AI Bias in Ad Targeting: How Algorithms Discriminate

Algorithmic discrimination in advertising is not a hypothetical that ethicists worry about. It has been measured by researchers, charged by a federal agency and settled in federal court, and the most important finding from that whole paper trail is the one most marketers have never heard: an ad campaign can discriminate even when the advertiser’s targeting choices are completely neutral. The delivery algorithm decides who actually sees the ad, and it makes that decision on its own patterns.

I write about this as a practitioner, not a lawyer - the same seat I write the rest of the ethics guides from. My interest is practical: the automated campaign types I cover in the paid media guides hand more and more of the who-sees-this decision to the platform, which means every advertiser now owns a risk they cannot see in their own settings. This post is what the documented cases actually established, where bias enters a modern campaign, and the audit I would run on any account that touches housing, jobs, credit or anything close to them.

Two layers, and the one everyone forgets

Discrimination can enter an ad system at two layers. The first is targeting - the audience the advertiser chooses. That layer is the one regulators fixed first, because it was the visible one: in March 2019, HUD charged Facebook with violating the Fair Housing Act, alleging the platform let housing advertisers exclude people classified as parents, non-American-born, non-Christian, interested in accessibility or interested in Hispanic culture, and exclude whole neighborhoods on a map. Those options are gone now, on every major platform, for the regulated categories.

The second layer is delivery - which people, inside a perfectly legal audience, the algorithm actually shows the ad to. This is the layer that survives every targeting reform, and the evidence for it is the study I would make every paid media manager read: “Discrimination through Optimization” by Ali, Sapiezynski and colleagues at Northeastern and USC. Running real ads with deliberately neutral targeting, they observed “significant skew in delivery along gender and racial lines for ‘real’ ads for employment and housing opportunities despite neutral targeting parameters” - driven, in their words, by market and financial optimization effects and by the platform’s own predictions about which ads are “relevant” to which groups. The machine looked at historical engagement, learned that certain job ads were mostly clicked by certain demographics, and reproduced the pattern. Nobody chose that. The optimization chose it, which is what optimization is for.

The settlement that made delivery bias official

The delivery layer stopped being an academic finding in June 2022, when the Justice Department sued Meta and settled within a week. The case record states the allegation plainly: Meta’s housing advertising system discriminated based on race, color, religion, sex, disability, familial status and national origin, and its delivery algorithms relied on characteristics the Fair Housing Act protects. The settlement required Meta to kill its “Special Ad Audience” tool - the lookalike variant it had already restricted for regulated categories - to remove targeting options relating to protected characteristics, and to build a Variance Reduction System that forces delivered housing-ad audiences to track the eligible targeted audience on sex and estimated race. Meta paid $115,054, the statutory maximum, and the numbers matter less than the architecture: a court-supervised system, with compliance metrics agreed in January 2023 and a third-party reviewer verifying them since, exists because the delivery layer discriminated by itself.

My takeaway for working marketers is uncomfortable but clarifying: the platforms’ own enforcement now concedes that engagement-optimized delivery reproduces demographic bias unless actively corrected, and the correction currently applies to housing, employment and credit ads in the US. Everywhere else - your SaaS campaign, your ecommerce prospecting - the same optimization runs uncorrected. Usually that is legally fine. Whether it is fine for your brand and your pipeline is a separate question, and it is yours.

Where bias enters a campaign you run today

Lookalike audiences and their seeds. A lookalike is a machine for finding more of what you already have, and it inherits every skew in the seed. Feed it a customer list built by years of biased distribution and it faithfully perpetuates the bias - not through any protected field, but through the thousands of proxies (location, interests, device, behavior) that correlate with protected traits. This is proxy discrimination, and removing the sensitive column does nothing to stop it, because the whole point of the model is to reconstruct patterns from everything else.

Fully automated campaign types. Advantage+ and Performance Max are delivery-layer optimization at maximum trust: you hand over creative and a conversion goal, and the machine decides audience composition entirely. I use and recommend these formats, and I hold both facts at once - they work, and they concentrate exactly the mechanism the Ali study measured, with less visibility than ever into who is being reached.

