event · January 2020
Robert Williams is arrested after a false facial-recognition lead
A false identification becomes an arrest, exposing the difference between a machine-generated lead and independent evidence.
Detroit police arrested Robert Williams in January 2020 following an incorrect facial-recognition lead in a watch-theft investigation. The ACLU, which represented Williams, describes a low-quality surveillance image, an erroneous candidate and a subsequent photo lineup. Its case account is an advocate’s reconstruction, rather than a judicial finding.[1]
From a lead to an accusation
The editorial question is whether later review adds independent evidence or merely repeats the initial suggestion. Imagine a spelling checker suggesting the wrong name, followed by a reviewer choosing that same name from a list assembled around it. Two stages need not provide two independent confirmations. This analogy illustrates dependence between checks; it does not reproduce the police investigation.
NIST’s demographic evaluation explains why candidate generation and identification must be distinguished. It does not reconstruct Williams’s case.[4]
The 2024 settlement
The June 28, 2024 order retains enforcement jurisdiction for four years. The agreement requires policy, training and audit changes; its attached directive requires additional independent reliable evidence for an arrest, beyond a facial-recognition lead plus lineup identification (§307.5–5.3). The agreement expressly states it is not an admission of liability. These requirements do not establish that implementation eliminated future errors.[2]
The participant’s perspective
In Sarah Cwiek’s June 2024 Michigan Public interview report, Williams describes police reliance on the technology: “They thought this was some type of magic they were using”. He welcomes safeguards while saying he wanted police use to stop. This is attributed testimony, rather than a benchmark result or proof of remedy effectiveness.[3]
Read the incidents guide and Meaningful Human Control to connect technical errors with human review and responsibility. A person in the decision chain is a starting point for analysis, not evidence that control was meaningful.
Sources
- Williams v. City of Detroit · Source record ref-williams-aclu-case · Back to claim ↑1
- Williams v. City of Detroit — final order and settlement agreement · Source record ref-williams-settlement-2024 · Back to claim ↑1
- “It didn’t make sense at all”: Wrongful facial recognition arrest in Detroit leads to landmark settlement · Source record ref-williams-interview-2024 · Back to claim ↑1
- Face Recognition Vendor Test (FRVT), Part 3: Demographic Effects · Source record ref-nist-frvt-demographics-2019 · Back to claim ↑1
Pages that link here
- AI Incidents: What Failures Actually Establish note
- Face Recognition Vendor Test (FRVT), Part 3: Demographic Effects paper
- Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification paper
- Inioluwa Deborah Raji person
- Joy Buolamwini person
- Robert Williams person
- Social Alignment concept
- Timnit Gebru person
Last updated 2026-10-12