holds: officers handed a narrative built from raw audio spend their time repairing it.

The part of the CCJ report that should worry every chief, though, is memory. CCJ warns that an AI draft generated from camera material can shape the officer's own recollection of what happened, creating a confirmation bias problem that undermines independent recall. Its recommended fix, where operationally feasible, is to have the officer document their own first-hand account before ever looking at the AI draft. Think about what that means. The best practice CCJ is asking camera-first agencies to bolt on is the exact workflow our platform starts with. When the officer dictates first, there is no machine version of events sitting in front of them to anchor on. The officer's memory is the input, not something the draft has a chance to overwrite.

The video problem

Some vendors have answered the audio problem by adding more raw data. Code Four, for example, generates reports from both body camera audio and video (Brewster, 2026). But more input the officer doesn't control is not the fix, and both of these new documents explain why.

POST is the most direct. Section 3.2 says that when a narrative is built from body-worn camera audio, video, or transcription, the officer's review has to account for what the AI output may not have captured or may have gotten wrong: observable conduct, visual cues, officer perceptions, safety concerns, prior knowledge, scene context, and the policy considerations that explain why the officer did what they did. It also requires that the final report reflect the officer's own observations and recollection (California POST, 2026). Put simply, POST just wrote down what every supervisor who has reviewed use-of-force footage already knows. The camera is not the officer.

It cuts both ways. A chest-mounted camera misses things the officer saw, like the hand that dropped toward the console from the angle of the driver's window, or the suspect the officer recognized from a call two weeks ago. But it also catches things the officer never saw, and that is where video-driven drafting gets genuinely dangerous. Say the footage picks up a knife on the floorboard that the officer never noticed during the stop. A video-driven narrative can fold that detail into the report as though it informed the officer's decisions. Under Graham v. Connor, force is judged from the perspective of a reasonable officer on scene, based on what that officer perceived in the moment. A report that credits the officer with perceptions they never had is not a more complete report. It is an inaccurate one, and it will come apart on cross.

CCJ treats video as a risk escalator too. Among the events it says should trigger an off-cycle reassessment of an AI report writing tool is the product adding multimodal video capabilities (Council on Criminal Justice, 2026). When a national framework lists adding video as a reason to stop and re-examine a tool, that should tell you something.

None of this means body camera footage has no place in the process. It absolutely does. On our platform, video, photos, interviews, and recordings support and corroborate the officer's account, and our investigative tools help detectives work through hours of footage and jail calls. But the narrative starts with the person who has to swear to it, and the footage serves that account rather than replacing it.

The prompt problem, and why we keep everything

The sharpest finding in the CCJ report is one most people outside the vendor world have never thought about: the system prompt. Every AI report tool runs on instructions written by the vendor that shape how the model turns input into a narrative. CCJ points out that those prompts materially shape the report but often aren't retained, versioned, or produced in discovery, and that neither California's SB 524 nor Utah's SB 180 fully closes that gap. It goes further and says a vendor's refusal to retain and produce prompts in discovery should be treated as a trigger to generally avoid the product, not a point to negotiate (Council on Criminal Justice, 2026). It recommends agencies treat the prompt, the initial AI draft, and the final report as one set, all subject to discovery.

This connects directly to the courtroom. CCJ raises a Sixth Amendment concern: officers may not be able to testify clearly about which parts of a report were AI-generated, which were edited, and what they actually remembered. It also flags Brady exposure when the initial draft or the prompt contains material information that differs from the final report. The Electronic Frontier Foundation found in 2025 that Axon's Draft One did not save the AI-generated first draft at all (Guariglia & Maass, 2025). In fairness, Axon's current documentation now references an original AI-generated draft retention report, likely in response to SB 524, so agencies should ask exactly what is kept and for how long. But draft retention alone still leaves the prompt question CCJ is raising.

We built for this from the beginning. Policereports.ai runs a two-tiered verification architecture: a generator drafts from the officer's own input, and a quality assurance checker validates that draft back against the original input, flagging fabrications, contradictions, speculation, and inconsistencies for the officer to resolve. We store every iteration of the process: the original input, the prompt, the generated draft, and the final signed report, all encrypted, timestamped, and auditable. When a defense attorney asks your officer who wrote what, the answer isn't a guess. It's in the record. And because the input is the officer's own account rather than ambient footage, the comparison between what the officer said and what ended up in the report is one a judge can actually follow.

