Investigation
Bend Privacy Alliance / Axon patents series, Part 11 of 11 · September 2026
Police departments are beginning to use generative AI to turn body-camera audio into official reports. The harder question is whether the path from evidence to narrative can still be reconstructed later.
When a Bend police officer uses Axon Draft One, the final report still carries the officer’s name.
But the path to that report looks very different from the one police departments have relied on for generations.
Under Bend Police Department Standard Operating Procedure A-2026-1.1, revised July 21, 2026, an officer docks a body-worn camera and waits for the recording to upload to Axon Evidence. The audio is automatically transcribed. The officer selects the evidence file, answers questions about the incident, charge severity, arrest status, and desired report length, then asks Draft One to generate a narrative.
The officer can review the result, add narrated information, and edit the text. Draft One may insert bracketed prompts where information is missing. Bend has also configured the system to place obvious errors into the draft that must be removed before the report can be copied. The officer then digitally acknowledges that Draft One was used, certifies the report as an accurate representation of their recollection, and pastes the narrative into the department’s Mobile Field Reporting system.
The process is designed to keep a human officer responsible for the final report, but it also creates something new: a police narrative whose first coherent version may have been written by a generative AI system rather than by the officer whose name appears on it. The deeper question is not simply whether AI can “hallucinate,” but whether the path from evidence to official narrative can still be reconstructed later. Accuracy, efficiency, authorship, and oversight all flow from that problem.
What Draft One actually does
Axon markets Draft One as an AI-assisted report-writing tool built around police evidence. Its current product materials publicly identify OpenAI’s GPT-4 Turbo as the model used by Draft One, although Axon’s own wording is inconsistent about the exact role that model plays.
Early descriptions of Draft One were relatively simple: body-camera audio is transcribed, and a generative model turns the transcript into a draft narrative.
The current product is broader. Axon’s Draft One generation documentation says officers can generate reports from one or more supported audio or video evidence items and add post-incident narration or typed context.
In supported Axon Records workflows, Draft One can go further. April 2026 release notes say it can incorporate information about people, vehicles, and property and ask officers targeted questions about missing or unclear details.
Taken together, those features produce a more complicated input chain: evidence and transcripts can be combined with structured report fields, model-generated questions, officer answers, and narration or typed context before the narrative is generated.
Not every agency uses all of those capabilities. Bend’s July 2026 procedure expressly says its trial does not use data from Computer Aided Dispatch or records-management software. Bend describes a narrower workflow centered on body-camera audio.
That distinction matters because there is no single Draft One deployment. The product Axon sells, the settings an agency enables, and the policy an agency adopts can all materially change what the software is permitted to do.
Accuracy is more than polished prose
Axon has published research arguing that Draft One can improve police-report quality.
In a double-blind report-quality study, evaluators compared officer-written reports with reports produced using Draft One and then reviewed and finalized by officers. Axon reported that Draft One-assisted narratives scored better on terminology and coherence and performed similarly on completeness, neutrality, and objectivity.
Those are meaningful results, but they are not the same as a real-world factual error rate.
A report can be coherent, neutral, and professionally written while still containing the wrong name, an omitted fact, a misidentified speaker, an altered sequence, or a detail the underlying evidence does not support.
Axon’s own documentation acknowledges that problem. Its Draft One FAQ warns that narratives can contain mistakes and that quality may be lower in chaotic incidents, when multiple people speak at once, in unsupported languages, in serious crimes, and during extended interviews. Axon instructs officers to review generated reports carefully rather than assume they are complete or accurate.
The public evidence reviewed for this article does not yet provide a large, independent, real-world Draft One dataset answering the most basic factual questions:
How often does a raw Draft One draft contain a material factual error?
How often do officers catch those errors?
How often does an error survive into the final submitted report?
Those questions are different from asking whether the prose is coherent, and they become harder because Draft One sits at the end of a chain.
Four places an error can enter
Draft One sits at the end of a chain, and an error can enter at multiple points.
The first is the source recording itself. Body-worn cameras do not capture everything an officer sees. Events happen outside the camera’s field of view. Relevant information may be learned later. An officer may notice something without saying it aloud.
