What happened
It was reported that in June 2022 the United States Department of Justice reached a settlement with Meta over the way Facebook delivered housing advertisements, in what the DOJ described as its first case challenging algorithmic discrimination under the Fair Housing Act. The case grew out of a 2019 charge by the Department of Housing and Urban Development, which alleged that Meta’s advertising system violated the Fair Housing Act. The DOJ’s complaint challenged three things: that Meta let advertisers target housing ads using protected characteristics; that its “Lookalike” or “Special Ad Audience” tool used a machine-learning algorithm that could consider characteristics such as race, religion and sex when finding users who resembled an advertiser’s chosen audience; and, most pointedly, that Meta’s delivery system “uses machine-learning algorithms that rely in part” on protected characteristics “to help determine which subset of an advertiser’s targeted audience will actually receive a housing ad.” In other words, according to the DOJ and prior research, even where an advertiser did not set out to discriminate, the algorithm deciding who saw the ad could skew the audience by race, sex and national origin.
Under the settlement, Meta agreed to stop using the Special Ad Audience tool by the end of 2022 and to build a new system, the Variance Reduction System, to reduce disparities in how housing ads are delivered by sex and estimated race or ethnicity. The new system was made subject to the DOJ’s approval, to an independent reviewer, and to court oversight running until June 2026, one of the first times Meta’s ad targeting and delivery has been supervised by a court. Meta also agreed to pay a civil penalty of 115,054 dollars, which the DOJ noted was the maximum available under the Fair Housing Act. In January 2023 the DOJ announced that Meta had built the new system and that the parties had agreed compliance targets. Meta, for its part, described the agreement as the result of more than a year of work to build “a novel use of machine learning technology” that would better match the audience of a housing ad to the population eligible to see it, and said it would extend the approach to employment and credit ads.
What an auditable version would have shown
The whole case came down to one number nobody was tracking in the open: how far an ad’s delivery leaned, by race or sex, away from the audience that should have seen it. The fix the DOJ went for says as much, because what it demanded was, put simply, to start measuring that gap and hold it under agreed limits. An auditable version would have measured it from the start, keeping a running count of who a housing ad actually reached against who was eligible to see it, split by the characteristics the law protects. With that in hand, a delivery system tilting along those lines would show up as a rising number someone was watching, not a problem that only comes out through a federal charge and a lawsuit.
Where the gap was
An algorithm decided which people actually saw a housing ad, and nothing on the public record showed whether it was putting those ads in front of some groups more than others, by race or sex. A MetricRecord would put a number on it: the gap between who received an ad and who was eligible for it, by sex and estimated race, which is more or less what Meta was later ordered to build and report. A ConstraintGate sets the limit that delivery must not skew past, and checks it as the ads go out, so a breach gets caught in-house instead of by a regulator years later. The system was doing what ad optimisation always does, chasing the audience most likely to click. What was missing was anyone measuring, or capping, what that chase was doing to fairness.
What governance should have looked like
When an algorithm decides who sees a chance at housing, a job or credit, the fairness of the distribution is not a side issue, it is the thing the law cares about. A system like that should measure, continuously, whether it is delivering opportunity evenly across the lines the law protects, and should be held to a limit it cannot quietly exceed, with the results kept where a regulator or a court can read them. What this settlement shows is that good intentions and clever optimisation are not enough on their own. The honest test of an ad-delivery system is a number it can show, not a promise it makes.
Failure Pattern: an ad-delivery algorithm allegedly sorted who saw housing ads in ways that tracked protected characteristics, and there was no published measure or constraint on that skew.
Governance Principle: where an algorithm distributes an opportunity like housing, the system should measure whether delivery skews along protected lines and be held to a limit on that skew, on the record.
The reference implementation of MetricRecord and ConstraintGate is open source. It lives at github.com/saffronandindia/headlights-oss, Apache 2.0 licensed and free to install. The repository is public now.
Sources
- Justice Department secures groundbreaking settlement agreement with Meta Platforms (DOJ, 21 June 2022)
- United States Attorney implements groundbreaking settlement with Meta Platforms (DOJ SDNY, 9 January 2023)
- DOJ settles with Facebook over allegedly discriminatory housing ads (CNBC, 21 June 2022)