Meeting SB 524 and the POST guidelines

Let me be clear about one thing. POST lists dictation-based tools like ours right alongside camera-based tools as covered by PC 13663, so the statute applies to us in full and we don't pretend otherwise. Here is how we meet it:

• Disclosure: the required statement that the report was written fully or in part using AI is applied to the report, along with the AI program used.
• Verification: the officer signs and attests that they reviewed the report and the facts are true and correct.
• First draft retention: the AI-only first draft is preserved for the same retention period as the official report, never overwritten on edit or export.
• Audit trail: we record who used AI to create the report and what source material went into it, for the full retention period.
• Vendor data restrictions: agency data is used only for the agency's purposes. We never sell or share it, and we never train on agency report data.

POST goes past the statute, and so do we. Its guidelines tell agencies to be skeptical of vendors claiming their system is hallucination-proof or bias-free, and to demand real-world evidence rather than marketing (California POST, 2026). You won't see that claim from us, because no honest vendor can make it. What we can do is make errors easy to catch and impossible to hide, and point you to agencies like Nampa PD, where two years of daily use and a prosecutor's office that worked through the process with them speak louder than any brochure.

Questions every chief should ask

Between POST and CCJ, agencies now have a procurement checklist written by people with no product to sell. Use it on every vendor, ours included.

1. Does the narrative start from the officer's account or from the camera? If it's the camera, how does your workflow protect the officer's independent recall?
2. If you use video, how do you keep details the officer never perceived from showing up in the narrative as if they did?
3. Do you retain the system prompt, the AI-only first draft, and the final report as a set, and will you produce them in discovery?
4. Will you notify us before changing the underlying model or the prompts?
5. Does anything verify the draft against the original input before the officer sees it?
6. Is any of our data used to train your models or improve the product for other customers?
7. What independent evidence do you have beyond your own customer surveys, and which agencies like ours can we call?
8. Do you claim your system can't hallucinate? If so, POST says to be cautious, and so should you.

The bottom line

POST and CCJ didn't hand the industry a ban. They handed it a standard, and the standard is clear: the officer's account comes first, the officer owns the report, and the agency has to be able to show every step of how it was made. Feed a model raw body camera audio and video and make the officer an editor, and you are fighting that standard at every turn, from memory contamination to perceptions the officer never had to prompts nobody can produce. Start with the officer, verify against their input, and keep every iteration, and you meet it by design.

That is the bar. Every vendor should welcome it. We built our company around it.

If you want to walk your current setup through the POST guidelines and the CCJ framework line by line, book a 30-minute walkthrough with a former officer or run your numbers with our free savings calculator.

Chris Ryan is a former Lieutenant and the Chief Product Officer at Policereports.ai.

References

Adams, I.T., Barter, M., McLean, K., Boehme, H., & Geary, I.A. (2024). No man's hand: Artificial intelligence does not improve police report writing speed. Journal of Experimental Criminology, 22, 137-154. https://doi.org/10.1007/s11292-024-09644-7

Brewster, T. (2026, July 22). Axon says AI police reports save time. Public records show they get facts wrong. Forbes. https://www.forbes.com/sites/thomasbrewster/2026/07/22/axon-says-ai-police-reports-save-time-public-records-show-they-get-facts-wrong/

California Commission on Peace Officer Standards and Training. (2026). Artificial intelligence guidelines. https://www.post.ca.gov

Council on Criminal Justice. (2026). Assessing AI in action: A case study on police use of AI-generated report writing tools. https://counciloncj.org/assessing-ai-in-action-a-case-study-on-police-use-of-ai-generated-report-writing-tools/

Guariglia, M., & Maass, D. (2025, July 10). Axon's Draft One is designed to defy transparency. Electronic Frontier Foundation. https://www.eff.org/deeplinks/2025/07/axons-draft-one-designed-defy-transparency

Graham v. Connor, 490 U.S. 386 (1989).