Not a solution: Meta's new AI system to contain discriminatory ads

This AlgorithmWatch article evaluates Meta's Variance Reduction System (VRS), an automated tool built to address discriminatory delivery of US housing ads under a Department of Justice settlement, based on Meta's compliance report prepared by auditor Guidehouse and interviews with privacy and AI researchers, including sources at the ACLU and Northeastern University.

The review finds the audit relied on spreadsheets supplied by Meta rather than independent access to its systems, and that it omitted key details such as a breakdown of ad impressions and the scale of users with unknown race or gender classifications. It notes Meta's own acknowledgment that its underlying classification model is less accurate for people of color, and cites an ad-delivery example in which a truck-driving ad reached 4,864 men but only 386 women on Facebook in Germany -- illustrating that the VRS addresses only US housing ads, leaving other ad categories and jurisdictions unaddressed.

It concludes that VRS functions as a limited fix rather than systemic reform, and that self-reported compliance auditing, without independent researcher access, risks becoming a form of "audit-washing" -- underscoring the need for enforceable frameworks like the EU's Digital Services Act to support external scrutiny.