Police AI Creeps Into Case Files With No One Watching

The pitch was paperwork relief. The product on offer was police AI creeping into the legal memory of the state.

At this year’s International Association of Chiefs of Police Technology Conference in Fort Worth, Texas, vendors sold tools billed as “the future of policing in the digital age,” while press access was barred, according to The Verge. The public framing was efficiency. The real fight is accountability. Our view: police AI is moving fastest into the exact places where human judgment should be slow, reviewable, and attributable.

Police AI vendors are automating the parts of policing that need a human name on them

The showroom menu was broad: facial-recognition cameras, automated license plate readers, body cameras, chatbots for non-emergency 911 calls, gunshot detection platforms, drones, and AI report-writing tools. That isn’t a collection of back-office helpers. It’s the skeleton of modern policing, from data gathering to the first written account of an incident.

The industry’s favorite line is familiar from enterprise software: let machines handle the busywork, let humans focus on the important stuff. In policing, that distinction collapses. A report is not busywork. A suspect summary is not busywork. A dispatch narrative is not busywork. These artifacts can later shape what attorneys and judges are able to test.

“A lot of it is sales gimmicks that don’t actually deliver on what the promise is,” Abrem Ayana, a police captain in Brookhaven, Georgia, told The Verge.

That skepticism matters because departments are being asked to trust products before public rules have caught up. The Verge reports that in the absence of federal oversight or industry standards, officials often have to take vendors’ word that tools are safe and work as advertised.

AI-written police reports can harden guesses into the official record

Axon’s Draft One shows the stakes. The tool generates police reports using a modified version of ChatGPT, trained for police-report drafting. Axon has said it is hallucination-free because “The creativity is turned down to zero,” according to Noah Spitzer-Williams, senior principal product manager at Axon’s generative AI division.

That claim deserves resistance. Even leading AI labs have not eliminated hallucinations from advanced models. The Verge cites an incident earlier this year in which Draft One wrote that an officer in Utah had morphed into a frog after audio from The Princess and the Frog played in the background at the scene.

The frog story is absurd. The legal problem isn’t.

A human officer can be cross-examined about what they saw, why they wrote one phrase instead of another, and what they left out. A generated report inserts another actor into that chain, one that cannot sit in a witness chair. If the AI drafts a polished version of events and the officer edits it lightly, the final document can conceal where machine output ended and human memory began.

The Verge reports that Draft One originally did not save an original copy of a submitted report. Spitzer-Williams said in a recorded roundtable that this was “by design” because “the last thing we want to do is create more disclosure headaches for our customers and our attorney’s offices.”

That sentence should chill every city attorney and public defender reading it. Axon later updated Draft One in December to let departments “retain and access the original, unedited AI-generated narrative,” according to Axon spokesperson Victoria Keough. Good. But the fact that the first design treated disclosure as a headache tells us where the market’s instincts point.

Fort Worth showed a police AI market racing ahead of oversight

The business case is obvious. Officers hate paperwork. Axon’s own 2024 study found the average police officer spends 40 percent of a typical shift writing reports. “We didn’t sign up to sit behind a keyboard,” said John Mackey, a patrol sergeant with Colorado’s Avon Police Department, which uses Field Notes, an AI report-writing tool made by Truleo.

Vendors know the pain point. They also know the buying channel.

The Verge describes a market where Axon, Motorola Solutions, and Flock Safety dominate much of the police technology stack, while newer startups crowd into conferences like IACP. In early 2024, Axon acquired surveillance technology company Fusus to launch Axon Fusus, a real-time crime center product. Axon already sold stun guns, body-worn cameras, automated license plate readers, Draft One, police drones through Axon Air, and an AI chatbot.

That vertical stack is the point. Data collection, storage, analysis, reporting, and decision support can all sit under one commercial umbrella.

“The entire game of all of these companies is to become the platform for policing,” said Andrew Guthrie Ferguson, a professor at Georgetown University Law School.

The money is moving with the pitch. Axon launched its AI Era Plan in late 2024, a subscription that gives customers access to current AI tools such as Draft One and future tools. Subscriptions rose 140 percent between the first quarter of last year and the same time this year, according to an earnings call transcript cited by The Verge. Axon’s AI product revenue grew 700 percent year over year.

