Thread by @math_rachel

Saving Face: Investigating the Ethical Concerns of Facial Recognition Auditing, from AIES 2020, by @rajiinio @timnitGebru @mmitchell_ai @jovialjoy Joonseok Lee @cephaloponderer

🧵with some quotes, but I recommend reading the full paper: arxiv.org/abs/2001.00964 1/

External algorithmic audits only incentivize companies to address performance disparities on the tasks they were publicly audited for

Microsoft & Amazon addressed gender classification disparity after audit, but still had huge performance gap by skin color for age classification

Audits have to be deliberate so as not to normalize tasks that are inherently harmful to certain communities.

Gender classification has harmful effects in both incorrect AND correct classification. Promotes stereotypes and excludes trans & non-binary individuals. 3/

Auditors need to consider the company's procedure/process, not just final performance on a benchmark.

Tax audits that evaluate a company's adherence to a compliance process (and not just submitted financial documents) led to better outcomes. 4/

Tension btwn representation vs privacy. Efforts to increase representation of marginalized communities can:
- lead to tokenism & exploitation
- compromise privacy
- perpetuate marginalization through population monitoring & targeted violence 5/

CelebSET as a benchmark should not be considered as a reward to game or a goal to strive for, but a very low bar not to be caught tripping over.

The humble goal of the algorithmic audit is to expose blind spots rather than validate performance. 6/