Detroit paid $300,000 to Robert Williams after faulty facial recognition matched his driver's license photo to a 2018 shoplifting suspect and police held him about 30 hours. The settlement bought enforceable limits on face matches, exposed why a bare match cannot prove probable cause at trial, and shows why appeals often focus on immunity. Facial recognition in this context means software that ranks a probe photo against stored images and returns possible leads. A lead is not an identification, and compensation cases turn on what police did next.
Official resources:
- Read the official guidance from Nist — Use this primary source to verify the official guidance.
- Read the official guidance from Ojp — Use this primary source to verify the official guidance.
Table of Contents
- How did a bad match become an arrest?
- What did the $300,000 settlement actually require?
- Why does a bare face match fail at trial?
- What evidence fights happen before and during trial?
- What can be appealed, and when?
How did a bad match become an arrest?
Detroit police arrested Williams at home in January 2020 after the system tied him to a Shinola shoplifting suspect from 2018. He was detained for about 30 hours before release, as the Associated Press reported in the Associated Press settlement report.
The risk was not abstract. NIST found many systems had 10- to 100-times higher false-positive rates for Asian and African American faces than for white faces. Its 2019 vendor test tied the worst errors to groups most exposed to false criminal matches.
What did the $300,000 settlement actually require?
Detroit settled on June 28, 2024, for $300,000 with reforms negotiated by the ACLU and University of Michigan Civil Rights Litigation Initiative. The money compensated Williams, while the policy terms controlled future arrests.
The enforceable terms bar Detroit arrests based solely on facial-recognition results, according to the ACLU of Michigan in the ACLU of Michigan policy summary. They also bar photo lineups based solely on a facial-recognition lead without independent reliable evidence, under four years of federal-court jurisdiction. For a possible claimant, the file should preserve:.
- the face-search report, candidate score, and source photo
- the independent evidence cited for arrest or lineup
- approval, witness, and custody records showing time held
Why does a bare face match fail at trial?
Federal model policy from the Bureau of Justice Assistance treats face-recognition results as advisory investigative leads only. The template states they do not alone establish probable cause or positive identification without further investigation, as explained in the federal model face-recognition policy.
That distinction shapes charging, suppression, and damages. Prosecutors still need conduct, witnesses, video, records, or other corroboration linking the defendant to the act. Civil plaintiffs can use absence of that corroboration to challenge the arrest decision.
What evidence fights happen before and during trial?
Defense access to system output is limited. A Florida appeal ruled a defendant was not entitled to photos of other facial-recognition candidates because he did not show a reasonable probability the trial result would change.
Plaintiff discovery can be broader against a city. In the Reno casino face-ID case, the court barred adding a wrongful-arrest claim against the officer on immunity grounds but held the city has no qualified-immunity defense. That ruling opened city-wide custom, training, and thousands-of-arrests discovery.
What can be appealed, and when?
Denial of qualified immunity permits an immediate interlocutory appeal only on legal questions, while fact disputes remain for trial. Municipal Monell liability receives no such immunity appeal, according to Touro Law Review and Suffolk University analysis in the qualified-immunity appeal analysis.
The practical effect is split tracks. An officer may seek early appellate review on whether the law was clearly established. A city must generally defend policy, training, and pattern evidence at trial, while Detroit facial-recognition use fell 91% after approvals tightened.