The Future of Policing on Display
I stood before a massive glass and brick building in the heart of Fort Worth, Texas. Thousands of people had gathered inside to see what was being called "the future of policing in the digital age." As a member of the press, I was not allowed to enter. But I met with attendees at nearby locations who told me what was being sold inside. And I learned that AI is threatening to take over the very core of policing in America.
The promise of AI at this year's International Association of Chiefs of Police (IACP) Technology Conference was about automating routine parts of the job. These routine parts also happen to be critical steps in the legal process. It is a similar sales pitch to the one that has been broadcast to businesses for years: let the machines handle the busywork so you can focus on more meaningful tasks. But in law enforcement, automating seemingly harmless "busywork" like filling out a police report carefully or reviewing a suspect's case history can have huge consequences on people's lives.
Among the AI products on display at the conference's showroom this May were facial-recognition cameras, automated license plate readers, body cameras, chatbots to handle non-emergency 911 calls, gunshot detection platforms, drones, and report-writing tools. As the country has dealt with law enforcement becoming disconnected from actual human police presence in neighborhoods, the industry is continuing to push automation.
The Rise of Real-Time Crime Centers
The decision-making process itself in police departments is increasingly being handed over to algorithms. A group of tech startups are now selling AI to police as a kind of automated air traffic control system. It is a centralized digital brain that can process the huge amounts of data being collected, often by other surveillance and automation tools sold by those same companies, and help departments allocate resources accordingly. Even police officers are not necessarily happy about these pitches.
"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 me. Without comprehensive federal oversight or industry standards, and because the technology is so new, law enforcement officials like Ayana often have no choice but to trust companies when they say their products are safe and work as advertised.
Police departments have used technology for decades to analyze data and, in theory, make better decisions in the field. In some well-known cases, it has backfired completely. CompStat and PredPol, for example, were two early experiments that tried to reduce human error through supposedly unbiased statistics. Instead, they made the problems they were meant to solve even worse. But while those early experiments failed to bring about a new era of unbiased policing as their supporters had hoped, human beings were still in charge of the most important decisions.
The sales pitch behind this new wave of AI products is that the mistakes of the past happened because there was not enough objective, real-time data. AI can, in theory, help bridge that gap by increasing the amount of public safety data collected and the level of analysis applied to it. Many public safety advocacy groups and legal experts, however, warn that bringing black box algorithms into law enforcement will reduce transparency and accountability at a time when public trust in the police is already dangerously low.
Jason Truppi, a former FBI special agent who specialized in cybercrime, told me that police are drowning in a sea of data. Truppi, wearing a pair of Meta Ray-Ban Smart Glasses, spoke quickly and excitedly in sentences filled with corporate buzzwords. In late 2020, he cofounded ForceMetrics, a software company that offers an "AI-powered decision-assist platform" for public safety agencies. The goal, according to the company's LinkedIn page, is to help agencies "increase operational efficiency and better serve their communities in real time."
All of the record-keeping systems that police departments have been using for the past two decades, from emergency call logs to parole record files to body camera footage databases, have created a burdensome information overload, according to Truppi. "All the systems of record [used by police departments] are essentially antiquated," he told me.
ForceMetrics offers police departments a platform called Velocity. According to the company's website, Velocity "uses AI to turn overwhelming amounts of public safety data into clear, actionable insights." In police-tech industry terms, Velocity is what is known as a real-time crime center, or RTCC. The New York City Police Department first adopted RTCCs over 20 years ago. They are designed to combine police data coming in from multiple sources like 911 dispatch, CCTV cameras, and license-plate scanners to give officers a summary of what to expect when they arrive at a scene. The idea is that the more real-time data you give officers, the less likely they will be to go in "guts and guns," as Truppi puts it. It is a casual term for when things go wrong and people get killed.
In the past, RTCCs were run by human analysts whose job was to collect all the incoming digital data, organize it, and send it to officers on patrol. But as Truppi points out, the growth of new data-collection technologies in policing over the years has made it nearly impossible for any department to keep up with the flood of information. By 2019, the NYPD was collecting about two years' worth of body camera footage every week, according to a transcript of a 2019 Committee on Public Safety hearing. That is too much for even the most dedicated human employee to analyze meaningfully.
Modern RTCCs like Velocity are designed to quickly find patterns in huge amounts of data with the goal of improving situational awareness for cops. According to Truppi, the "unfortunate events" that have so badly damaged Americans' trust in police departments in recent years, especially during the pandemic, can largely be blamed on a lack of what he calls "a data-driven approach" to policing.
Nina Loshkajian, a fellow at the New York University Center on Race, Inequality, and the Law, is skeptical of that claim. "The reality is that police departments had already been using predictive algorithms, which companies touted as data-driven, for years before calls to defund the police revved up in 2020," she told me. "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."
