JACKSONVILLE, Fla. – As artificial intelligence becomes a bigger part of police investigations nationwide, critics warn that the technology can still make serious mistakes.
News4JAX obtained arrest video showing a Fort Myers man being taken into custody after facial recognition software wrongly identified him as a child abduction suspect.
Robert Dillon, 52, is now suing multiple law enforcement agencies in federal court, alleging his civil rights were violated after he was arrested and jailed before charges were later dropped.
Body-worn camera video from August 2024 shows a Lee County Sheriff’s Office deputy at Dillon’s Fort Myers home telling him Jacksonville Beach police suspected him of luring or attempting to abduct a child. Police said Dillon was identified through facial recognition software, which they described as a match.
Dillon was among more than two dozen people who have been wrongfully arrested after being identified through facial recognition technology since 2019, according to his attorneys. In Dillon’s case, the criminal charges were eventually dropped after further investigation determined police had arrested the wrong man.
Dillon has filed a federal civil rights lawsuit against the Jacksonville Beach Police Department, the Jacksonville Sheriff’s Office — which his lawyers say ran the facial recognition search — and the Pinellas County Sheriff’s Office, which also houses AI technology.
Police said the facial recognition system returned what they described as a 93% match.
Ann Liebschutz, a government affairs specialist, said AI tools have improved rapidly in recent years. Still, she said mistaken arrests can be avoided when investigators apply proper standards before seeking an arrest warrant.
“The professional duty of care that needs to be exercised when procuring an arrest warrant can eliminate any of these challenges we have with the technology identifying the wrong person,” Liebschutz said.
She said she expects fewer mistaken arrests as agencies use more diligence. But Dillon’s attorney, ACLU’s Nate Freed Wessler, said facial recognition remains unreliable in real-world policing and that similarity scores can be misunderstood.
“The technology over time has gotten better in test conditions, but that doesn’t reflect how it’s actually used in the field in real life,” Freed Wessler said. He said the 93% figure is “meaningless” and not the same as saying it is 93% likely the person is the suspect.
Liebschutz also said law enforcement agencies are increasingly using social media and online images as potential offender databases, meaning the more photos and information people post publicly, the stronger the data set can become for facial recognition matching.