AI police reports are now being written, at least in first draft, by software that never attended the scene, and departments across the United States are adopting the technology fast enough that regulators are struggling to keep pace. Axon’s Draft One, the dominant product in this category, has generated more than 100,000 incident report drafts since its 2024 launch, built directly on top of the body-worn camera footage that Axon itself supplies to the majority of American police departments.
The pitch is straightforward: officers spend up to 40 percent of their working hours on paperwork, staffing shortages are acute nationwide, and AI can hand that time back. The complication is that a police report is not an ordinary business document. It is evidence, and in some cases it is the single most consequential piece of writing that determines whether a person keeps their liberty.
What has changed in the past year is not the technology itself but the regulatory response to it. California became the first state to require disclosure when officers use AI to help write official reports, and that law took effect at the start of this year. This piece explains how AI police report writing actually works, what the new disclosure rules require, and where the genuine tension sits between the efficiency gains departments are chasing and the due process concerns the technology raises.
How AI Police Reports Are Actually Generated
Draft One works by uploading the audio from an officer’s body-worn camera, transcribing it automatically, and using a large language model, built on OpenAI’s GPT-4 Turbo according to Axon’s own product documentation, to turn that transcript into a structured narrative in the format a police report typically follows. Axon says the system is calibrated with “creativity turned off” specifically to prevent the model from speculating or embellishing beyond what the audio actually captured, and that officers cannot submit a report until they have reviewed the draft, filled in any gaps the audio did not cover, and signed off on its accuracy.
That officer-in-the-loop design is the industry’s core defence against the most obvious objection: that a machine should not be allowed to describe what happened during an arrest. In practice, agencies that have adopted AI police report tools report time savings in the range of 40 to 67 percent on report writing, according to figures Axon has published from early adopter departments including Fort Collins, Colorado and Lafayette, Indiana. Whether that reviewed-and-signed final product is meaningfully different from a report an officer wrote unassisted is the question regulators, defence attorneys, and civil liberties groups have spent the past two years trying to answer.
California’s New Disclosure Law: What It Actually Requires
California’s Senate Bill 524, authored by state Senator Jesse Arreguín and signed by Governor Gavin Newsom, took effect January 1, 2026, and applies to every law enforcement agency in the state. Any official report generated fully or partially using AI must now carry a disclosure statement on every page identifying that AI was used and naming the specific program, along with the signature of the officer who prepared it. “AI hallucinations happen at significant rates, and what goes in a police report can influence whether or not the state takes away someone’s freedom,” Arreguín said when the bill was signed, framing the requirement explicitly around due process rather than efficiency.
The law’s most consequential provision may be its recordkeeping requirement rather than its disclosure requirement. Agencies using AI to generate a first draft or official report must maintain an audit trail identifying who used the tool and must retain that first draft for as long as the final report is kept on file, and a draft created with AI assistance cannot itself stand in as an officer’s sworn statement.
Kate Chatfield, executive director of the California Public Defenders Association, called the transparency requirement a due process matter directly: “Everyone in the legal system, judges, juries, attorneys, and the accused, deserve to know who wrote the police report,” she said. Utah passed a narrower disclosure requirement the same year, and civil liberties advocates expect more states to follow.
Why the Audit Trail Requirement Exists
The recordkeeping mandate directly targets a specific technical design choice that had already drawn criticism before SB 524 passed. According to the Electronic Frontier Foundation’s investigation into Draft One, the software was built to erase the AI-generated first draft once an officer exports the edited version into a records system, leaving no record of which portions of the final report were written by the model and which were written or changed by the officer. A senior Axon product manager confirmed the design choice directly in a public roundtable, saying the company deliberately does not store the original draft because doing so would create what he called disclosure headaches for police departments and prosecutors.
The practical consequence, EFF argues, is that an officer whose courtroom testimony contradicts their written report has a built-in explanation available that cannot be independently checked: the AI wrote that part. California’s audit trail and draft-retention requirements are a direct legislative response to that specific gap, designed to make it possible for a judge, defence attorney, or internal auditor to reconstruct exactly what the AI generated versus what the officer changed, something the software’s original design made deliberately difficult.
The Case for the Technology
Axon’s own account of the problem it is solving is not implausible on its face. Staffing shortages across American policing are well documented, and departments describe report writing as one of the least popular, most time-consuming parts of the job, one that keeps officers at a desk rather than out responding to calls. Axon commissioned an independent study, reviewed by 24 outside experts including district attorneys and inclusion scholars, comparing officer-only reports against reports produced with Draft One assistance. According to Axon, the results found Draft One reports equal to officer-only reports on comprehensiveness, neutrality, and objectivity, and rated more highly on terminology and internal coherence.
