Domain Atlas / Hiring & employment screening AI
Meta Job-Ad Delivery: the guardrail and the layer below
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings comes from a published baseline, not this deployment's own record. Evidence base: 12 published baseline.
Meta Platforms' advertising system decides, for every job ad, which users inside the advertiser's eligible audience actually receive impressions, optimising predicted relevance and engagement against auction economics — so the allocation runs before anyone applies and the people it does not reach generate no application, no rejection and no adverse-action record anywhere. The platform's own published data, as recorded in the December 2022 class charge and the December 2023 joinder release, gives the scale: roughly 239 million US Facebook users, more than 30,000 US job ads published daily, women 54 percent of users interested in job hunting and people 55 and over more than 28 percent of them. The governance record runs on two layers that do not coincide: a restricted advertiser portal created by private settlement in March 2019, and a delivery-optimisation stage that independent peer-reviewed audits measured weeks later.[4]
What happened
A job ad written to reach everyone still has to reach someone in particular, and on a platform that someone is chosen.
Between November 2016 and September 2018, civil-rights organizations, the Communications Workers of America, individual workers and consumers brought five legal actions alleging that Facebook's advertising system let housing, employment and credit advertisers exclude users by race-associated interests, by sex, by age and by narrow geography, including through the Lookalike Audience tool. On 19 March 2019 Facebook settled all five at once. The joint statement it co-signed is a real structural remedy and reads like one: housing, employment and credit ads confined to a separate restricted portal; no gender, age or multicultural-affinity targeting; a fifteen-mile minimum geographic radius and no postal-code targeting; Lookalike Audiences for those ads stripped of gender, age, religion, postal-code and group-membership inputs; detection and rerouting of covered ads created outside the portal; advertiser anti-discrimination certification; testing rights for the plaintiffs and regular implementation meetings; and a commitment to engage researchers on algorithmic bias. Implementation was due by 30 September 2019. Six months later the Equal Employment Opportunity Commission's side of the story landed too: on 25 September 2019 the ACLU announced reasonable-cause determinations against seven employers — Capital One, Edward Jones, Enterprise Holdings, Drive Time Auto, Nebraska Furniture Mart, Renewal by Andersen and Sandhills Publishing — for unlawfully excluding women and older workers from the audiences of their own Facebook job ads, part of roughly 66 charges filed since 2018. Those findings ran against advertisers, over their own targeting choices; they produced no platform remedy, they are administrative rather than judicial, and their conciliations have no public outcomes.
In April 2019, weeks after the settlements took effect, Ali, Sapiezynski, Bogen, Korolova, Mislove and Rieke published the result that reframes the whole record. Buying paired ads with identical neutral targeting on ordinary advertiser accounts, they measured them being delivered to strongly gender- and race-skewed audiences — by the delivery stage itself, through the platform's own relevance predictions and, separately, through budget and market effects, because under a budget constraint the demographics that cost less to reach take more of the impressions. The guardrail the settlements had just installed governed what an advertiser may ask for. The skew they measured arose one layer below, in what the platform does with the request. In 2021 Imana, Korolova and Heidemann closed the obvious objection — that skew might just be the labour market — by running paired ads for the same job at companies with different de facto workforce gender mixes; they confirmed gender skew in job-ad delivery that qualification differences do not justify, and found no comparable skew on LinkedIn.
The one delivery-layer control with enforceable numbers arrived through a different door. On 21 June 2022 the Department of Justice sued Meta under the Fair Housing Act over housing-ad delivery, and on 27 June the court entered a settlement: the Special Ad Audience tool terminated, a $115,054 civil penalty — the statutory maximum at the time — and a Variance Reduction System required, a controller that measures the demographic gap between an ad's eligible audience and the audience its impressions actually reached and adjusts the remaining delivery to shrink it. A compliance-metrics agreement followed on 9 January 2023, with four-monthly reporting, an independent third-party reviewer and court oversight through 27 June 2026. The controller samples eligible-audience demographics, estimates race and ethnicity from surname and geography with differential-privacy noise on the aggregates, and keeps no individual-level race data. Guidehouse Inc. published five verification reports between June 2023 and 30 October 2024. In the last published period, 1 May to 31 August 2024, 95.1 percent of housing ads above a thousand impressions met the ten-percent sex-variance threshold against a 91.7 percent requirement, 85.5 percent met the ten-percent estimated race and ethnicity threshold against an 81.0 percent requirement, and Guidehouse's own recomputation matched Meta's reported coverage at a 0.0 percent difference. Every one of those numbers is about housing ads. Meta reports in its own newsroom having extended the same controller to US employment and credit ads by October 2023, alongside Canadian housing and employment; for that extension there are no published compliance metrics, no employment figures and no third-party verification.