Bend Police had acknowledged this limitation years before Draft One entered its workflow. In Policy 420, its body-worn camera policy, the department cautions that a recording may not represent the complete incident. During the City Council’s 2021 body-camera discussion, Chief Mike Krantz similarly emphasized that camera footage is only one piece of evidence: it may fail to capture everything an officer saw, capture things the officer did not perceive, and must be interpreted in context.
The second layer is transcription. Axon warns that transcripts can contain misheard words, misspellings, and missing context. Overlapping speech, background noise, accents, movement, and difficult proper nouns all complicate automated transcription.
The third layer is the generative model. Draft One has to transform transcripts and other inputs into a coherent police narrative. That requires selection, organization, omission, and phrasing. A model can turn ambiguity into prose that sounds more definite than the underlying evidence.
The fourth layer is human review. An officer may catch an error, or may not; the same is true of a supervisor, and even a correction can be incomplete.
That is why spectacular anecdotes about AI errors are less important than they first appear. A bizarre mistake is often easy to detect. The more consequential problem is a plausible mistake that survives because the generated prose sounds ordinary.
Even one widely circulated Draft One anecdote illustrates the danger of treating viral examples casually. News reports have repeated a story about an AI-generated police report claiming that an officer “turned into a frog.” But local reporting conflicts over which Utah department actually generated the report, and Axon itself deliberately inserts absurd statements into some Draft One drafts when an agency enables its Obvious Errors safeguard.
Axon’s administration guide gives examples of intentionally inserted absurdities involving trained squirrels, walking the plank, rock-paper-scissors, and secret hideouts inside paintings.
Those examples are review tests, not hallucinations. Bizarre errors are still possible, but the source of an apparent error matters.
Does Draft One actually save time?
The strongest independent evidence so far complicates one of Draft One’s central selling points.
A preregistered randomized controlled trial at the Manchester Police Department in New Hampshire tested Draft One with 85 officers and 755 reports.
The study, “No man’s hand: artificial intelligence does not improve police report writing speed”, found no statistically significant reduction in total report completion time.
The researchers tested multiple alternative specifications and still found no statistically significant effect. A separate one-year difference-in-differences analysis covering 6,084 reports likewise failed to find a statistically significant reduction.
The distinction between writing speed and total workflow time is important.
Police reports are more than narratives. Officers still enter structured information, review recordings, correct transcripts, edit generated text, and sometimes revise reports after supervisor review.
Draft One can generate prose quickly without necessarily reducing the time required to finish the report.
A follow-up study found an interesting disconnect between perception and measurement. Many Manchester officers and supervisors liked the system or believed it improved reporting even though the earlier experiment found no statistically significant total-time savings.
Among treated officers, 52 percent reported problems including missing details, poor alignment with body-camera timelines, difficulty with complex reports, or extensive editing requirements.
Officers could still reasonably find the tool useful. It may reduce the mental burden of staring at a blank page, improve formatting, or simply feel easier to use without shortening the full reporting workflow.
Anchorage, Alaska, reached a similar operational conclusion outside an academic experiment.
According to Alaska Public Media, Anchorage Police ran a free Draft One trial for roughly three months in 2024 and chose not to continue. Department officials said they had hoped for significant time savings but found that officers still had to review and correct the AI-generated text and add visual observations that were never spoken aloud.
In Anchorage, the expected savings disappeared. Axon has a fair response to both examples: these agencies were evaluating an early 2024 version of Draft One, and the product has changed substantially since then.
The Manchester and Anchorage results should not be treated as a final verdict on the 2026 product. They establish something narrower: generating a narrative faster is not the same as completing a police report faster.
Draft One can change what happens before the report is written
One of the most interesting findings from Manchester had little to do with report-writing speed.
Officers were trained to verbalize their observations and actions more completely because Draft One would later rely on the body-camera transcript.
Other agencies describe similar practices.
Arvada, Colorado, says on its Draft One transparency page that officers receive training on making minor adjustments to interviews and contacts to produce higher-quality reports.
