Robert Dillon, 52, a Fort Myers commercial crabber, was arrested at home in August 2024 on a child-luring charge he didn’t commit, in a city he’d never visited. A facial recognition system had flagged him as a 93% match to a suspect caught on surveillance footage 300 miles away.
The system is FACES, the Face Analysis Comparison and Examination System run by the Pinellas County Sheriff’s Office, with over 38.5 million images and access by at least 196 law enforcement agencies. The source image wasn’t a digital export from the surveillance camera; it was a photograph of a McDonald’s computer screen displaying the footage. Screen glare, reduced resolution, color distortion: the image had been degraded twice before the algorithm ran.
Corporal Scott O’Connell had evidence that should have killed the case. A license plate reader search showed neither of Dillon’s vehicles in Duval County during the incident window. Dillon had told O’Connell he’d never been to Jacksonville Beach and described a distinctive scar running from his hairline to his nose. Both facts were omitted from the warrant affidavit.
Here’s what the ACLU press release doesn’t bury: a 93% confidence score measures digital proximity between two mathematical templates, not the probability that two images show the same person. Dillon said it plainly: “Says it’s 93 percent accurate. Far as I’m concerned, it’s 100 percent inaccurate.” He’s one of 15 known people in the US wrongfully arrested on a facial recognition mismatch. If you’re building AI tools for law enforcement workflows, your confidence score documentation is a liability exhibit now.
The lawsuit, filed June 10 in US District Court for the Middle District of Florida, names five defendants including Jacksonville Sheriff T.K. Waters and Pinellas County Sheriff Bob Gualtieri. Charges were dropped after two months. No agency has apologized.
Nathan Zakhary