Rekognition's only known law enforcement client does not use “confidence threshold”, undermining Amazon's defense against research showing the software's bias
Faced with two independent studies that found its facial recognition software returns inaccurate or biased results …
Context & Ripple Effects
Amazon has spent months defending Rekognition against two independent studies showing inaccurate or biased results by arguing that accuracy is a matter of configuration — after the ACLU's Rekognition test, the company said it was reasonable for government users to decide how to weigh matches, and AWS chief Andy Jassy repeated in a staff meeting that regulators should specify how the tech is used. The defense treats the confidence threshold as the safeguard.
Gizmodo now reports that Rekognition's only known law enforcement client does not actually use a confidence threshold — meaning the one safeguard Amazon points to is absent in its sole visible policing deployment, and the responsibility framing loses its factual footing.
First-order effects
- Amazon's 'operators set the threshold' defense is directly contradicted by its own customer base: with no known law enforcement user applying the safeguard, blame-shifting to government configuration no longer holds as an argument against the bias findings.
Second-order effects
- Critics and lawmakers gain concrete evidence that default product behavior, not operator misuse, drives real-world deployments — pressure shifts from how governments configure Rekognition to what Amazon ships, and to the low-cost accessibility Forbes documented that lets anyone build facial recognition tooling (for pennies on a laptop).
Third-order effects
- If the pattern holds, scrutiny moves from usage policies to deployment defaults — pushing toward regulation of facial recognition systems themselves rather than trusting buyer-side settings, the governance gap Jassy asked regulators to fill.
The trend: Facial recognition accountability is shifting from operator-responsibility arguments toward scrutiny of vendor defaults and product-level safeguards.