# Amazon Rekognition - erroneous facial recognition
Global, 2018-2020 | Severity: N3

## What Happened
In 2018, the American Civil Liberties Union (ACLU) conducted a test of Amazon's Rekognition facial recognition software, matching photos of members of the US Congress against a database of 25,000 public mugshots.
Using the system's default confidence threshold, the test produced 28 false positive matches, incorrectly identifying lawmakers as people who had been arrested.
The false matches disproportionately affected members of color, raising concerns regarding performance disparities, threshold selection, validation, and the use of facial-recognition systems in consequential contexts such as law enforcement.
Under intense public pressure, Amazon implemented a temporary moratorium on police use of Rekognition in 2020.

## How MEG 1 acted
MEG 1 would classify facial recognition used for consequential identification as a high-risk decision-support function.
DAI validation would require testing accuracy across relevant demographic groups, confidence thresholds, operating conditions, and deployment contexts rather than relying on aggregate performance claims.
ISR controls would prohibit a facial-recognition match from independently triggering adverse action.
Any candidate match would require independent human verification and corroborating evidence before it could influence detention, investigation, access denial, or another consequential decision.
Threshold configuration and model limitations would have to be explicit, auditable, and appropriate to the deployment context.

## MEG 2 Analysis - Legal Framework
* **Agent level:** N3 (individualized/advanced) - biometric identification system capable of matching individual faces against reference databases, with potentially significant consequences when used in law-enforcement or other high-impact contexts.
* **Cause (6.1):**
  * **(a) system defect** - insufficiently reliable facial matching for consequential use, including sensitivity to confidence thresholds and documented disparities in false-match outcomes.
  * **(b) automated decision error** - the system produced false identity matches that could create serious adverse consequences if treated as authoritative evidence without independent verification.
* **Attachment of liability:** Under the MEG 2 allocation model, organizations deploying facial-recognition outputs in consequential decisions would bear responsibility for verification, threshold selection, and downstream action. Manufacturer accountability would depend on defects in system performance, validation, documentation, or representations concerning appropriate use.

## Procedural Mechanism
* **6.2 (architectural human confirmation):** A facial-recognition match must never constitute sufficient authority for an adverse or coercive action. Consequential use must require independent human verification and appropriate legal authorization.
* **7.1 (independent forensics):** Searches, confidence thresholds, candidate matches, human verification steps, and subsequent actions must be independently logged so false matches and decision pathways can be reconstructed.
* **9.4 (least privilege):** Facial-recognition output must remain a bounded decision-support signal and must not independently authorize identification, detention, investigation, or other consequential actions.