The Big Players and the Gold Rush
Truppi's company is competing with two of the biggest names in the modern police-technology industrial complex: Motorola Solutions and Axon Enterprise. Both companies make their own RTCCs, as well as many of the data-collection and surveillance technologies those RTCCs depend on.
In early 2024, Axon, which was originally called TASER, bought surveillance technology company Fusus to launch a RTCC that was officially branded as Axon Fusus. By that time, Axon was already a well-known seller of stun guns, body-worn cameras, and automated license plate readers. The company also offers a popular AI-powered report-writing tool called Draft One, drones for police departments through a program called Axon Air, and even its own AI chatbot.
Axon and Motorola are part of a very small group of companies competing to effectively control the entire modern police technology stack, from collecting data at crime scenes to the strategic decision-making abilities of AI-powered RTCCs. Police departments today often sign multiyear contracts with these providers. In return, the providers offer free trial periods for new technology, along with what are known as sole-source procurement agreements. These agreements allow them to keep selling new products to departments without having to bid against competing offers from other vendors.
In late 2024, Axon launched its AI Era Plan, a subscription that lets customers pay a flat annual fee to get access both to the company's current AI tools, like Draft One, as well as others it might launch in the future. AI Era Plan subscriptions jumped by 140 percent between the first quarter of last year and the same time this year, according to a transcript of a company earnings call with investors. "We are seeing AI move from early interest to a standard part of how large agencies think about their future technology stack," Axon President Joshua Isner said in that call. "We are determined to become the AI company in public safety, and we are well on our way." According to the transcript, Axon's AI product revenue grew 700 percent year over year.
While bigger companies like Axon, Motorola, and Flock Safety currently dominate the police technology-industrial complex, they are facing growing competition from a wave of newer tech startups that were exhibiting at the IACP tech conference in Texas. "The entire game of all of these companies is to become the platform for policing," says Andrew Guthrie Ferguson, a professor at Georgetown University Law School and the author of multiple books on policing and technology. "We're seeing a gold rush into selling [AI] technology to police with the promise that it will all make their jobs easier and more efficient."
That gold rush has also brought in an influx of outside investors. About one-quarter of attendees on the showroom floor at the conference were from "equity firms looking to invest in the latest tech," according to Amber Schroader, a tech entrepreneur I spoke with in Fort Worth during the event. "That was a surprise."
Stay updated
Get the day's AI and automation news in your inbox. No spam, unsubscribe anytime.
The sales pitch has been working.
The Appeal of AI Report Writing
Draft One and other AI-powered report-writing tools have significant appeal at a time when the average police officer spends 40 percent of a typical shift writing reports, according to a 2024 study conducted by Axon. Many of those reports are for routine incidents like traffic stops and noise complaints. "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-powered report-writing tool made by a company called Truleo. "That wasn't why I became a police officer."
Draft One comes with design features meant to force a degree of human oversight. The system will intentionally leave certain details blank, for example, forcing officers to go in and fill them in manually. The platform is built on a modified version of ChatGPT trained specifically to generate police reports. According to the company, it is hallucination-free. "The creativity is turned down to zero," Noah Spitzer-Williams, senior principal product manager at Axon's generative AI division, has said. That claim should be taken with a very large grain of salt, however, since even frontier labs like OpenAI, Anthropic, and Google have not yet figured out how to completely eliminate hallucination from even their most advanced models. And indeed, in one infamous incident from earlier this year, Draft One wrote that an officer in Utah had turned into a frog after picking up audio from the Disney movie The Princess and the Frog, which had reportedly been playing in the background at the scene.
It is easy to laugh at that incident, but real-world outcomes from AI-written police reports could be deadly serious. When a human officer writes a report, they can be cross-examined in a courtroom to figure out important details like their state of mind at the time, or why they included certain details and omitted others. By definition, it is impossible to subject black box algorithms to the same level of scrutiny.
In the case of Draft One, it was also originally impossible to determine which parts of a report were generated by the AI and which by the human officer once the report had been submitted, save the officer's own memory. That was a feature, not a bug. In a recorded roundtable discussion published online shortly after Draft One was launched in 2024, Spitzer-Williams said the platform "by design" does not save an original copy of a report after it has been submitted, "because [the] last thing we want to do is create more disclosure headaches for our customers and our attorney's offices... it's actually never stored in the cloud at all so you don't have to worry about extra copies, you know, floating around." In other words, if a report generated by Draft One ended up in court and was found to contain erroneous details, there was no way for attorneys or judges to know for certain if those were input by the officer or by AI.
Draft One was updated in December to allow police departments "to retain and access the original, unedited AI-generated narrative," according to Axon spokesperson Victoria Keough. The change was implemented "as [law enforcement] agencies, prosecutors, policymakers, and legislatures have established clearer expectations and requirements for AI-assisted report writing."
