When the Los Angeles Police Department started feeding decades of arrest records into predictive algorithms, it believed it was upgrading how it approached crime. What it actually did, researchers now argue, was automate its existing prejudices.
The growing use of artificial intelligence in law enforcement is raising alarms among criminologists and legal scholars who say these systems absorb the biases already embedded in historical police data and repeat them at scale. Far from offering a neutral, data-driven view of crime, the tools risk entrenching the same inequalities they claim to transcend.
At the core of the problem is how these systems learn. AI tools for predictive policing are trained on past arrest and crime records, which means they begin with a picture of crime shaped by previous policing decisions rather than by crime itself. Areas that received heavier patrol in earlier decades show more recorded incidents. The algorithm reads that density as a future risk signal and directs officers back to those same neighborhoods. More arrests follow. The data is updated. The cycle repeats.
Sarah Brayne, a sociologist at the University of Texas at Austin, spent time inside the LAPD studying how the department used big data systems. In research published in the *American Sociological Review*, she found the department was increasingly using data to predict crime rather than respond to it. The data flowing into those systems was skewed from the start: crimes in public places were more likely to be recorded, and patrol resources were disproportionately concentrated in Black neighborhoods. Once that history was absorbed by the algorithm, Brayne wrote, "the predictions appear impartial." They were not. The models, she concluded, create "a feedback loop in which they not only predict events such as crime or police contact, but also contribute to their future occurrence."
The problem sharpens when that underlying data was produced under conditions of institutional misconduct. Researchers at New York University's AI Now Institute studied 13 jurisdictions across the US and found that in nine of them, police departments had trained predictive systems on records gathered during documented periods of unlawful or biased policing. Chicago is the starkest case. The city's police department was under federal investigation for discriminatory practices at the same moment it adopted a risk-flagging algorithm to identify people likely to be involved in shootings. The residents federal investigators had identified as targets of biased policing overlapped substantially with those the algorithm flagged. The list eventually grew to more than 400,000 names, 56 percent of them Black men.
Chicago shut down the program in 2019 after its own inspector general raised concerns about its effectiveness. Los Angeles scrapped a similar initiative, Operation LASER, the same year. A 2023 analysis of Geolitica, formerly known as PredPol, one of the most widely deployed predictive policing tools in the country, found a success rate below half a percent across more than 23,000 predictions made in Plainfield, New Jersey.
Not everyone in the field believes the technology is fundamentally broken. Dr. Dimitrios Kollias, an associate professor in multimodal artificial intelligence at Queen Mary University of London, argues that modern AI systems do not simply replay their training data. "They can combine learned information in new ways and generate outputs that were not explicitly present in those examples," he told the BBC. He acknowledges dirty data is a genuine concern but argues systems can be tested for unfair outcomes and redesigned when they fail. As those systems grow more autonomous, he adds, ensuring they remain under meaningful human control becomes a priority in its own right.
The consequences of algorithmic bias run well beyond law enforcement. A 2019 study published in the journal Science examined a healthcare AI tool used to allocate additional patient support across US hospitals. The system measured patient need by predicting the cost of their care. Because less was already being spent on Black patients than on equally sick white patients, often due to poorer access to care, the algorithm systematically directed resources toward white patients. When researchers recalibrated it to predict illness directly rather than cost, the share of Black patients selected for extra support would have risen from roughly 18 percent to 47 percent. The fix was straightforward once someone looked for the problem. That is precisely the case researchers make for independent oversight.
Legislators are beginning to respond. In January 2024, seven US Congress members asked the Justice Department to pause grant funding for predictive policing programs until their discriminatory impact could be assessed. A White House policy from March 2024 requires federal agencies to independently test rights-impacting AI tools, though the mandate does not extend to state and local police departments. In Britain, 17 civil society organizations called the same year for a full ban on predictive policing.
Dr. Carolyn McKay, an associate professor in digital justice at the University of Sydney Law School, says cleaning up biased data is not a matter of removing a few stray errors. The corrupted records are not sitting at the edges of an otherwise sound dataset. They run through the foundation the algorithm was built on. Even systems redesigned to flag locations rather than individual people can produce the same outcomes, she warns, because the neighborhoods they target were shaped by years of concentrated policing long before any algorithm got involved.
The root of the problem, most researchers agree, is not the machine. It is the human judgment baked into the data the machine inherited. The algorithm did not decide what suspicious behavior looked like, or which crimes were worth recording. People did. Until those decisions are independently examined, AI will keep reflecting them back.
