Document Category: Facial Recognition

Because each stage of Facial Recognition (FR) carries a risk of error, due process and Brady require disclosure concerning every step (pp. 6–7). The five steps are: (1) the probe photo used, (2) the database selected, (3) any photo editing performed, (4) the algorithmic search, and (5) human review (pp. 6, 9–13). The brief explains that FR has particularly high error rates when applied to people of color, women, elders, and children (p. 8). Low-quality or edited probe photos increase error (pp. 10–11), and many FR databases are skewed by overrepresentation of minorities (pp. 12–13). Algorithms operate as “black boxes” with differing reliability (p. 13), while human review is subject to the same biases as eyewitness identification (p. 13). Because the risk of error varies at each stage and may be exculpatory, defense counsel is entitled to full discovery of the FR process, including the algorithm and analyst, both of whom function as impeachable “witnesses” (pp. 26–27). Defenders can use this brief to argue for comprehensive discovery of FR methods and to frame challenges to the admissibility or reliability of FR-based identifications.

This document was filed in support of a motion to compel discovery of the underlying source code, parameters, error rates, input data, results, reports, analyst, and confidence scores of the Facial Recognition program used in a criminal case. The expert explains the steps involved in a facial recognition search and areas where subjective human decisions — and cognitive bias — are likely to impact the outcome of the search.

Civil rights complaint about problems with facial recognition technology – pgs. 9-12 collect research about errors with when images are of low quality, angles are different, resolution is bad; pgs. 12-15 collect research showing facial recognition algorithms are racially biased; pgs. 15-16 collect data about jurisdictions that have banned use of facial recognition technology because of its flaws