The Deleted Face: Why Facial Recognition Must Erase You First — SmarterArticles
The Deleted Face: Why Facial Recognition Must Erase You First<br>August 3, 2026
On a January afternoon in 2020, Robert Williams pulled into the driveway of his home in Farmington Hills, Michigan, and was arrested on his own front lawn while his wife and two young daughters watched. Detroit police accused him of stealing watches from a Shinola boutique. The case against him was, in essence, a single thing: a grainy frame of in-store surveillance footage that an algorithm had decided looked like the photograph on his driver's licence. He was held for roughly thirty hours in an overcrowded cell, made to sleep on a concrete floor, and questioned over a crime committed by a man he had never met and did not resemble in any way that a human eye, given a moment of honest attention, would have confirmed. When detectives finally laid the surveillance still beside his face, even one of them seemed to concede the obvious. The computer, Williams later recalled being told, must have got it wrong.
It is a story that has, by now, hardened into a parable. Williams was the first person in the United States known to have been wrongfully arrested because of a face recognition match. He would not be the last. Porcha Woodruff, eight months pregnant, was arrested in Detroit in February 2023 for a carjacking and held for around eleven hours, though nothing in the surveillance or witness accounts described a visibly pregnant woman; the photo lineup put before the victim used an eight-year-old mugshot rather than her current driver's licence photograph. Her charges were dismissed. Nijeer Parks spent ten days in a New Jersey jail for a shoplifting and assault he could not have committed, having been thirty miles away making a money transfer at the time. Robert Dillon, a fifty-two-year-old from Fort Myers, Florida, was arrested in August 2024 for allegedly trying to lure a child from a fast-food restaurant in Jacksonville Beach, a city he had never visited, three hundred miles from home, after police ran a grainy image of the suspect through an AI-assisted facial recognition system that returned him at 93 per cent “confidence”. Charged with a third-degree felony, he saw the case dropped more than two months later, once his attorney showed he had been at work. The ACLU and the ACLU of Florida sued on his behalf on 10 June 2026. By the ACLU's tally there are now at least fifteen such cases. Nearly all of the wrongfully arrested were Black.
The familiar way to tell this story is as a tale of error. The system, we say, made a mistake. It misidentified. The accuracy was poor, the dataset unrepresentative, the threshold miscalibrated. Fix the maths, broaden the training data, audit the vendors, and the harm recedes. This is the framing of most policy debate, most journalism, and a good deal of the technical literature. It is also, argues a paper presented at the 2026 ACM Conference on Fairness, Accountability, and Transparency and published in its proceedings, a profound misreading of what these systems actually do.
The paper, which appears under the title “Frankenstein in the Pipeline: Computational Epistemicide in Facial Recognition” and circulates as the arXiv preprint 2606.07628, makes a claim more unsettling than miscalibration. Its author, the Brazilian computer scientist Nina da Hora, takes Mary Shelley's creature not as a parable of unintended consequences but as a description of method: a body disassembled, reassembled from parts, legitimated by the procedure that made it. Facial recognition, da Hora contends, does not merely misidentify people from Black and non-Western communities. It performs something closer to an act of erasure. Through a sequence of ordinary engineering steps, the technology takes the face as a living, relational surface and progressively narrows it to whatever can be held still as data, then measures the residue against a norm that is, in its statistical bones, predominantly white, frontal, and European. To be recognised by such a system, the argument runs, anyone whose face departs from that norm must first be remade in its image. Da Hora gives this process a deliberately heavy name: computational epistemicide. The killing, by computation, of a way of being known.
It is a phrase designed to make you flinch, and it should. But before deciding whether it is overheated, it is worth doing something the policy conversation rarely does. It is worth looking, carefully and without squeamishness, at what actually happens to a face when a machine sets out to recognise it.
The Pipeline That Eats a Face
A modern face recognition system is not a single model that gazes at you and knows your name. It is an assembly line, and like all assembly lines it works by subtraction. At each station, something is removed, normalised, or thrown away, until what remains is a thing that can travel.
The first station is detection. Before a system...