Brandon Garrett, a professor at the Duke University School of Law who has studied the implications of AI systems for due process, is wary of the technology. "The idea that you'd be making up data, which is what generative models do, to be used in court, is really, really troubling," he says. "We would never tell a police officer, 'Just be creative and come up with a story about what you saw at the crime scene.' Of course not: They're supposed to objectively record as best as they can and document what they saw at the crime scene. But generative models are designed to create."
Lessons from Predictive Policing
In the wake of the 2008 financial crisis, LA police chief Charlie Beck took inspiration from Wal-Mart and Amazon's personalized shopping algorithms and wrote that police departments should use similar tools to predict crime. Starting in the 2010s, "predictive policing" programs were widely implemented in cities across the country. But far from creating a new era of fairness and justice in policing, the algorithms in many cases had exactly the opposite effect. Since the models had been trained to detect patterns from historic crime data, the biases hidden within that training data were perpetuated under the guise of mathematical objectivity.
PredPol, for example, was based on an algorithm originally used to predict the geographical distributions of earthquake aftershocks. The idea was that the same general principle could be applied to predicting crime: the tighter the correlation between a certain area and a particular criminal pattern, so the thinking went, the higher the likelihood that same pattern will continue into the future. This allowed the AI to identify crime hotspots, which personnel-strapped police departments could focus more attention on.
But PredPol and similar programs failed to account for some key facts. For example, more crimes tend to be reported in poorer neighborhoods, which in many major cities are populated primarily by people of color, leading to a higher police presence and arrest rate than those found in other areas. The algorithm had no way of understanding that the fact that there was a higher crime rate in one neighborhood than in another, more affluent area was largely the product of a complex history of social, political, and racial biases and policies. It just ingested the data it had been given, leading to a more intensive focus on historically over-policed areas: a self-perpetuating cycle.
This was clearly illustrated in 2016, when AI researchers Kristian Lum and William Isaac tested a predictive policing algorithm using historic drug crime data from the Oakland Police Department. The algorithm recommended dispatching police "almost exclusively to lower income, minority neighborhoods," Lum wrote in a follow-up article, even though public health data at the time showed that illegal drug use was widely distributed across the city.
The same pattern emerged wherever predictive policing programs were implemented. "The use of predictive policing systems can make the future look a lot like the past," Ángel Díaz, an associate professor at Loyola Law School, told me. "Because a lot of the data you're pulling is from the world as understood by biased policing practices, the patterns that exist in that data will be drawn out by the computer and might help inform future policing practices." In 2024, four Democratic US senators urged the Department of Justice to halt all future grants to law enforcement agencies for predictive policing programs, citing evidence that such programs "are prone to over-predicting crime rates in Black and Latino neighborhoods while under-predicting crime in white neighborhoods."
Predictive policing has therefore become taboo in the modern police-tech industrial complex, a cautionary tale about conflating statistics with objectivity. PredPol changed its brand name to Geolitica in March of 2021. "We don't use the 'p word' at all," Truppi told me, "because it failed."
The Regulatory Vacuum
Experts say a future of policing based on increasingly fine-grained personal data collection and AI-driven policing is frightening. As the decision-making power of AI within policing grows, so too will the inscrutability of the justice system itself, according to Díaz. "The biggest thing that worries me is that we are rapidly expanding how much data is being collected about all of us," he told me. "The reality is that the more data you have about any given person, the easier it is to reverse engineer a reason to target them; the more data you have about each individual, the easier it is to transform them into the subject of an investigation."
Facing budget cuts and staffing shortages, and bombarded by sales pitches in every direction, police departments are now facing the same kind of pressure as private companies to adopt new AI tools. These tools, they are promised, are free of the flaws found in earlier programs like PredPol and CompStat. And as Brookhaven's Captain Ayana mentioned, all of this is happening inside a regulatory vacuum, with law enforcement leaders left to their own discretion to separate the gimmicks from the legitimately safe and useful tools.
Such transparency is made much more difficult when the data is controlled by private vendors, such as Axon, whose business models rely on maintaining the secrecy of their proprietary AI tools. And if there is one lesson that can be drawn from the broader AI race, it is that the race to dominate market share often comes at the expense of safety. For the moment though, in the absence of any broad governance, police departments are left to their own devices to choose from a growing roster of tech vendors. The decisions they make today will impact how decisions are made within their departments tomorrow.
When I asked Stephen Redfearn, the chief of Colorado's Boulder Police Department, about the future of AI within law enforcement, he told me: "It's going to continue to be kind of a roller coaster for a while, while people get more comfortable with it."
This reporting was supported by a grant from the Tarbell Center for AI Journalism.