Fairness, Exposure, and User Control

Same targeting, same bidding strategy, same budget, all running at the same time. Everything the advertiser controls, held fixed. A 2019 field study bought real employment and housing ads on Facebook that way. That left the platform's own delivery stage as the only thing free to vary. It varied a great deal.

“In the most extreme cases, our ads for jobs in the lumber industry reach an audience that is 72% white and 90% male, our ads for cashier positions in supermarkets reach an 85% female audience, and our ads for positions in taxi companies reach a 75% Black audience, even though the targeted audience specified by us as an advertiser is identical for all three.” — Muhammad Ali and colleagues, 2019.

Housing ads run under the same conditions split the same way, from audiences over 72% Black users down to as little as 51% Black users.

Now look at where the disparity sits. An advertiser auditing its own campaign settings would find them identical across the three job ads, because they were identical. The split happened after the settings, in a stage the buyer does not write and the recipient never sees. An audit that stops at the targeting parameters inspects the one part of the pipeline that was demonstrably clean.

Protected-attribute removal and decorative controls are two forms of the same substitution: the appearance of a remedy in place of one. Drop the attribute and the proxies stay where they are, while the audit gets harder. Ship a “show less” button that barely alters the policy and the user gets a lever attached to nothing.

The group average has a measured case against it. Gender Shades audited three commercial gender classifiers in 2018. Its abstract: “We evaluate 3 commercial gender classification systems using our dataset and show that darker-skinned females are the most misclassified group (with error rates of up to 34.7%). The maximum error rate for lighter-skinned males is 0.8%.” — Joy Buolamwini and Timnit Gebru, 2018.

34.7% against 0.8% is a gap that appears only when sex and skin type are crossed. Report accuracy by sex alone, or by skin type alone, and it flattens into something a launch review would wave through.

The benchmarks were part of the problem. The two audited there, IJB-A and Adience, were 79.6% and 86.2% lighter-skinned. A system could post a strong aggregate score on them and still fail the group they barely sample.

Ask the same question at every stage, because the answer keeps changing. Data and exposure: who was ever visible, who interacted, and which histories are missing — the point where a benchmark that is 79.6% lighter-skinned decides what "accurate" will mean. Candidate generation: which groups or providers can enter the set at all. Ranking and position: how attention is allocated within the relevant candidates. Outcome and burden: who benefits, who is harmed, and who must do the extra work; in the Facebook study the burden fell on people who never learned that a job ad existed. Control and remedy: whether a person can inspect, correct, opt out, appeal, or choose another mode.

That last stage can be measured against a written standard rather than a preference. The Digital Services Act makes the non-profiled option compulsory: “In addition to the requirements set out in Article 27, providers of very large online platforms and of very large online search engines that use recommender systems shall provide at least one option for each of their recommender systems which is not based on profiling as defined in Article 4, point (4), of Regulation (EU) 2016/679.” — Article 38, Regulation (EU) 2022/2065.

Article 27(3) then settles where the control has to live. The switch between available options must be directly and easily accessible from the interface section where the information is being prioritised. Between them, the two articles give the decorative-control failure something concrete to fail against. A control buried three menus away from the feed it governs does not satisfy Article 27(3). A personalisation-only recommender does not satisfy Article 38 at all.

User fairness can involve relevance, error, opportunity, burden, accessibility, or control. Provider fairness can involve position-weighted exposure relative to relevance, merit, need, or another justified allocation rule.

No fairness metric defines justice by itself. Which target is the right one depends on the surface, on who stands in what relationship to whom, on the legal context the system runs in, on which candidates were eligible at all, and on the harm done by showing something as much as by leaving it out.

The two measured cases above pick out two different targets, and neither would have caught the other. The Facebook study is about who received an ad the advertiser wanted delivered to everyone. Gender Shades is about error rates on a group the benchmark barely contained.

Harms come first in a fairness review, together with the people who carry them: recipients, providers, workers, and communities that never see the interface. Step 2 locates the allocation points — data, eligibility, retrieval, ranking, layout, and feedback. That is what the Facebook design did, holding the advertiser's parameters fixed until only delivery was left.

Step 3, select justified metrics. "Context-justified" is not a placeholder. A regulator has written one down. New York City's rule implementing Local Law 144 of 2021 took effect on 5 July 2023. It defines the "impact ratio" as the selection rate for a category divided by the selection rate of the most selected category. And it refuses to take that as a group average: “Clarifying that the required “impact ratio” must be calculated separately to compare sex categories, race/ethnicity categories, and intersectional categories;” — New York City Department of Consumer and Worker Protection. The intersectional slice is required, not encouraged.

Step 4, test interventions. LinkedIn published a fairness-aware re-ranking framework in 2019 and A/B tested it on a live hiring surface: “Our approach resulted in tremendous improvement in the fairness metrics (nearly three fold increase in the number of search queries with representative results) without affecting the business metrics, which paved the way for deployment to 100% of LinkedIn Recruiter users worldwide.” — abstract of LinkedIn's fairness-aware ranking paper, 2019. Close to a threefold increase in the number of search queries returning representative results. No measured loss on business metrics. Then rollout to 100% of LinkedIn Recruiter users worldwide, across a member base of more than 630 million. That is the shape a tested intervention reports in: an effect size, a stated utility outcome, and a named surface.

Step 5 is remedy — correction, appeal, reset, and monitoring that somebody owns by name.

Report candidate recall, position-weighted exposure, user utility, outcome, complaints, and control effectiveness across relevant groups. Evaluate both absolute levels and changes from the previous policy. When provider exposure is constrained, document the allocation principle and test consumer utility, capacity, quality, and long-term supply. A mathematically balanced ranking can still be socially or commercially unjustified.

Control effectiveness is measurable, and the measurement is unkind. A 2024 study checked whether New York City's audit-and-notice remedy had moved anything: “In this study, 155 student investigators recorded 391 employers’ compliance with LL 144 and the user experience for prospective job applicants. Among these employers, 18 posted audit reports and 13 posted transparency notices.” — Wright and colleagues, 2024. Of the impact ratios that were posted, only nine of 386 fell below the 0.8 four-fifths threshold. So 18 of 391 employers acted, and the audits they did post almost never report a failing number. That is a disclosure regime measured rather than assumed.

The baseline requirement has a worked example too. The experiment ran on Twitter, and it began by holding a control group off the algorithm: “We provide quantitative evidence from a long-running, massive-scale randomized experiment on the Twitter platform that committed a randomized control group including nearly 2M daily active accounts to a reverse-chronological content feed free of algorithmic personalization.” — Huszár and colleagues, PNAS, 2022. Every amplification figure was then read against that held-out feed. The comparison covered 3,634 legislator accounts in seven countries and 6.2 million news articles. In six of the seven countries the mainstream political right was more amplified than the mainstream political left. There was no evidence that far-left or far-right groups were amplified more than moderates. Without the control group, the same exposure counts would have described a state and attributed it to nothing.