The Hillsborough County Sheriff’s Office in Florida has told deputies to clearly narrate events and observations because Draft One relies on the audio.
That creates a feedback loop:
officer knows AI will later consume the body-camera audio → officer changes what they say → transcript changes → generated report changes
More contemporaneous verbal documentation could preserve details that might otherwise be forgotten, but the practice also raises accountability questions. The body-camera recording is no longer just evidence created before the report; it may also be partly shaped for later machine consumption. Nothing in the evidence reviewed here establishes that this practice is improper or biased, only that Draft One can affect police documentation earlier than the moment an officer clicks “Generate Draft.” What happens next depends heavily on local rules.
There is no single Draft One deployment
The biggest mistake in discussing Draft One nationally may be treating adoption as a yes-or-no question.
Bend’s trial is comparatively restrictive.
Its procedure excludes Measure 11 crimes, homicide, sex crimes, and officer-involved deadly force. Probationary officers may not use the system. Draft One-generated reports may not be attached, quoted, or directly relied upon when preparing search-warrant affidavits. Bend requires an acknowledgment, enables obvious errors, creates an audit log, and says the initial AI-generated draft is maintained.
Other agencies make very different choices.
Fargo Police Department allows Draft One for any reportable incident. Its policy uses a 10 percent modification safeguard intended to push officers toward meaningful review.
Orem Police Department expressly includes warrant affidavits among the official law-enforcement narratives that may be created with Draft One.
Colorado Springs Police Department permits AI assistance with probable-cause affidavits, search warrants, arrest warrants, and court orders, but only through a special Narrate Instead workflow.
Pasco Police Department requires an officer acknowledgment connecting the final report to the officer’s recollection, willingness to testify, and a declaration under penalty of perjury.
Hillsborough County launched Draft One only for non-criminal calls for service.
Seattle Police said in a 2025 oversight response that it had no plans to adopt AI report writing because the King County Prosecuting Attorney’s Office would not accept AI-assisted police narratives.
The same product can therefore operate under radically different rules. Saying an agency “uses Draft One” tells the public little unless we also know how it is configured and governed.
The result is not an absence of oversight so much as fragmented oversight. One agency may rely on a dedicated SOP, another on tenant-level settings, another on prosecutor restrictions, and another on state law. Case exclusions, disclosure, draft retention, auditability, warrant use, and review requirements are therefore distributed across different institutions rather than governed by one common standard.
What survives from evidence to final report?
The strongest unresolved issue may be less about generation than reconstruction. A Draft One report can involve at least eight separate layers:
- the source recording;
- the machine transcript;
- officer narration or typed context;
- the original AI-generated draft;
- system-added prompts or verification material;
- officer edits;
- the final police report;
- audit metadata.
Axon’s auditing documentation says Draft One can log the user, IP address, Evidence IDs, incident characteristics, generation events, and acknowledgment.
Since December 2025, U.S. agencies can also enable retention of the original unedited AI draft.
That was a significant product change. Earlier critics, including the Electronic Frontier Foundation and the King County Prosecuting Attorney’s Office, objected that Draft One did not preserve the original AI output. Their criticism accurately described the earlier product state, but it is no longer a complete description of Draft One in 2026. The new retention system still does not preserve everything.
Axon says the retained draft is the untouched AI-generated version. Intermediate officer edits are not preserved by that Draft One retention feature.
That distinction matters for policy design. Preserving the first AI draft is more informative than preserving only the final report, but it is still not the same as preserving the full path between them. A stronger provenance record would include the first AI draft, final report, audit trail, source-evidence identifiers, material officer inputs, and—where technically available—intermediate revisions.
Some officer inputs may disappear even earlier. Axon’s March 2025 Records release notes say post-incident narration audio and its transcript are not saved as evidence. The audit trail can indicate that narration was used without preserving the actual content of that narration.
Colorado Springs’ own policy describes the same problem. Under its Narrate Instead workflow, the officer’s narration audio and transcribed text are not stored in Evidence.com or the audit logs.
That means an official narrative can potentially contain information derived from:
officer narration → temporary transcript → AI draft → final report
without preserving the actual verbal input that entered the model.
