NewsNouvelles techniques

Hundreds of Thousands of Police Reports Now Start as Axon Draft One Text While California Forces Agencies to Keep the First Machine Draft

A police report used to be the one document a defense lawyer could treat as a human being’s memory of a bad hour. That assumption is now a product setting. Axon’s Draft One listens to body-worn camera audio and writes a first-person narrative the officer can edit and sign. Forbes reported on July 22, 2026, that the company was telling customers the tool, together with the rest of its report-writing suite, had already helped produce more than 600,000 police reports across about 600 departments, with a claimed 300,000 hours saved. The hour figure comes from customer surveys, not from a stopwatch a court can cross-examine. The report count is the number that should worry anyone who still thinks this is a pilot.

The AEGIS Alliance is not interested in whether officers dislike typing. They do. The fight is what happens to the first paragraph the model offered, and what happens when that paragraph is wrong. California tried to answer the first question in statute. Public records have started to answer the second.

The sentence the statute forces onto the page

Governor Gavin Newsom signed Senate Bill 524 on October 10, 2025. It took effect January 1, 2026, as Penal Code section 13663. When a covered agency uses artificial intelligence to draft an official report, in whole or in part, the report has to identify every AI program used. It has to state, prominently, « This report was written either fully or in part using artificial intelligence. » The officer has to sign, on paper or electronically, that the facts were reviewed and are accurate. The agency has to keep the first machine draft for as long as it keeps the official report, and it has to keep an audit trail that ties the draft to the person who ran the tool and the camera file that fed it. Vendors are barred from selling or sharing the incident data an agency poured into the model.

The Los Angeles County Sheriff’s Department put the same rules in a December 15, 2025, bulletin and added the point defense lawyers will quote: an AI draft, other than the final signed report, is not the officer’s statement. That distinction matters under Brady. A sentence the model invented and the officer deleted is not a private scratch pad. It is a version of the government’s story.

California’s Commission on Peace Officer Standards and Training followed with an Artificial Intelligence Guidelines document that walks agencies through section 13663 alongside procurement, training, and oversight. The guidelines treat errors in machine prose as a reason for disclosure, not as a reason to hide the draft. Utah’s SB 180, effective in May 2025, already required disclosure and an officer’s certification. California went further and demanded the first draft survive. Prosecutors in Connecticut, and the King County Prosecuting Attorney’s Office in Washington, have imposed their own limits without waiting for a matching statute.

What the filing cabinets already show

Forbes went after the time-savings pitch with emails and reports, not with a press release. In Lafayette, Indiana, the department that helped sell Draft One, officers testing a companion tool called Form One wrote that the software did not save them time and that it introduced inaccuracies they then had to edit. Form One is supposed to fill names, identification numbers, and license plates from the same audio. One Lafayette note said a form that took about 30 seconds by hand took about three minutes once the machine’s mistakes were cleaned up. Another officer wrote that Form One dramatically increased the time it took to finish reports. The department had not committed to keeping it.

Manchester, New Hampshire, the first agency to test Draft One, dropped it in 2024. Lieutenant Matthew Barter told Forbes it was easier for officers to type the report themselves. Anchorage, Alaska, walked away the same year after finding no time savings. Richmond, California, produced a cache of AI-assisted reports that included school calls where no crime occurred, deaths described as natural, and graphic sexual-assault narratives. The problem is not that a model can describe an ugly scene. The problem is that a polished first-person voice can also attach the wrong name, put an officer on a radio call the officer was not on, or smooth a chaotic recording into a confession the tape does not support.

Daniel Clark, a prosecutor in King County, Washington, described that class of mistake in plain trial language. It is the kind of error that challenges an officer’s credibility on the stand. His office has refused AI-generated reports. Ian Adams, a criminology professor who has studied the tools, told Forbes there is a gap between the sales perception and the measured reality. A report writer bolted onto a camera contract turns an accountability device into another pipe, the same problem The AEGIS Alliance has tracked in machines that leak what people type.

The switch that still starts in the off position

Draft One does not watch the video. Axon says so. It transcribes audio, the officer picks an incident type, and a generative model writes the narrative. The company says creativity is turned down so the software will not invent flourishes. Officers are told to fix names and sign. The U.S. Department of Justice COPS Office described the same pipeline in January 2025: audio in, narrative out, human review, then a push into the records system. A camera can show two empty hands while the microphone records a shout. The model never sees the hands. An officer who treats the draft as a finished memory is signing the microphone’s story.

For more than a year after launch, the product did not keep that first draft. The Electronic Frontier Foundation quoted an Axon product manager saying the company did not store the original text, by design, because stored drafts were « disclosure headaches » for agencies and prosecutors. Mother Jones, working from Lafayette emails in August 2025, reported that departments could drop or customize the footer that announced AI use. In December 2025 Axon added a United States setting that lets an agency retain the unedited machine narrative. The company’s own help pages, current through the summer of 2026, say that setting is off by default so an organization does not start storing drafts before it has updated its policy. Outside California, a department can still run the disappearing-draft workflow if its lawyers allow it.

A defense guide updated on September 12, 2026, says the Hillsborough County Sheriff’s Office in Florida has used Draft One since February 26, 2026, and that the practical discovery fight is still the body-camera file, the agency policy, and whatever audit entry shows that a draft was requested. If the first text was never kept, there is nothing to produce. That is not a technicality. It is the design.

The price of a paragraph

Draft One rides contracts Axon already owns: Tasers, cameras, evidence.com storage, records software. Forbes reported an « AI Era » plan listed at $199 per officer per month, and about $750 million in potential multi-year AI contract value booked in the year after Draft One launched in late 2024. In first-quarter 2026 results, Axon said revenue from AI products grew more than 700 percent year over year. A June 2026 College Station, Texas, council packet showed a negotiated quote well below that list price for a small bundle of licenses. The discount does not include the cost of a plea bargain built on a sentence nobody can compare to the first draft.

Truleo and smaller vendors are selling their own report writers. Draft One is simply the one clipped to the cameras already on the uniform. Readers who follow the Police Accountability desk know how a department’s paperwork becomes the official past. A tool used on hundreds of thousands of reports is a tool used on people who will never see a jury.

What a signed page still has to survive

No widely reported trial has thrown out a Draft One narrative as junk science. That silence is not a blessing. Most cases plead. The first serious test looks like discovery, not a seminar on model weights. Did the agency keep the first draft? Did every page carry the California sentence? Did the officer listen to the audio again before signing? Did the narrative name a witness the tape never did? Those questions are ordinary. They became hard because the vendor spent a year making the answers optional.

Officers should not spend a shift recreating a form a clerk could design. The public should still be able to set the computer’s first paragraph next to the paragraph a sworn officer signed. California now requires that comparison. POST has written the compliance manual. Axon sells the switch. The switch still ships off. Departments that leave it off are deciding that hundreds of thousands of reports do not need an author anyone can audit.

The AEGIS Alliance will keep this file on the Technology desk and inside US News, next to other cases where official paper outran the facts, including the Atlanta verdict that priced a bad use of force at $21 million. Primary documents are Axon’s Draft One page, les retention settings guide, les EFF account of SB 524, les text of Penal Code section 13663, POST’s Artificial Intelligence Guidelineset les Forbes public-records review.

Jeffrey Childers
Journalist, editor, cybersecurity and computer science expert, social media management, roofing contractor.

Articles similaires

Laisser un commentaire

Votre adresse e-mail ne sera pas publiée. Les champs obligatoires sont indiqués avec *

Bouton retour en haut de la page