The Risky Marriage of AI and Police Reports - Governing

The department was testing new artificial intelligence software that uses body-camera audio to generate a first draft of a police report. During a call, the AI picked up a nearby TV playing the movie “The Princess and the Frog” and merged the animated fairy tale into the official narrative. The officer reviewing the draft corrected the error before the report was filed.

It's an easy story to laugh at (and most of the coverage did), but it speaks to the prudence and foresight of Utah state Sen. Stephanie Pitcher. Six months earlier, legislation she sponsored had become the country’s first law regulating the use of AI in law enforcement reporting. The bill (SB 180) requires any police report produced with AI to carry a disclosure notice and the officer to certify the accuracy of the final document.

California soon followed with an even more ambitious law (SB 524) requiring audit trails and retention of any AI-generated police report drafts. The two models — Utah’s light-touch disclosure rule versus California’s comprehensive auditing mandate — offer contrasting templates for other states as they weigh the civil-liberty and public-safety implications of this rapidly changing technology.

As a former prosecutor, Sen. Pitcher has reviewed enough body-camera videos and police reports to know that the two don’t always tell the same story. She had watched body cameras go from an experimental technology to a factor in the majority of criminal cases. If AI hallucinates an incriminating detail and an officer signs off on it, that false information becomes part of the official legal record. “I felt like it was something that it’s better for us to get ahead of, instead of playing catch-up on,” she told the Salt Lake Tribune.

Body cameras have come a long way since their widespread adoption a decade ago. Once pitched as a tool to monitor the police and public in equal measure, the devices are increasingly turning their gaze outward, transforming from standalone recording devices into multipurpose intelligence platforms. In 2024, Axon, the Arizona-based vendor that enjoys a near-monopoly on the body-camera market, released software that produces narratives from body-camera audio using the same class of large language model that powers ChatGPT. A handful of competitors have since developed their own automated reporting tools.

One problem is that human communication is highly nuanced, relying on body language, tone and other contextual signals to convey meaning. For example, profanity can be used in both positive and negative ways. By analyzing a transcript in isolation, AI can struggle to differentiate the two. Furthermore, it can miss important visual details, like a bruise on a victim's arm or the precise layout of a crime scene.

Cognizant of these shortcomings, California lawmakers went further than Utah. SB 524 requires disclaimers on every page of an AI-generated report, plus an exhaustive audit trail identifying who used the AI and what footage it relied on, along with retention of the original AI-generated draft. “We're not going to gamble with personal liberty,” state Sen. Jesse Arreguín said as the governor signed his bill into law. “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.”

The Problem With Inflexible Mandates

Of the two bills, Utah’s approach is narrower and more pragmatic, providing structure without strangling innovation.

Offloading paperwork to AI so humans can spend more time on police work is a worthwhile goal that will require piloting new capabilities, learning what works and sharing those lessons. Inflexible mandates could have a chilling effect as the technology matures, saddling already short-staffed departments with unnecessary compliance costs — a tradeoff California police unions raised about the bill.

Utah's SB 180 sidesteps that trap. By establishing a baseline that allows each agency to adopt its own AI policy, it centers transparency and accountability while allowing individual agencies to work out the operational details. A framework that incentivizes both a rural sheriff's office and a metropolitan police force to experiment is better than forcing each into a one-size-fits-all mandate.

One thing Utah and California both get right is keeping a human in the loop. AI can flag footage, suggest leads and even draft reports, but the decision to arrest someone or file a charge must be done by a person. Even then, automation bias — the tendency to defer to computers — can lull officers into a false sense of security and give investigators tunnel vision. Last year, a Tennessee woman spent five months in jail in North Dakota based on an erroneous facial match despite never having traveled there.

While that investigation did not involve body-camera video, this past December police in Edmonton, Alberta, became the first in the world to pilot body-camera facial recognition that continuously checks passersby against a watchlist of some 7,000 people. At a large concert or sporting event, where terrorism is a serious threat, this kind of live biometric dragnet could save lives. But at a political protest in a public park, the same scene evokes a more troubling vision of state control. Limiting live facial recognition to violent suspects, for instance, or barring it from First Amendment-protected activity, may be necessary to prevent mission creep and protect individual rights.

Balancing Public Safety and Civil Liberties

Too often, we wait until a new technology has been deployed before we think about its implications. In Heber City, the shape-shifting officer was hard to miss. Subtler distortions, such as a mistaken street name or misordered sequence of events, present a more acute risk. If AI condenses a chaotic scene into a cleaner narrative or smooths over uncertainty, it could affect the outcome of a trial.

Prosecutors and courts are not waiting for legislatures to act. In Seattle, the King County Prosecuting Attorney's Office announced that it won’t accept reports written with the help of AI, warning that the technology could compromise the evidentiary record. Guidance from the Prosecutors' Center for Excellence suggests that AI-generated reports should clear the Daubert standard, the reliability test most courts apply to scientific evidence. Clearing that hurdle could prove difficult for vendors reluctant to disclose proprietary technical details.

As body cameras evolve from passive recording devices to nodes in an active surveillance network, the balance between public safety and civil liberties grows more precarious. Giving law enforcement powerful new tools without creating an Orwellian police state will require active, engaged policymaking. The only certainty is that the jurisdictions that act now, before litigation forces the issue, will be in better shape than those playing catch-up.

Logan Seacrest is a resident fellow in criminal justice and civil liberties at the R Street Institute.

Governing's opinion columns reflect the views of their authors and not necessarily those of Governing's editors or management.

Related Articles