Another configuration wrinkle is easy to miss. Axon’s current administration documentation says agencies can assign roles that are allowed to access transcripts and metadata and use Draft One while lacking permission to view or hear the underlying evidence media.
So the product technically supports a workflow where someone can generate a report from the transcript without being authorized to watch or listen to the recording itself.
A local policy may require more rigorous review, but the platform does not inherently guarantee it.
The patent describes a more traceable system
Axon’s patent portfolio makes the provenance question more interesting.
A pending application, WO2025106596A1, “Automatically generating a report using audio data”, maps unusually closely to several documented Draft One behaviors.
The patent describes a generative model receiving a transcript and report-generation instructions, generating a report, inserting verification information, and potentially preventing transmission or storage until required review occurs.
That resembles Draft One’s current Obvious Errors safeguard, which Bend’s SOP confirms is enabled locally. The patent also describes something more granular than anything clearly documented in the current commercial product.
It contemplates source indicators linking portions of generated report text to corresponding portions of the transcript and underlying audio—potentially at the level of individual words, clauses, or sentences.
Conceptually, the patent describes a chain from a generated sentence to the exact transcript segment and corresponding audio time range. Current Draft One documentation establishes evidence-level auditability. It can show which evidence items were selected. Officers can review transcripts beside the draft. Agencies can retain the original AI output.
But the public product documentation reviewed for this article does not establish a feature where a prosecutor, defense attorney, supervisor, or court can click a generated sentence and see the precise transcript passage and audio timestamp that produced it.
A patent is not proof that a capability was commercialized or deployed. But it does show that Axon has contemplated a much finer-grained provenance architecture than the current public Draft One documentation clearly exposes.
Draft One is beginning to learn the agency itself
In September 2026, Axon disclosed another development that pushes the authorship question further.
Its ACEIP law-enforcement transparency portal now lists a feature called Draft One Customized Rules.
Axon says the system learns recurring agency-specific style and formatting preferences from past Draft One drafts and the corresponding finalized reports.
According to Axon’s detailed use-case description, the system identifies recurring patterns in terminology, formatting, and structure, filters them, and uses the resulting agency-level rules to make future drafts better match local reporting conventions.
That creates another feedback loop: an AI draft is edited by an officer, the finalized report becomes part of the pattern the system learns from, and future drafts may increasingly resemble the agency’s historical reports. Officers may spend less time correcting the same formatting conventions, and reports may become more consistent.
The feature also creates a governance question:
If an agency has problematic habits of phrasing, framing, or emphasis, when does a recurring institutional habit become an AI “preference”?
Axon says only recurring generalized rules are retained and describes privacy protections around the process. The current public materials do not establish that Bend uses this feature.
Still, the development shows Draft One evolving from a tool that writes from incident evidence into a system that can also learn how a particular agency prefers to write.
Prosecutors and lawmakers are moving before the courts
Reported appellate law squarely addressing Draft One remains scarce, but institutions are not waiting.
In September 2024, the King County Prosecuting Attorney’s Office told local law-enforcement agencies it would not accept AI-assisted police narratives. The office cited accuracy, preservation, discovery, and officer-credibility concerns.
Some of the preservation concerns were later partly addressed by Axon’s original-draft retention feature, while the credibility issue remains different.
A police report can matter long before a jury ever sees it. Reports inform charging, probable cause, detention, plea negotiations, suppression litigation, discovery, sentencing, and trial preparation.
Andrew Guthrie Ferguson’s 2025 Northwestern University Law Review article, “Generative Suspicion and the Risks of AI-Assisted Police Reports”, argues that AI-assisted reporting changes not only how police write but how legal institutions understand the facts of a case.
California has already legislated directly. California Penal Code § 13663 requires official AI-assisted law-enforcement reports to identify the AI program, disclose AI use, and be signed by the officer as true and correct.
The law also requires retention of the first AI-generated draft for as long as the final report and requires an audit trail identifying the person who used the AI and the source audio or video used to create the report.
California also expressly says the AI-generated draft is not the officer’s statement.
