This week on Weird and Wacky Wednesdays: The Machine Said So
Last week I wrote that the evidence more or less collects itself now, and that the human contribution is mostly redundant. Cameras, gunshot detectors, an ankle bracelet, all telling the same story from four directions.
This week we look at the new ways evidence is collected. Sometimes it collects the wrong person, and then armed officers arrive at that person’s car window working from the assumption that machines do not make mistakes.
Nobody in the stories this week did anything wrong.
A Letter, Two Guns, and a Six Week Old Baby
In February, a Flock camera in Sherwood, Arkansas read the plate on an SUV, got one character wrong, and matched it to a stolen vehicle. Officers stopped the car and ordered the couple inside out at gunpoint. Their six week old baby stayed in the car seat in the back while its parents were handcuffed at the roadside.
It took several minutes to sort out. An officer offered this:
“Those cameras are placed everywhere, and they hit license plates. I’m not gonna say they’re completely perfect, because, you know, that’s modern technology.”
He was being honest. But look at what the sentence gives away. Everyone using these cameras already knows they are unreliable. Knowing that, the standing response to a hit is still to approach a family car with firearms drawn and check afterward.
Some numbers for scale. Roseville, California went back through two years of Flock alerts and found that 71 percent of them involved a misread plate. One camera there could not reliably tell an 8 from a 9, and a single innocent driver got flagged at least six times. The Institute for Justice has now catalogued at least 27 wrongful stops, detentions and arrests traced to these cameras. In close to two thirds, the guns came out before anyone noticed the error.
A Bag of Doritos
Last October a student named Taki Allen was sitting outside Kenwood High School in Baltimore County, Maryland, after football practice, eating chips. The school’s AI gun detection software flagged the bag as a possible firearm.
Police arrived, Allen was ordered to the ground, then handcuffed and searched.
“I was just holding a Doritos bag,” he said afterward. “It was two hands and one finger out, and they said it looked like a gun.”
County officials called for a review of how the school used the system and how a snack ended in handcuffs. I suspect the review will land where every other story here lands, which is that nobody in the chain treated the alert as something a person was supposed to confirm.
One Case I Would Put Before a Judge
Robert Dillon is a 52 year old commercial crabber in Fort Myers, Florida. In August 2024 he was arrested for attempting to lure a child at a fast food restaurant in Jacksonville Beach, a city more than 300 miles away that he says he has never set foot in.
According to the lawsuit he filed in June with the ACLU, a sheriff’s office employee ran grainy surveillance images through facial recognition software. It suggested Dillon. A restaurant employee then picked his photo out of a lineup. On that, police got a warrant.
The complaint makes a point about lineups I had not seen put so cleanly. When facial recognition returns a false match, it returns someone who genuinely resembles the suspect. Put that photo into a lineup alongside random fillers and the witness gets handed the one face that already looks right. It ends up in the very lineup police then use to corroborate the software, and the corroboration comes back looking independent.
Police ran a licence plate reader search on Dillon’s vehicle. It came back with no hits anywhere near that restaurant around the time of the offence. That search supported his innocence, and it was not disclosed to the court that issued the warrant.
Charges were eventually dropped and the record wiped. He had pledged the title to his truck to make bond, lost income, and still gets stopped in public about it. Nobody has apologized. The ACLU counts him as one of fifteen people in the United States known to have been wrongfully arrested on a face match.
Briefly, Canada
We are not running Flock networks here yet. Flock says it has no partnerships with Canadian municipalities, police forces or airports, and the RCMP says it is not buying them.
We have similar surveillance, however. The RCMP runs ALPR in several divisions and, by its own description, captures plates indiscriminately at up to 3,000 images an hour.
British Columbia got to this early. In November 2012, Commissioner Elizabeth Denham released Investigation Report F12-04 on the Victoria Police Department’s ALPR program, finding that data on plates producing no match, which is nearly everything these systems capture, had to be deleted immediately rather than kept or passed to the RCMP. Fourteen years old and it identified the right problem on the first try.
What I keep returning to in the Dillon file is the plate search. The surveillance network had already produced the evidence that would have cleared him, and it was not disclosed. Cold comfort but worth remembering is that because these systems record everything, everything includes where you actually were.
If a camera ever flags your plate, be polite, keep your hands visible, do not argue on the shoulder of the road, and call a lawyer. The place to prove the machine was wrong is later, in a room with a judge in it.
See you next week. Try not to look like a firearm.
