The Discrimination Happens Before Anyone Applies: AI Job Ad Targeting and Delivery
Bias audits examine the screening model. The screening model only ever sees people who applied, and who applied is decided by an optimization system nobody in legal has looked at. The largest filter in most hiring funnels sits above the funnel.
Recruitment Was Always Covered
There is a persistent assumption that employment discrimination law begins at the application. It does not. The ADEA specifically addresses printing or publishing a notice or advertisement for employment that indicates a preference, limitation, specification, or discrimination based on age. Title VII reaches recruitment practices that limit or classify applicants in ways that deprive them of employment opportunities. Recruiting is a covered practice on its own terms.
What changed is the mechanism. A newspaper ad reached whoever bought the paper. A programmatic recruitment campaign reaches whoever an auction model predicts will click, and that prediction is trained on historical engagement patterns that encode exactly the demographic distribution of who has historically been recruited into that role.
The result is a system that can produce a delivery pattern indistinguishable from deliberate age or sex targeting while every human involved made only budget and creative decisions. The doctrine that reaches this is disparate impact, which asks what the practice did rather than what anyone intended.
Three Distinct Failure Modes, Three Different Defenses
1. Explicit targeting selections
Age bands, gender selection, or proxies chosen in the campaign builder. This is the smallest category and the least defensible — it is a preference stated in a job notice, recorded in the ad account, and trivially discoverable. Several platforms now restrict these options for employment ads, but restrictions vary by placement, by API path, and by whether the campaign was flagged as a special ad category in the first place.
2. Proxy targeting
Interest, behavior, education, or affinity segments that correlate tightly with a protected characteristic. Targeting recent graduates, or interests concentrated in one age cohort, or exclusions that remove a demographic in effect. The defense here is business necessity for the specific criterion, and 'it performed better' is a cost argument rather than a job-relatedness argument.
3. Optimized delivery skew
No demographic selection at all, but the delivery algorithm allocates impressions unevenly. This is where most exposure actually sits, and the honest defense is not that the employer had no role — it is that the employer measured delivery, found the skew, and took corrective action such as reach-based buying, broad audience configuration, or supplementary channel spend to rebalance exposure.
Why the Applicant Pool Cannot Validate the Campaign
The instinct when someone raises this is to check the applicant demographics and report that the pool looks reasonable. That analysis is circular. The applicant pool is composed of people who saw the ad; if delivery skewed by twenty points, a pool that mirrors the delivery skew looks internally consistent and proves nothing about whether the opportunity was fairly advertised.
The correct comparison is between the delivered audience and the labor market availability for that occupation and geography. Public occupational data provides a usable baseline for most roles. A recruitment campaign for a job where the available workforce is broadly distributed across age bands, delivering three quarters of its impressions to a single band, is a finding — regardless of how the resulting applicants performed.
This asymmetry is also why the exposure survives an otherwise clean bias audit. Statutory audit regimes for automated employment decision tools measure impact ratios at the selection stage. A tool can produce perfectly balanced selection rates on an audience that was never balanced to begin with, and the audit will report a pass.
The lookalike audience is the one to look at first
Optimized delivery skew is at least arguably emergent. A lookalike audience seeded on current employees is not — it is a documented instruction to find more people resembling a population the employer already has, executed against the employer's own uploaded list. If the incumbent workforce is demographically skewed, the campaign is engineered to reproduce that skew, and the seed file, the upload timestamp, and the audience definition are all sitting in the ad account. Where a recruiting team uses one, it is worth a specific decision with a specific justification rather than a default setting inherited from the marketing team.
The Delivery-Side Audit
1. Inventory every paid recruitment channel
Social platforms, programmatic job distribution, sponsored job board slots, retargeting pixels firing on career pages, and any agency-run spend. Agency-managed campaigns are the most frequent blind spot because the configuration lives in someone else's account and the employer never sees the targeting history.
2. Pull reach and impressions by age band and gender per campaign
Not clicks, and not applications — delivery. Most ad platforms expose this natively. Do it per campaign and per role family, because aggregate account-level numbers average away exactly the roles where skew concentrates.
3. Compare against occupational availability, not against your funnel
Use external labor market data for the occupation and metro area as the denominator. Write down the source and the vintage. This is the single analytical choice that determines whether the audit is meaningful or self-confirming.
4. Export targeting configuration and audience definitions
Selected and excluded segments, special ad category flags, lookalike seeds and their source lists, placement restrictions, and the optimization objective for each campaign. Configuration is what a plaintiff subpoenas; knowing what it says before they do is the point of the exercise.
5. Change the buy where you find skew, and record the change
Reach-based or broad-audience buying, removal of proxy segments, supplemental spend in channels reaching underrepresented cohorts, and periodic re-measurement. The documented remediation is what converts a bad finding into evidence of a functioning compliance process.
6. Put recruitment advertising inside the HR governance perimeter
In most organizations job ads are bought by marketing under marketing's optimization norms and never reviewed by anyone with employment law responsibility. Extending the same approval and audit path used for screening tools to recruitment media is a structural fix that costs almost nothing.
Frequently Asked Questions
We flagged our campaigns as an employment special ad category. Are we covered?
That control removes certain targeting options and limits some optimization signals, which meaningfully reduces the explicit and proxy targeting failure modes. It does not eliminate delivery skew, because the auction still allocates impressions by predicted engagement within whatever audience remains. Treat the flag as necessary and insufficient, and keep measuring delivery after enabling it.
Our recruiting agency runs all the ads. Does the exposure transfer to them?
Contractually you can allocate cost, but the statutory obligation stays with the employer, and an entity that procures workers for an employer can itself meet the definition of an employment agency with its own liability. The practical step is to require delivery reporting and targeting configuration disclosure in the agency contract, because without it the employer cannot audit a practice it remains answerable for.
Is there a safe threshold for delivery imbalance?
There is no statutory delivery threshold. The four-fifths rule is a selection-rate convention, not a reach standard, though it is a reasonable internal trigger for investigation. What matters more than a single number is whether the imbalance is large, persistent across campaigns, unexplained by the underlying labor market, and left unaddressed after being observed.
Do we have to collect demographic data on applicants to do this analysis?
No — and that is what makes this tractable. Delivery breakdowns are reported by the ad platform in aggregate and require no new personal data collection by the employer. You are analysing an aggregate audience report against public labor market statistics, not building an applicant demographic database.
What is the fastest meaningful change for a small team?
Switch recruitment campaigns off engagement-optimized objectives and onto reach-based delivery with broad audiences, then pull the age and gender reach breakdown thirty days later and compare it to the prior period. It is a settings change, it addresses the dominant failure mode directly, and it produces the before-and-after record that demonstrates the issue was noticed and acted on.
Audit the Top of the Funnel or the Rest Is Decoration
An employer can run a defensible bias audit on its screening tool, document human oversight, and still have advertised the role to a demographically narrow slice of the labor market. Nothing downstream repairs that, and nothing downstream detects it either.
The data needed to check already exists in the ad account and requires no new collection. The only real question is whether anyone with an employment law obligation has ever been shown it.