The employment claim sits exactly in that gap, and it sits there as an allegation. On 1 December 2022 Real Women in Trucking filed a class charge of discrimination with the EEOC, No. 570-2023-00655, through Gupta Wessler PLLC and Upturn, Inc., pleading that Meta's delivery algorithm discriminates by sex and age in deciding who receives job ads — disparate treatment, disparate impact, the advertising-discrimination provisions and an employment-agency theory, under Title VII and the Age Discrimination in Employment Act. Its exhibits are Facebook's own ad-library tables for ads whose eligible audiences were all-gender and all adult ages: a commercial-driving ad shown 94 percent to men and 5 percent to women, a mechanic ad above 99 percent to men, a school administrative-assistant ad at 2 percent men and 6 percent users 55 and over, a health-facility hospitality ad at 83 percent women. Those tables exist in public only because the ad library publishes per-ad delivery demographics for ads classified as "Issues, elections or politics" — so the evidence for the charge is an artefact of a classification that has nothing to do with employment. AARP Foundation joined on 19 December 2023 on behalf of older workers, citing a corpus of more than seventy-five postings. Meta's public response is that "Addressing fairness in ads is an industry-wide challenge," citing its collaboration with civil-rights groups, academics and regulators. No tribunal has found Meta's delivery algorithm unlawful in employment.
As of the run date the record is open on both ends. The court's oversight term over the housing settlement expired on 27 June 2026; the Department of Justice case page was last updated on 21 January 2025 and lists no verification report after 30 October 2024; no public post-expiry disposition has been located. The charge remains pending — counsel's live case page, read on 28 August 2026, lists it that way and no public EEOC determination exists — inside an enforcement environment that moved underneath it, with the executive order of 23 April 2025 directing agencies to deprioritise disparate-impact liability and a reported internal EEOC directive ordering disparate-impact-only charges closed by 30 September 2025, past which charges also alleging intentional discrimination reportedly proceed.
The sociotechnical reading
Every other case in this domain begins with an application. A person applies, a system scores the application, someone acts on the score, and a rejection is written down somewhere. That sequence is what makes hiring discrimination tractable: it produces a record, and a record can be audited, subpoenaed, and counted. This case begins one step earlier and has none of it. The system here allocates attention rather than adjudicating candidates: for each impression of each job ad it chooses, from inside an eligible audience the employer has already defined, who receives it. The people it does not reach never apply anywhere. They are not rejected. They generate no adverse-action record, no disparate-impact statistic, no charge of their own. The deployment's characteristic harm is invisible to every instrument the field built for hiring discrimination, and that is a fact about the position of the system in the pipeline, not about anyone's diligence.
The record's second signature is that each remedy landed one layer above the measured failure. The March 2019 settlements were a real, enforced structural change, and they governed the request: what an advertiser may specify. The peer-reviewed audit published weeks later measured skew surviving identical neutral requests, arising in delivery. The transparency regime that lets an outsider read per-ad delivery demographics governs political ads; employment ads sit outside it, so the charge's own evidence is a set of job ads that happened to be classified political. And the one delivery-layer control that carries numbers, an independent reviewer and a court is anchored to housing ads under housing law, reaching employment as an extension the operator reports and nobody has verified. Three times over, the instrument and the thing it was meant to govern are one step apart. The map draws that: the portal is a separate component from the substrate it hands work to, the handoff between them is its own pathway, and the checks that would close the loop for employment ads — an administrative determination, and a reconciliation of published figures against the platform's own delivery data — are drawn at zero, because the record documents that neither has been performed for this ad category.
What follows is a compensating structure, and it is the most instructive thing here. If the operator publishes nothing about employment delivery, the detection layer has to be outside the operator, and it is: two peer-reviewed audits bought with ordinary advertiser accounts, and an advocacy corpus scraped from a public library. Both channels work. Neither has authority. The auditors' finding can reach the employer — who can then read academic literature about a campaign it cannot inspect — and it reaches the operator only by becoming a legal claim, which is what happened, and which has been pending for years. Meanwhile the channel that does have authority, the court-supervised compliance loop, points at the sibling ad category and stopped publishing in 2024. This is the shape the diagram is really about: a system whose only competent sensors sit outside it, whose only instrumented control governs a neighbouring domain, and whose harm class produces no evidence of its own.
It is worth saying plainly what the record does establish and what it does not, because the case is easy to overstate in both directions. Established: the settlement terms and their implementation; the termination of a targeting tool by court order; a remedial controller that, where it is measured, beat every threshold it was given and survived an independent recomputation to a 0.0 percent difference — a genuinely good result, and the strongest fact about oversight in this file. Also established, and attributed to two peer-reviewed studies: that in the 2019-to-2021 era, before any such controller existed, identically targeted job ads were delivered to skewed audiences by the delivery stage. Attributed to the EEOC: that seven named employers violated federal law through their own targeting choices. Alleged and disputed: that Meta's delivery algorithm relies on gender and age, or discriminates in employment. The per-ad percentages are single-ad snapshots from a pleaded corpus and are not platform rates. The audits' magnitudes belong to the pre-controller platform and do not transfer forward without saying so. And no outcome for any job seeker is computed anywhere in this atlas — whether someone saw a posting, applied, or was hired is a boundary quantity, recorded here and propagated nowhere, because the population this deployment screens is, by its own structure, the population about which nothing is measured.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.