“We are determined to become the AI company in public safety, and we are well on our way,” Axon President Joshua Isner said.

For XOOMAR readers tracking how AI procurement differs by sector, compare this with $20M Plant AI Bet Sends Applied Computing Into Big Oil. Industrial AI raises hard questions. Police AI raises harder ones because the output can sit inside the justice process. For separate coverage involving police technology risks, see China, India-Linked Hackers Raid Balochistan Police.

Old policing data gets new speed inside RTCCs and surveillance stacks

The new police AI pitch depends on real-time crime centers, or RTCCs. These systems gather streams such as 911 dispatch, CCTV cameras, and license plate scanners, then package summaries for officers before they arrive.

ForceMetrics sells one called Velocity, which “uses AI to turn overwhelming amounts of public safety data into clear, actionable insights,” according to the company’s website cited by The Verge. Co-founder Jason Truppi, a former FBI special agent specializing in cybercrime, argues police departments are drowning in aging systems and fragmented records.

There’s a real operational problem here. By 2019, the NYPD was collecting around two years’ worth of body camera footage every week, according to a Committee on Public Safety hearing transcript cited by The Verge. No human review team can meaningfully process that flood.

But speed changes power.

• Before: RTCCs relied on human analysts to gather, organize, and send information to officers.
• After: Modern systems such as Velocity aim to extract patterns from huge data pools and deliver rapid operational summaries.
• Risk: A flawed lead can travel faster, look more neutral, and become harder to unwind once it enters the workflow.

Nina Loshkajian, a fellow at the New York University Center on Race, Inequality, and the Law, pushed back on the idea that more data solves the legitimacy crisis.

“These algorithmic systems did not prevent violent encounters between police and civilians then, and we shouldn’t be tricked into thinking they’ll make a meaningful difference in the future.”

That is the core critique. Police AI does not start from a clean slate. It is built inside existing enforcement patterns, using existing records, attached to existing surveillance tools. If the underlying inputs reflect unequal attention, the outputs can give that attention a machine-polished justification.

The honest case for police AI still fails without courtroom disclosure

The strongest argument for police AI should be taken seriously. Departments have too much video, too many records, too much paperwork, and too few hours. A narrowly built tool that organizes body camera footage, flags inconsistencies for human review, or drafts low-stakes administrative summaries could be defensible.

But the line has to be bright.

If an AI system influences a criminal case, defendants must know it was used. Defense lawyers must be able to inspect its role. Courts must know whether a human account was machine-generated, machine-edited, or machine-suggested. Vendor assurances cannot replace adversarial testing.

Brandon Garrett, a professor at Duke University School of Law, put the problem plainly.

“The idea that you’d be making up data — which is what generative models do — to be used in court, is really, really troubling.”

Efficiency cannot outrank due process. Public power requires public proof.

Cities should pause high-risk police AI until the public can inspect it

City councils, county boards, and police oversight bodies should require public hearings before departments buy or expand high-risk police AI. That means no quiet expansion through multiyear contracts, free trial periods, or sole-source procurement agreements that let vendors keep selling new products without competing bids.

The minimum safeguards are not complicated:

• Disclosure: Criminal defendants must be told when AI shaped a report, lead, summary, or narrative.
• Auditability: Independent reviewers need access to test accuracy claims and failure modes.
• Retention: Original AI outputs should be preserved when they touch a case.
• Bias testing: Departments should publish testing standards before deployment.
• Human responsibility: Sign-off must carry real accountability, not rubber-stamp approval.
• Forbidden uses: Cities should ban AI-generated probable cause, secret risk scores, and tools that courts or defendants cannot examine.

The public should not discover that AI changed policing through a courtroom surprise. If departments want machines inside the justice process, they need to answer first to the people those machines may judge.

Impact Analysis

• AI is entering core policing functions that can affect investigations, prosecutions, and civil rights.
• Lack of federal oversight leaves police departments and the public dependent on vendor assurances.
• Automating reports and evidence narratives could make errors harder to trace, challenge, or correct.