By Stuart Kerr, Technology Correspondent, LiveAIWire

Algorithms police the police in a growing number of cities, but a wave of 2026 wrongful arrest cases shows the same facial recognition systems being pitched as oversight tools are still generating false matches that put innocent people in jail. In June 2026, the American Civil Liberties Union filed suit on behalf of Robert Dillon, a Florida crabber who spent a night in jail after an AI facial recognition system matched grainy surveillance footage to his face.

Dillon lives five hours from the crime scene and had never visited the city where it occurred. He is one of at least 15 people publicly known to have been wrongfully arrested in the United States after police relied on a facial recognition match, according to the ACLU.

The Dillon case lands at an awkward moment for the broader push to use AI as an oversight tool inside policing itself. The same technology being deployed to flag officer misconduct, predict where crime will occur, and audit internal affairs complaints is drawn from the same family of pattern matching systems that just cost an innocent man his freedom for a night and left him fighting a court case a year later.

How Algorithms Police the Police From the Inside

Predictive policing software such as PredPol captured public attention for forecasting where crime might occur, but a parallel and less visible use of the same underlying technology has been redirected inward. In cities including Los Angeles, Chicago, and London, machine learning models have been tested to flag unusual patterns in officer conduct, monitor complaint data, and generate early warning alerts for internal affairs divisions. The pitch is straightforward: a system that never gets tired, never covers for a colleague, and never lets a pattern of complaints go unnoticed for years the way human oversight sometimes has.

These systems do not operate in a vacuum, and that is the problem. They are trained on historical complaint and misconduct data, and research on algorithmic bias in other high-stakes decisions has repeatedly shown that models trained on historically skewed data tend to reproduce that skew rather than correct for it. An internal affairs algorithm trained on decades of underreported or suppressed misconduct complaints inherits the blind spots of the system it is meant to fix.

What Happens When the Oversight Tool Gets It Wrong

The Dillon case shows what that failure looks like in practice, and it is not an isolated one. According to the ACLU’s June 2026 press release, a Jacksonville Sheriff’s Office employee ran grainy surveillance photos through an AI-assisted facial recognition program, which returned Dillon as a possible match. Police built an arrest warrant on that match and a single eyewitness photo lineup, without disclosing that automatic license plate readers showed no record of Dillon’s car anywhere near the scene, and without telling the court that the actual suspect had been described as a regular customer at the restaurant while Dillon had never been to the area in his life.

“The night I spent in jail after they arrested me for a crime I did not commit still haunts me to this day,” Dillon said in the ACLU’s statement, adding that he is still rebuilding his life more than a year later. Nate Freed Wessler, deputy director of the ACLU’s Speech, Privacy, and Technology Project, said Florida police departments “owe it to Mr. Dillon to make amends,” and warned that unreliable facial recognition technology is hurting people nationwide.

The same ACLU release documents a second Florida wrongful arrest tied to the identical statewide facial recognition system, and a separate case in which a North Carolina man spent three months in jail after a Jacksonville Sheriff’s Office misidentification cost him his job, his home, and custody of his children. No law enforcement agency involved has apologized to Dillon or acknowledged the pattern.

Algorithms Police the Police, But Who Polices the Algorithm

The Law Commission of Ontario has published detailed analysis of AI use in Canadian law enforcement, warning that algorithmic oversight tools can replicate existing biases because most misconduct data originates from internal reporting systems that are themselves vulnerable to underreporting. That warning is consistent with the wrongful arrest pattern documented separately by the ACLU. Facial recognition technology in particular has been shown in independent testing to produce higher false match rates for women and people from ethnic minority backgrounds, a disparity that compounds when the same flawed matching process is used to hold police accountable rather than the public.

None of this means algorithmic oversight of policing is worthless. A system that reliably flags a genuine pattern of excessive force complaints faster than a human records clerk could is a real improvement, and few would argue for returning to purely manual internal affairs review. The distinction that matters is between AI used to surface patterns for human investigators to verify, and AI treated as sufficient grounds on its own to arrest someone or discipline an officer. Departments that blur that distinction are repeating the exact error that put Robert Dillon in a jail cell, just aimed inward at their own officers instead of outward at the public.

What This Means for Anyone Living Under Algorithmic Policing

For most people, the practical exposure to this technology comes not through internal affairs software but through the same facial recognition systems now implicated in the Dillon case and at least fourteen other documented wrongful arrests. LiveAIWire’s earlier reporting on facial recognition in UK and US policing found that court rulings and regulators have repeatedly flagged accuracy disparities that fall short of the standard a human identification procedure would be required to meet, yet departments continue deploying the technology with limited independent audit.

The same accountability gap shows up in how AI risk scores are used against defendants rather than police. LiveAIWire’s coverage of AI risk assessment tools used in parole and sentencing documented how the COMPAS system assigned higher risk scores to Black defendants with similar histories to white defendants, a pattern that mirrors the underlying data problem now surfacing in facial recognition oversight tools. Algorithms police the police and the public with the same flawed inputs, and the consequences fall hardest on people who have the least capacity to challenge an algorithmic error in court.

Where Accountability Goes From Here

The Dillon lawsuit is asking for more than damages. It demands specific policy changes from the Florida agencies involved, including a requirement that facial recognition matches be treated strictly as investigative leads rather than grounds for arrest, which is already the written policy of departments including the New York Police Department.

Whether that standard becomes enforceable nationwide, rather than a best practice some departments ignore, will determine whether algorithms policing the police becomes a genuine accountability mechanism or one more system that fails the people with the least power to contest it. For now, the evidence points in one direction: a facial recognition match is a starting point for investigation, not a substitute for it, and every department that treats it otherwise is one grainy photo away from its own Robert Dillon case.

About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, emerging technology, and their impact on business, society, and everyday life. LiveAIWire publishes original AI journalism every weekday at liveaiwire.com.