Ai facial recognition jailed the wrong grandma for five months

Angela Lipps was spooning cereal into a toddler’s mouth in Tennessee when a SWAT team pointed rifles at her kitchen window. Five months later she walked out of county lockup with no home, no dog, and no apology—collateral damage from a facial-match algorithm Fargo police never bothered to second-guess.

The match that wasn’t

Detectives fed grainy ATM stills into Cognitec FaceVACS, a German-built engine marketed to U.S. sheriffs as “99 % accurate.” The software spat out a single name: Angela Lipps, 42, babysitter, zero priors. A human detective agreed, ticking boxes for “same nose width” and “similar hair part.” Nobody checked mileage logs, cell-site pings, or purchase records. A federal warrant already labeled her a fugitive, so Tennessee judges denied bail. The algorithm’s word became the court’s.

While Lipps sat in a jumpsuit, her court-appointed lawyer, Emily Schaeffer, did what the cops wouldn’t. She subpoenaed timestamped CCTV from a Kroger in Chattanooga showing Lipps buying diapers at the exact moment bandits emptied cash drawers in Fargo. The bank slips were sitting in police evidence lockers the whole time—1,200 miles of asphalt away from the crime scene.

Nine wrongful arrests and counting

Lipps is the ninth known U.S. citizen exonerated after facial-recognition hallucination, according to the Georgetown Center on Privacy & Technology. The common thread: investigators treat the heat-map overlay as gospel, then build a case around the match instead of testing it. Defense teams call the maneuver “AI-washing”—outsourcing probable cause to a black box nobody can cross-examine.

Cognitec’s own validation reports warn of 1-in-1,000 false positives on white female faces under ideal lighting. ATM night footage is far from ideal. Yet departments pay $25,000 per license precisely because sales decks promise to “eliminate human error.” The promise becomes a legal shield: officers swear they “independently corroborated,” but the corroboration is just their own eyeballing of a low-res thumbnail.

Silicon valley’s quiet export

I saw the same pitch in São Paulo last year, where Cognitec reps demoed the tech for Brazilian military police. The slide deck was identical—same accuracy claims, same liability disclaimer in 4-point font. When I asked about false positives, the engineer shrugged: “That’s a courtroom problem.” Translation: once the algorithm ships, the risk is socialized, the profit privatized.

Back in Tennessee, Lipps now sleeps on her sister’s couch. The state has no compensation statute for wrongful algorithmic arrest; she’d have to sue the sheriff personally, a five-year slog with no guarantee. Fargo PD quietly updated its policy manual—officers must now “attempt secondary evidence”—but no disciplinary file was opened. The software license was renewed in March.

Meanwhile, 6,000 U.S. law-enforcement agencies run facial recognition with zero federal oversight. Every day they feed another million mugshots into the cloud, training tomorrow’s model on today’s mistakes. The loop feeds itself: error becomes data, data becomes authority. And somewhere a grandmother is still looking for her dog.