Online advertising runs on prediction: platforms estimate how likely each person is to respond to each ad, then use those estimates to decide who sees what. Researchers reported in 2019 that this delivery step alone can skew an audience by gender and race, even when the advertiser's targeting settings are neutral. In June 2022 the U.S. Department of Justice settled a Fair Housing Act case against Meta over exactly this issue, and the settlement required a new delivery system with an outside reviewer.
Key takeaways
- Targeting is the advertiser's choice of an eligible audience; delivery is the platform's choice of who within that audience actually sees the ad.
- A 2019 study of real housing and employment ads found significant delivery skew by gender and race with neutral targeting settings.
- The DOJ and Meta settlement was entered June 27, 2022, and ended Meta's Special Ad Audience tool for housing ads.
- Meta's replacement, the Variance Reduction System, adjusts ad delivery using aggregate demographic measurements of who has seen an ad.
- In Canada, the federal privacy commissioner's guidance on behavioral advertising limits use of sensitive information and requires a working opt-out.
How an ad gets chosen: prediction plus an auction
When a page loads, many advertisers compete for one slot. A platform typically ranks them by a value that combines the bid with a predicted outcome, such as the chance of a click. The following is an illustrative example with invented numbers, not any platform's real formula. Advertiser A bids USD 2.00 per click and the model predicts a 1 percent click chance, for an expected value of USD 0.02 per view. Advertiser B bids USD 0.50 with a predicted 5 percent chance, for USD 0.025. B wins, although A bid four times as much.
The prediction step is where learned patterns about users enter. If past clicks differ by group for reasons unrelated to what an advertiser wants, the system can learn those differences and act on them.
Targeting versus delivery
An advertiser can limit an audience by location, interests, or age. Within the allowed audience, the platform still chooses who is shown the ad. Muhammad Ali and five co-authors studied this second step in a 2019 paper on discrimination through optimization. They found that Facebook's delivery skewed along demographic lines, tied to optimization for market and financial effects and the platform's predictions of ad relevance. Both the advertiser's budget and the content of the ad contributed. For real employment and housing ads, they observed significant skew by gender and race even with neutral targeting parameters. Their argument was that regulators and platforms should examine delivery optimization, not only advertiser targeting.
The Meta settlement
The Department of Justice alleged that Meta's ad system violated the Fair Housing Act, including through algorithms that used protected characteristics to decide which users saw housing ads. A proposed settlement was announced June 21, 2022 in the Southern District of New York and entered June 27, 2022. Reports of the terms say Meta paid the maximum civil penalty available under the Act, USD 115,054, stopped offering its Special Ad Audience tool for housing ads, and would not provide housing targeting options that describe protected characteristics.
The remedy for delivery was a new system, called the Variance Reduction System (VRS). Meta describes it as an offline reinforcement learning framework. After a housing, employment, or credit ad has been shown to a large enough group, the system compares the age, gender, and estimated race or ethnicity of people who saw it with those of the eligible audience. If the two differ, it changes a pacing multiplier that raises or lowers the ad's chance of winning later auctions. It receives only aggregate measurements, and race and ethnicity are estimated with a privacy-enhanced version of Bayesian Improved Surname Geocoding.
What the agreement required, as reported
| Element | What was reported |
|---|---|
| Scope | All housing ads on Meta platforms; Meta said it would extend the system to employment and credit ads |
| Measure | Gap between the eligible audience and the audience that saw the ad, by sex and estimated race or ethnicity |
| Target | Variance at or below 10 percent for 91.7 percent of housing ads by sex and 81.0 percent by estimated race or ethnicity |
| Verification | An independent reviewer, Guidehouse, selected by the parties |
| Oversight | Court oversight and regular compliance review through June 27, 2026 |
The scheduled oversight window has now passed. No closing report or court filing on the final compliance results was located for this article, so how the targets were met is not stated here.
Canada: privacy rules for behavioral advertising
Canadian rules approach the same industry through personal information. The Office of the Privacy Commissioner of Canada publishes guidelines on privacy and online behavioral advertising, last modified August 11, 2025 according to the page. They say opt-out consent can be reasonable when people are told the purposes clearly at or before collection, can opt out easily, the opt-out takes effect immediately and persists, the data is limited as far as practicable to non-sensitive information, and it is destroyed or de-identified quickly. Health and financial data are listed as generally sensitive. Tracking technologies that users cannot decline, such as device fingerprinting, should not be used, and children should not be tracked.
Questions to ask about an AI advertising claim
- What does the model predict, and what outcome is it optimized for?
- Does the claim describe targeting, delivery, or both?
- Who measured the result, and was it compared with a baseline?
- Are housing, job, or credit ads handled under special rules?
- What personal data feeds the model, and can a person opt out?
This article gives general information, not legal advice. Questions about a specific campaign or compliance duty belong with a qualified lawyer or the regulators' own pages.
FAQ
Is the platform or the advertiser responsible for who sees an ad?
Both make choices. The advertiser sets the eligible audience, and the platform's delivery system picks who inside it is shown the ad. The 2019 study and the Meta case concern the second step.
What did the Meta settlement change?
It ended the Special Ad Audience tool for housing ads, limited housing targeting options, and required a delivery system, the VRS, checked by an independent reviewer.
Does the VRS use individual race data?
According to Meta's description, it works from aggregate measurements, and race and ethnicity are estimated with a privacy-enhanced method rather than collected for each person.
Does the Meta settlement apply in Canada?
No. It is a U.S. federal case under the Fair Housing Act. Canadian advertisers and platforms are covered by privacy law and other statutes, and the privacy commissioner's guidance is one reference point.
Related reading
- Whether AI makes people lazy
- How conversational interfaces developed
- AI in letter writing
- Alternatives to AI
More in AI Basics.
Last verified: October 2026
Sources
- Ali M et al., Discrimination through optimization: How Facebook's ad delivery can lead to skewed outcomes, Proceedings of the ACM on Human-Computer Interaction, 2019
- Groundbreaking Resolution to Address Discriminatory Delivery of Housing Advertisements (text of the U.S. Department of Justice announcement, January 2023), Benton Institute, 2023
- Meta Platforms settlement summary, National Association of Realtors, 2025
- Advertising fairness: Variance Reduction System, Meta AI, January 9, 2023 (developer description)
- Guidelines on privacy and online behavioral advertising, Office of the Privacy Commissioner of Canada, modified 2025