Artificial intelligence systems deployed in law enforcement risk deepening discrimination against marginalized communities by converting historical enforcement patterns into self-reinforcing feedback loops, according to research published on October 25, 2026. Experts point to predictive policing, facial recognition, and skewed training data as accountability hurdles.
Sarah Brayne and the Los Angeles Police Department Big Data Study
Consider a cartographer handed a list of past crimes and asked to shade where future offenses might occur to guide police patrol routes. More infractions get logged in those shaded zones simply because more officers are watching them. Unshaded areas appear quiet because nobody is looking. When the next map is drawn, those same neighborhoods are shaded even darker in a compounding cycle.
Sarah Brayne, a sociologist at the University of Texas at Austin, spent time inside the Los Angeles Police Department studying its adoption of big data as published in her research paper in the American Sociological Review.
Brayne found that the department increasingly relied on data to anticipate crime rather than merely react to it.
At the same time, the incoming data was skewed: public-place infractions were far more likely to be observed, and police focus landed disproportionately on Black neighborhoods.
Brayne noted that once that skewed input enters an algorithm, the resulting models establish a feedback loop in which they not only predict events such as crime or police contact, but also contribute to their future occurrence
across targeted communities.
Biased Police Records Flaw AI Training Data
Researchers at New York University warned that automated models frequently rely on what they term dirty data
—police records shaped by unlawful, corrupt, or biased practices, including rigged crime statistics and the wrongful arrests of innocent people according to their findings.
Their study revealed that in nine out of 13 US jurisdictions examined, historical police data stemming from periods of unconstitutional or biased enforcement remained accessible to train these algorithms.
An artificial intelligence tool cannot discern whether a historical record is flawed. It treats the information simply as a pattern to repeat.
Analysts point out that technical adjustments offered by commercial vendors fall short unless an independent auditor verifies how the underlying data was generated before deployment.
Dr Dimitrios Kollias, an associate professor in multimodal artificial intelligence at the Queen Mary University of London, argues that computational systems have advanced beyond simple data retrieval.
The old-map analogy is useful for explaining how historical bias can be inherited, but it has its limits. Modern AI systems do not simply retrieve or copy their training examples; they can combine learned information in new ways and generate outputs and solutions that were not explicitly present in those examples.
Dr Dimitrios Kollias, associate professor in multimodal artificial intelligence at the Queen Mary University of London, via BBC
Kollias acknowledges that algorithmic bias stemming from corrupted records remains an important concern. He maintains that unfair outcomes are not predetermined.
At the same time, Kollias cautions that advanced autonomy introduces distinct hazards. As AI systems become more capable of generating novel decisions and acting with greater autonomy, we need to ensure that their behaviour remains safe, reliable and under meaningful human control,
he noted in his interview commentary.
Technical Limitations in AI Lead to Wrongful Arrests
Automated license plate readers, risk assessment algorithms, and facial recognition tools have demonstrated persistent technical limitations. In high-stakes criminal justice settings, these flaws have contributed to wrongful stops, erroneous arrests, and harsher sentencing recommendations across multiple documented implementations.
Researchers emphasize that technical patches alone cannot resolve systemic inequities rooted in decades of discriminatory law enforcement practices, pointing instead toward reforms in how agencies collect data and allocate public resources beyond traditional arrest metrics, while regulatory approaches remain fragmented across jurisdictions.