Arrested Because of AI: When 93% Certainty Destroys a Life
A Florida man was wrongfully arrested after a facial recognition system identified him with a 93% match. A story that raises urgent questions about AI's future role in criminal justice.
There is a number that should give us pause: 93%. It is not certainty. It is not proof. And yet, in Florida, it was enough to put handcuffs on an innocent man. Robert Dillon found himself behind bars, charged with one of the most serious offenses imaginable: attempting to lure a young girl away from a McDonald's. The problem, which only came to light afterward, is that Dillon lives nearly 500 kilometers from the scene of the incident and maintains he has never been there in his life.
It all allegedly started with a facial recognition system. Investigators compared a photograph of him against an image pulled from surveillance cameras. But it was not even a direct frame from the footage: it was a photo taken of a screen displaying the security video, with image quality degraded a second time over. The software processed that grainy image and returned a verdict: 93% match. From that point forward, according to the civil lawsuit Dillon has filed against the police department, the investigation followed a single track, never circling back to examine contradictory details.
The signs were there, though. There was no record of his vehicle passing through the area on the relevant dates. His own statement of non-involvement prompted no thorough follow-up. Despite all of this, Dillon was arrested, spent a night in jail, had to post bail, and for months his name remained publicly linked to a devastating accusation. The charges were eventually dropped, but the reputational and personal damage had already been done.
Dillon's story is not an isolated incident. Organizations such as the American Civil Liberties Union and the MIT Media Lab have spent years documenting how facial recognition systems show significantly higher error rates when analyzing the faces of dark-skinned men or low-quality images. The National Institute of Standards and Technology, in its FRVT study published from 2019 onward with subsequent updates, found enormous variation in the performance of different algorithms, with false positive rates that in certain contexts can reach deeply troubling levels.
The core of the matter, however, is not purely technical. It is cultural. When a human investigator makes a mistake, their judgment can be challenged, questioned, and dismantled piece by piece. When an algorithm makes a mistake, the number it produces tends to harden into truth. Ninety-three percent sounds like near-certainty, but that figure does not measure the likelihood that a person is guilty: it measures only how similar two mathematical models are according to that system's internal criteria. It is a subtle but critical distinction, one that often dissolves the moment the data lands on the desk of someone who needs to make a decision.
Looking at the near future, the Dillon case arrives at a moment when the debate over AI use in law enforcement is intensifying. The European Union, through the AI Act that came progressively into force between 2024 and 2026, has classified real-time biometric systems among high-risk applications, imposing strict requirements for human oversight. In the United States, by contrast, regulation remains fragmented, left to individual municipalities and states, producing deeply uneven results.
The future of artificial intelligence in criminal justice will depend largely on whether those who use it can genuinely understand what an algorithm can and cannot say. An indication is not proof. A mathematical correlation is not a conviction. Robert Dillon knows this firsthand. And his story, told through the lawsuit filed against the police, may well become one of the landmark cases in the global conversation about how AI enters courtrooms and jail cells alike.
Sources: Civil lawsuit filed by Robert Dillon against Florida law enforcement authorities; NIST FRVT Report; European Union AI Act (EU Regulation 2024/1689); ACLU report on the use of facial recognition in law enforcement.