Optimization goals. Optimizing for cheap engagement skews delivery harder than optimizing for conversions, because engagement predictions lean heaviest on demographic pattern-matching. The study saw skew vary with the ad’s content and the auction’s economics - which groups are cheap to reach shapes who gets reached.

AI-generated creative. The newest entry point. Image models asked for “a CEO” or “a nurse” have well-documented stereotype defaults, and generated ad creative quietly decides who looks included in your ad - which then feeds the engagement patterns the delivery algorithm learns from. Creative review for this sits with the same approval gate as the brand safety system.

Personas built by a model. If your buyer personas came out of a model without real customer evidence, they arrive pre-loaded with the training data’s assumptions about who buys what. Targeting built on invented personas launders stereotype into strategy.

The audit I would run

None of this needs a data scientist. An afternoon and honesty will do.

1. Categorize correctly. If any campaign touches housing, employment or credit - including adjacent offers like “find apartments” content or hiring-brand ads - declare the special ad category. Misclassification to keep better targeting is the one mistake on this list that is a legal problem all by itself.
2. Interrogate the seed lists. For every lookalike or customer-list audience: where did this list come from, and does it under-represent anyone my product actually serves? A skewed seed is a decision you are making, even if you never look at it.
3. Read the delivery demographics you do get. Platforms show age and gender delivery splits for most campaigns. Look quarterly. A software ad delivering 85% to men is a fact worth knowing, whatever you decide to do about it.
4. Review generated creative as a set. One AI image is fine. Fifty AI images in which every professional is the same demographic is a pattern your audience will notice before you do. I review generated creative in batches for exactly this reason.
5. Write decisions down. Who approved the audience, what the seed was, why the exclusions exist. The FTC’s substantiation expectations and the EU AI Act’s documentation habits both point the same direction: the teams that can show their reasoning are the ones that survive scrutiny cheaply.

Why I treat this as more than a compliance topic

The legal exposure concentrates in the regulated categories, but the mechanism does not respect category lines. I treat delivery bias as bad marketing before it is anything else: an algorithm that decided nurses’ job ads go to women is the same algorithm deciding your product is not “relevant” to a segment that would have bought it, and no dashboard will ever flag the customers you never reached. Optimization finds the cheapest version of the pattern it already knows. Markets you have never served look expensive to it forever, which quietly turns yesterday’s customer base into a ceiling.

And the trust cost runs on the same rail as dark patterns: audiences do not distinguish “the algorithm did it” from “the brand did it,” because from the outside there is no difference. The brand shipped what the machine assembled.

FAQ

Can I be liable for discrimination the platform’s algorithm caused? In the regulated categories, advertiser and platform obligations both exist, and the advertiser’s start with correct categorization and honest audiences. Outside those categories, US liability is unlikely for delivery skew alone - but I would not let the legal answer end the review, for the market-ceiling reasons above. For a real answer on a real campaign, ask a lawyer, not a blog.

Do the special ad category restrictions fix the problem? They remove the targeting-layer tools and, on Meta, the Variance Reduction System now constrains delivery for housing and more. “Fix” is strong: the underlying optimization still learns from engagement patterns, and the corrections exist only where regulators forced them.

Is using lookalike audiences unethical? No. A lookalike from a representative seed, outside regulated categories, is ordinary marketing. The ethical work is in the seed audit and in noticing when “people like our customers” has hardened into “only people like our current customers.”

Does GDPR or the EU AI Act ban any of this? GDPR constrains the data feeding the models, and the AI Act’s marketing-relevant duties are mostly transparency, covered in my EU AI Act checklist. Neither is primarily an ad-delivery statute, but both raise the cost of not knowing how your targeting works.

This post is part of the AI marketing ethics guides, next to dark patterns on manipulation and brand safety on the creative side. The campaign types that concentrate the delivery decision are in the paid media guides, and the wider practice starts at the complete guide to AI marketing.