Utah has taken a different approach. Its law requires agencies using generative AI to maintain policies on authorized tools and uses and requires disclosure and author certification for AI-created records, but it does not impose California’s first-draft-retention requirement.
Oregon has no equivalent AI-police-report statute identified in this research.
Oregon’s existing criminal-discovery law broadly includes reports and electronically stored information, but it was not written with intermediate AI-generated police narratives in mind.
That leaves unresolved questions about how an Oregon court would classify the original Draft One output, audit records, or unpreserved officer inputs.
Those questions should not be collapsed into a categorical claim that failure to disclose every AI draft is automatically a Brady violation. An original AI-generated draft could become relevant to criminal discovery or Brady obligations if it contains material exculpatory or impeachment information, but the public law reviewed here does not establish that every Draft One draft must always be disclosed in every case.
The accurate conclusion is not that AI police reports “will not hold up in court,” but that police agencies, prosecutors, and legislatures are creating rules before courts have settled most of the technology-specific questions.
When the officer is no longer the first author
The conventional police-report model is straightforward in theory:
officer observes → officer remembers and reviews evidence → officer writes
Draft One inserts an additional process:
evidence + transcript + structured information + officer context + model instructions → AI narrative → officer reviews and adopts
The officer remains responsible, but responsibility and authorship are not identical. California now makes that distinction explicit: agencies must retain the first AI-generated draft, while the law says that draft is not the officer’s statement. The final signed report occupies a different status because the officer reviews, adopts, and certifies it.
Draft One therefore separates two ideas that traditional report writing often collapses into one: who first composed the narrative and who ultimately assumes responsibility for it.
That distinction matters because police reports serve more than one function.
They are tools for moving cases through the criminal-justice system. They are also records of what an officer observed, remembered, considered important, and chose to put into words.
A generative system can improve grammar, organization, terminology, and consistency.
It can also make uncertainty less visible.
A rough officer-written report may reveal hesitation, incomplete recollection, or awkward sequencing. A language model is designed to produce coherent prose.
Whether that changes how readers perceive the certainty of the underlying event is an empirical question that still deserves study.
There is another reason authorship matters. Axon tells prosecutors that police reports may later be used to refresh an officer’s recollection in court. That creates a question researchers have not yet answered for Draft One: whether reviewing and adopting a machine-generated narrative affects later memory differently from writing the first narrative oneself.
The concern is plausible given what is known generally about memory and suggestion, but the public evidence reviewed here does not establish that Draft One distorts officer recollection. It remains a question for testing rather than a conclusion.
Bend’s cautious experiment
Bend offers a useful place to return because its policy shows what a comparatively cautious Draft One deployment can look like.
The department limits the trial to less serious cases. It bars probationary officers. It prohibits direct Draft One reliance in search-warrant affidavits. It requires disclosure. It enables obvious errors. It creates an audit log. It says the initial AI draft is maintained.
It also expressly limits the system to body-camera transcript input rather than CAD or RMS data.
Those are substantial safeguards, but they do not answer every provenance question. The SOP does not say whether every officer edit is versioned. It does not say whether narrated additional details are preserved. It does not establish sentence-level linkage between generated statements and source audio. And it does not answer how prosecutors or defense attorneys will ultimately receive Draft One materials in discovery.
Bend is useful precisely because it is not a worst-case example: even a cautious policy cannot reconstruct information the underlying system does not preserve.
The report should be auditable, not merely readable
The debate over AI police reports is often framed as a competition between efficiency and error, but that frame is too narrow. Draft One can produce polished prose, officers may find it easier to work with, and agencies can add safeguards; some departments restrict it heavily while others allow it broadly.
The deeper issue is what happens to the evidentiary chain when a generative system becomes the first writer.
A police report can influence whether someone is charged, detained, searched, prosecuted, offered a plea, or believed in court.
If artificial intelligence becomes part of the process that converts evidence into an official narrative, accountability should require more than a human signature at the end. It should require a record of how the narrative came to exist.
Axon is building increasingly sophisticated ways to make Draft One sound more like an agency’s own reports. The corresponding standard for accountability should be just as clear:
If a system can transform evidence into an official police report, the path from evidence to report should be reconstructable.