AI Wrote the Police Report: Axon's Draft One and the Flock Abortion Search Expose the New Surveillance Chain

The paper trail of one of the most controversial police surveillance cases in recent American history was, in part, written by a machine. On September 2, 404 Media reported that the Johnson County, Texas Sheriff’s Office — the agency that searched Flock’s nationwide network of more than 80,000 license plate cameras for a woman who had self-administered an abortion — used Axon’s Draft One generative AI tool to draft the official incident report about the investigation. The report itself contains the disclosure, plainly stated: “I acknowledge this report was generated using Draft One by Axon.”

The story would be troubling enough if it stopped at the surveillance. But the detail that should stop every policy maker, prosecutor, and defense attorney cold is this: two automated systems now sat in the same investigative chain. One AI-powered system — Flock’s automated license plate reader network — located the woman. Another AI system — Axon’s Draft One — then wrote the official record of how that happened. The machine found the person, and the machine wrote the history.

What Draft One Actually Does

Draft One is Axon’s generative AI report-writing tool, and it is conceptually simple. When an incident concludes, an officer uploads body camera audio. Draft One transcribes it and, within about five minutes, produces a preliminary narrative of the event formatted as a police report. Officers are supposed to review, edit, and digitally sign the draft before it becomes official. According to Axon’s marketing, agencies using the tool report roughly 67% time savings on report writing — hours of paperwork reclaimed per officer per week.

That pitch has landed. Draft One has become one of the fastest-spreading AI tools in American policing, embedded in the same contracts that agencies sign for Axon’s cameras and evidence-management ecosystem. For departments drowning in documentation, the value proposition is obvious. A 2026 study in the Journal of Experimental Criminology examining AI-assisted police reports found no meaningful improvement in report quality over human-written ones — but quality was never the selling point. Speed was.

The problem is what a police report is. It is not a memo. It is the foundational document of the criminal legal system: the version of events that prosecutors rely on to file charges, that defense attorneys cross-examine against, that judges and juries weigh, and that journalists and civil-rights groups request under public records laws. When that document is machine-drafted from a machine-collected evidence base, every downstream actor inherits the machine’s framing.

The Johnson County Chain, Reconstructed

The underlying case first made headlines in May 2025, when 404 Media revealed that Johnson County deputies had queried Flock’s nationwide camera network — cameras that sit in thousands of neighborhoods across the country — for the whereabouts of a woman who had taken abortion pills. The Electronic Frontier Foundation later obtained records showing the entire investigation was initiated at the behest of the woman’s abusive partner. Deputies discussed charging her and, after consulting the District Attorney’s office, learned the state could not statutorily charge her for taking the pill.

The newly obtained AI-generated report, secured by journalist Cameron Probert of the Tri-City Herald and shared with 404 Media, fills in the rest of the chain. Sergeant Damien Bethel used Draft One to write his report on the incident. The AI’s summary of the deputies’ deliberations reads like a chilling compression of the event: deputies “discussed the legal implications of the situation, noting that while the girlfriend self-administered the abortion pill, there were no criminal charges applicable under current Texas law,” and “considered the possibility of a civil lawsuit against the pharmaceutical company that supplied the pill.”

Another section of the report documents that the deputy who ran the Flock searches never visited the woman’s house. He looked her up in a commercial database called TLO to get her addresses and vehicle information, then entered her vehicle into Flock. His justification, in his own words: he worried the reporting partner might have hurt her, and he wanted to “use resources available” to find her and her children. No mention of family concern appears anywhere in the reports — the only person who expressed worry about the woman in the police narratives is the abusive partner himself.

Meanwhile, Flock CEO Garrett Langley has spent months telling law-enforcement customers a different story — that the search was a welfare check driven by family concern, a claim contradicted by every police document in the case. Leaked audio obtained by 404 Media shows he was still repeating it to police customers as recently as August.

Why the AI-Authored Report Matters More Than Either Tool Alone

Individually, each system has already been debated. Flock’s camera network has drawn national criticism for turning routine driving into a searchable database, with documented cases of misuse — including a Wisconsin officer who allegedly used Flock to check whether his ex-girlfriend had visited an abortion clinic. Draft One has drawn criticism for introducing errors that officers must then catch and remove, as Forbes has reported, and for the epistemics of letting a language model draft the state’s official memory.

Combined, they create something qualitatively new. Consider the discovery problem: the Johnson County Sheriff’s Office has refused to release the body camera footage — the raw evidence Draft One was built on — in response to separate public records requests from both 404 Media and the EFF. The public is left with the AI’s summary of audio it cannot hear, describing searches performed by a system it cannot audit. The transcript becomes the truth, and the truth becomes inaccessible.

There is also the accountability laundering problem. When Keven George of the sheriff’s office was asked about Draft One’s role, he emphasized the human-in-the-loop safeguards: officers review, edit, fill in missing details, and digitally sign. “The Deputy has to digitally sign before the narrative can be used,” he said. That is true, and it is also the point. The signature launders the machine’s framing into an officer’s sworn account. If the model omits a hesitation in a deputy’s voice, drops a disputed phrase, or smooths an ambiguous exchange into a clean narrative, the reviewing officer is unlikely to catch what they never noticed in the first place. Studies of AI-assisted drafting consistently show humans inherit the machine’s omissions.

The Regulatory Gap

No federal framework governs AI-drafted police reports. Draft One deployments are governed by a patchwork of internal department policies — when they exist at all. The EFF has published a guide for requesting records about Draft One use specifically because agencies have been reluctant to disclose which reports are machine-drafted, how often officers edit the drafts, and what error rates look like in practice.

The Johnson County case is a preview of the questions courts will soon have to answer routinely: Does a machine-drafted report qualify as an officer’s personal recollection for impeachment purposes? Can a defense attorney cross-examine a transcript? If the underlying audio is withheld but the AI summary is released, what does due process require? None of these questions have settled answers, and every day more reports are drafted by systems that were never designed with the rules of evidence in mind.

What happened in Johnson County is not a hypothetical about future AI risk. It is the present, documented reality: an AI camera network found a woman, and an AI writing tool composed the official story of how. The governor on this system, right now, is a five-minute review by the same officer who used the tools. That is not oversight. That is a signature.

Sources for this article were reported by 404 Media, AI Weekly, the Electronic Frontier Foundation, and Axon’s own product documentation.