Wrong face, five months in jail: how one ai match ruined angela lipps
Angela Lipps was heating chicken nuggets for the four kids she babysits when the knock came. Officers stormed the trailer park home in Tennessee, rifles up, cuffs out. Their algorithm had spoken: the 32-year-old was a fugitive bank robber in North Dakota. She had never set foot there. She had never even flown.
That was 14 February. She saw sunlight again on 26 July, five kilos lighter, jobless, homeless, and still waiting for an apology no one intends to give.
The nine-minute trial inside a patrol car
Fargo detectives fed grainy CCTV stills into a face-matching engine sold by Idemia. One score popped: 0.89, high enough for probable cause in Cass County. A human detective glanced at Lipps’ driver-licence photo, nodded, and signed the warrant. Extradition paperwork moved faster than the software’s own confidence interval. No one requested bank pings, geolocation, or travel records. Why bother? The machine had voted.
Across state lines, a Tennessee judge rubber-stamped ‘no bail’; after all, the docket labelled her a flight risk. The public defender appointed to her only met Lipps after she had spent 18 days inside, sharing a pod with women awaiting murder trial.

Alibi buried in plain sight
It took one subpoena to vaporise the case. Her Chase account showed a debit card swipe at 11:07 a.m. on 3 December—buying menthols at a Murfreesboro Exxon—while, 1,500 km north, a masked woman emptied tellers in West Fargo. Timestamp, GPS, loyalty card: all matched. Prosecutor Tamara Olson filed dismissal papers the same afternoon. Lipps’ release came 37 minutes later, through a sally-port door she barely recognised.
She returned to a repossessed trailer, eviction notice fluttering like a flag of surrender. Her beagle, Roscoe, had been euthanised by the county shelter after the landlord dropped him off. Lipps keeps the collar in a Ziploc bag.

Pattern hidden in plain sight
Her ordeal is not an outlier; it is entry nine on a spreadsheet kept by the Georgetown Privacy Center. Since 2019 at least eight other Americans—mostly Black men—have done jail time because a faceprint scored above an arbitrary threshold. None received compensation. Two are still fighting second-tier charges prosecutors refuse to drop, afraid of admitting fault.
Idemia’s marketing brochures promise ‘99% accuracy under ideal conditions’. Bank cameras are not ideal. Lighting is harsh, angles oblique, lenses smeared. Yet departments pay $250,000 plus annual maintenance for the mirage of mathematical certainty and cheaper line-ups.

Humans are the weakest link
Computer-vision researchers call the phenomenon ‘automation bias’: people trust a read-out more than their own eyes. A 2022 MIT study found reviewers overturn only 2% of high-match alerts, even when told the software errs one in five times on dark-skinned females. The figure drops to zero if the examiner feels time pressure—every detective’s default state.
Fixes exist. European airports demand a second human check and a time-stamped alibi query before border gates open. U.S. police guidelines, where they exist, recommend nothing binding. The FBI’s own handbook allows arrests on ‘computer identification’ alone, a policy written in 2005 and never updated.
Cost of living while innocent
Lipps now works night shifts scrubbing hospital floors; background checks flag the dismissed charges anyway. She owes $4,300 in jail fees—yes, Tennessee bills inmates for room and board—and $1,180 in towing costs. Fargo police spokesperson Shannon Fassett said the department ‘followed standard procedure’ and declined further comment. Idemia did not reply to repeated emails.
She has consulted attorneys about a civil suit but learned qualified immunity shields cops unless ‘no reasonable officer’ would have acted the same. Reasonableness, it turns out, is whatever a squad of peers signs off on.
Meanwhile the algorithm keeps scanning, keeps scoring, keeps smiling its pixelated smile. Somewhere today another detective is glancing at a match, nodding, reaching for handcuffs. The only thing easier than blaming code is trusting it.
