The Match That Wasn't
In September 2021, two men carjacked a driver at the Mall of America, robbed a group of women, and used the stolen cards at a nearby Walmart. Bloomington Police ran the surveillance stills through facial recognition software and got a name: Kylese Perryman. Detective Andrew Risdall built the case around that match. Nobody bothered to run a lineup. Nobody called the eyewitnesses. Nobody, apparently, noticed that the actual suspect had tattoos and stood roughly five inches shorter than Perryman, who had neither.
Perryman spent five days in jail, thirty more on home monitoring, and fifty-two total waiting for someone to look at the file. He had timecards from his night shift. He had cell phone location data. He had a home-surveillance video of himself at a family birthday party at the exact hour of the robbery. The Algorithm had a probability score. Guess which one the department trusted first.
The System Worked Exactly as Advertised
This is the part that should bother you more than the arrest itself: the software did not malfunction. Dr. Manjeet Rege, who directs the University of St. Thomas's Center for Applied Artificial Intelligence, told reporters that facial recognition is designed to "generate a lead," not deliver "a positive match." That distinction — lead versus match — is the entire disclaimer buried in the fine print of every vendor pitch deck sold to every police department in the country. It is also, one presumes, the distinction that a detective is supposed to remember exists before he arrests someone off the back of it.
Sources within the Law Enforcement Community note that facial recognition "routinely fails to accurately identify people of color," which is less a controversial finding at this point than a settled fact everyone keeps discovering for the first time, one lawsuit at a time.
Fifty-Two Days, Two Settlements, One Apology-Shaped Object
Perryman's charges were dropped after fifty-two days, once his defense attorney did the lineup, alibi-checking, and evidence-reviewing that the department had skipped the first time around. He sued Hennepin County and the City of Bloomington in 2023. The county settled for $10,000 in 2024 — a figure that works out to a little under $200 a day of a man's life. Bloomington held out two more years before agreeing to its own settlement this week, along with a promise to change how it alerts other agencies about facial-recognition-derived suspects.
Sources confirm the city did not disclose the new settlement amount. And yet — as if this were not enough — the "policy change" amounts to promising to be more careful about the exact process that put an innocent man in jail, rather than promising to stop using a tool that generates leads dressed up as certainties.
What a Lead Is Not
Nobody at Bloomington PD has publicly explained why Detective Risdall treated a software probability score as case-closed evidence, in a development that will surprise no one who has been paying attention to how facial recognition gets used the moment it lands on a real detective's desk. The disclaimer says "lead." The arrest report said otherwise.
Perryman got his settlement, eventually, in the way people eventually get settlements: years later, for less than they lost, after the system that failed them first got to keep deciding the terms.
