The order
On July 7, D.C. Superior Court Judge Neal Kravitz ordered prosecutors to disclose, to the court and to the defense, exactly how facial-recognition technology was used to identify Marquis Foster, 43, arrested July 1 in the June 8 shooting near Howard University that left one man wounded in the chest, arm, and hip (DC Witness). Investigators ran crime-scene footage through Clearview AI's roughly 70-billion-image database, then cross-referenced the resulting candidate against MPD's own mugshot records before an eyewitness lineup (Biometric Update). Foster's attorney, Elizabeth Weller, told the court she'd "never heard of" that two-step matching workflow before. The parties reconvene July 22.
The order itself is narrow — it compels disclosure, not suppression. But it lands squarely on a jurisdiction with no rules at all: the D.C. Council has never passed legislation governing police use of facial recognition, and MPD has no published policy specific to third-party tools like Clearview. The department's own citywide camera network is explicitly barred from running face recognition — MPD's public CCTV policy states flatly that "the system does not use face-recognition or other biometric technologies" (MPD) — yet that same restriction apparently does not extend to detectives feeding still images into a private vendor's database. That gap, not the existence of the tool, is what Kravitz's order is really surfacing.
A national pattern, not a one-off
D.C. isn't unusually reckless here; it's just unusually exposed right now. Two weeks earlier, the New Jersey Supreme Court ruled unanimously in State v. Miles that prosecutors must disclose the name, manufacturer, and publicly available error rates of any facial-recognition software used in an investigation — even when the state never intends to introduce the match as trial evidence — while stopping short of forcing vendors to hand over source code absent a "particularized need" (Reason). That's a workable template: disclose the workflow and its known error characteristics, protect the trade secrets, and let judges weigh reliability case by case. Kravitz's order tracks the same logic, and D.C., lacking any statute, is essentially building the rule from the bench.
The case for caution — stated fairly
The strongest argument for tight controls isn't hypothetical. Facial-recognition misidentification has already produced real wrongful arrests, and Clearview specifically has been at the center of them. Randal Reid, a Georgia resident who had never set foot in Louisiana, spent nearly a week in jail in 2022 after a Jefferson Parish detective relied on a Clearview match with no corroborating evidence; the parish settled with Reid for $200,000 in May 2026 (Biometric Update). The ACLU has documented at least seven such wrongful arrests nationwide, nearly all involving Black individuals (ACLU) — consistent with NIST's own testing, which found accuracy varies meaningfully by demographic group across nearly 200 commercial algorithms (NIST). A technology with that error profile, deployed with zero statutory guardrails in the nation's capital, is a legitimate cause for judicial skepticism.
Why a ban would be the wrong fix
But the Reid and Detroit cases share a common thread that argues for disclosure-and-corroboration rules, not prohibition: in every documented wrongful arrest, facial recognition was used as the sole basis for a warrant, with no independent verification. That's a process failure, not proof the underlying tool is useless. In the Foster case, by contrast, the Clearview lead was followed by a cross-reference against MPD's own booking records and, reportedly, a victim identification — the kind of layered verification that critics of unsupervised FRT use have long demanded. Detroit's 2024 settlement in the Robert Williams case moved in exactly this direction, barring arrests built on an FRT match alone and requiring corroborating evidence first — without banning the technology itself.
Facial recognition is a lead-generation tool, not a verdict. Courts should regulate it as an investigative input subject to disclosure and corroboration — not treat every use as inherently suspect.
A categorical ban forfeits real public-safety value — a violent-crime suspect identified and off the street within three weeks of a shooting — to guard against a failure mode (uncorroborated, black-box matching) that targeted disclosure rules already address. D.C.'s Council should use the Foster litigation as the occasion to legislate what Kravitz is currently having to improvise from the bench: mandatory disclosure of which FRT vendor was used and its known error rates, a requirement that no arrest rest on a match alone, and periodic accuracy audits along the lines NIST already runs. That would give defendants the transparency Miles and now Foster are fighting for, without discarding a tool that, used correctly, closes cases NIST's own data says human investigators alone often can't.