Individual officers quoted in Axon’s own marketing describe the change in blunt, personal terms: one early-access officer said the tool let them spend an evening with their family rather than finishing paperwork at home, a complaint about police work that predates AI by decades. None of that evidence is independent of the vendor, which is exactly why California’s disclosure and audit trail requirements matter regardless of whether the underlying technology performs as advertised: the law does not require departments to stop using AI, only to make its use visible and reviewable after the fact.
The Case Against Deploying It Faster Than Oversight Can Follow
The most pointed independent critique of AI police report writing does not primarily concern the quality of the AI’s prose. It concerns what happens when the same industry bundling body cameras, tasers, and report-writing software into single long-term contracts becomes the dominant supplier of the tools that generate, store, and can selectively withhold the evidence used against defendants. Multiple pilot departments, including Manchester, New Hampshire and Anchorage, Alaska, discontinued Draft One after concluding it produced no measurable time savings for their officers, a result that sits awkwardly next to the efficiency figures the vendor publishes from other agencies.
Connecticut paused statewide rollout of AI report writing software in 2026 pending further review by state’s attorneys and police chiefs, even though the state police had already signed a ten-year, $120 million contract with Axon that includes the option to add the tool. The King County prosecuting attorney’s office in Washington state went further, declining to accept any police narrative produced with AI assistance at all, writing in an internal memo that it did not yet trust the products currently on the market, language that reads as an acknowledgement that oversight has not caught up with adoption rather than outright rejection of the technology’s eventual potential.
What This Means for You
If you live in a jurisdiction where police have adopted AI report writing tools, the practical questions worth asking your local department or city council are specific ones: does the agency have a written policy governing when and how AI-generated drafts are used, does that policy require disclosure on the final report, and is the original AI draft retained in a form that a defence attorney could actually obtain through a records request. California residents now have a statutory right to at least the first two of those answers. Elsewhere in the country, the answer currently depends entirely on which department you happen to live under, since no federal standard exists and state laws remain a patchwork.
For anyone facing a criminal charge where a police report played a role, knowing whether AI assisted in writing it is no longer a theoretical question. It is a fact that, in a growing number of states, the report itself is now legally required to disclose, and one that a defence attorney should be asking about even in jurisdictions where the law does not yet require it to be volunteered.
The Pattern Repeating Across Institutions
The tension in AI police report writing, a genuine time-saving benefit set against a due process risk that falls hardest on people with the least power to contest it, mirrors a pattern showing up across other parts of the justice system as AI tools are adopted faster than the legal frameworks meant to govern them. LiveAIWire’s coverage of AI sentencing bias in predictive risk tools found the same underlying dynamic: a proprietary system making consequential decisions about a person’s liberty, deployed with limited independent audit, and defended largely by vendor-commissioned research rather than external oversight.
It also connects to a broader accountability gap in how facial recognition and other algorithmic policing tools are deployed with minimal transparency. LiveAIWire’s reporting on algorithms used to police the police themselves documented wrongful arrests traced to unaudited facial recognition matches, a failure mode that shares the same root cause SB 524 is trying to address in report writing: a technology treated as sufficiently reliable to act on before independent evidence confirmed that it was.
The broader digital rights response to that pattern, tracked in LiveAIWire’s coverage of the growing accountability infrastructure being built to audit AI systems used against the public, argues that the burden of proof should sit with the institutions deploying these tools, not with the individuals they are used against.
Where This Goes Next
The most likely near-term development is more states following California and Utah’s lead on disclosure, since the underlying concern, that a police report shapes prosecutorial and judicial decisions and should not be treated as more reliable simply because it reads more smoothly, is not unique to California’s legal culture.
What SB 524 does not resolve is the more structural question EFF and other critics have raised: whether a single vendor supplying body cameras, evidence storage, and report-writing AI to a majority of American police departments creates an accountability gap that disclosure requirements alone cannot close, since a report can fully comply with every disclosure rule and still be produced within an ecosystem where the same company controls both the evidence and the tool that summarises it.
The parallel that runs through nearly every professional field currently adopting generative AI to draft first-pass documents, whether police reports, legal briefs, or news copy, is the same one playing out here: the technology is genuinely capable of producing a serviceable first draft quickly, and the institutions adopting it are still building the oversight infrastructure needed to confirm that speed has not been purchased at the cost of accuracy, exactly the tension LiveAIWire has traced in coverage of declining editorial and professional standards as AI-generated first drafts become the default starting point across knowledge work generally. Police report writing is simply the version of that story where the stakes are someone’s freedom rather than a reader’s trust.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity, and the social impact of emerging technology. He publishes daily at LiveAIWire.com.