Actionable Auditing: Investigating the Impact of Publicly Naming Biased Performance Results of Commercial AI Products

Although algorithmic auditing has emerged as a key strategy to expose systematic biases embedded in software platforms,  we struggle to understand the real-world impact of these audits, as scholarship on theimpact of algorithmic  audits on increasing algorithmic  fairness and transparency in commercial systems is nascent.  To  analyze the impact of publicly naming and disclosing performance results of biased AI  systems, we investigate the commercial impact of GenderShades,  the  first algorithmic audit of gender and skintype performance disparities in commercial facial analysis models.  This  paper  1)  outlines  the  audit  design and  structured  disclosure  procedure  used  in  the  Gender Shades study, 2) presents new performance metrics from  targeted  companies  IBM,  Microsoft  and  Megvii(Face++)  on  the  Pilot  Parliaments  Benchmark  (PPB)as of August 2018, 3) provides performance results onPPB by non-target companies Amazon and Kairos and,4) explores differences in company responses as shared through  corporate  communications  that  contextualize differences  in  performance  on  PPB.  Within  7  months of the original audit,  we find that all three targets re-leased new API versions. All targets reduced accuracy disparities between males and females and darker and lighter-skinned subgroups, with the most significant update occurring for the darker-skinned female subgroup, that underwent a 17.7% - 30.4% reduction in error between audit periods. Minimizing these disparities led to a 5.72% to 8.3% reduction in overall error on the Pi-lot Parliaments Benchmark (PPB) for target corporation APIs. The overall performance of non-targets Amazon and Kairos lags significantly behind that of the targets, with error rates of 8.66% and 6.60% overall, and error rates of 31.37% and 22.50% for the darker female sub-group, respectively.