The hiring-AI problem: bias audits, notices, and the law already watching hiring AI
The short answer
Hiring AI is the most-watched AI you run
A single resume screener can owe United States federal statutes, New York City's Local Law 144, the Illinois Human Rights Act, and the EU AI Act's high-risk regime at the same time. No other system in your estate is under this many spotlights, which is why it earns its own topic and its own record.
What you will be able to do
- Identify every hiring-related AI system in your organization by pulling from your AI systems inventory (built in (see Topic 0.2)) and naming, for each, exactly which stage of the hiring funnel it touches: sourcing, screening, ranking, assessment, or interview analysis.
The ideas, one by one
The category is the whole funnel, not the resume box
Targeted job ads, candidate sourcing, screening, assessments, and video-interview analysis are all named as employment AI. Map your systems stage by stage or you will govern one box and miss three.
"The vendor built it" is not a defense
Mobley v. Workday established that a hiring-AI vendor can be liable as the employer's agent, and that a company delegating hiring to software is delegating to an agent the same as to a human. Liability is shared; treat vendor hiring AI as your own system with your own audit rights.
A fairness statement is not a bias audit
The document that failed in court says "designed to be unbiased." A real audit reports per-group selection rates, impact ratios, the four-fifths result, the auditor, and the population. Read the numbers, demand the intersectional and full-funnel cuts, and treat a bare "pass" as unaudited.
Proxy discrimination is why inputs are not enough
Illinois banned ZIP code as a stand-in for a protected class because a model can discriminate through correlated variables it never labels as protected. Audit outcomes, not just inputs.
Notice is the cheapest obligation and the most skipped
Name the tool, say what it evaluates, offer a human alternative, deliver it before the tool runs. A few lines of plain text separate a documented good-faith process from an ambush.
Human oversight must be able to override
A recruiter who cannot see why the tool rejected someone and cannot reverse it is decorative, and the tool is really deciding, which is the exact configuration the agent theory targets.
Where you set the bar is a decision you must own
A hiring model outputs a score; a human picks the cutoff, and that choice moves every selection rate and impact ratio. Document who chose the threshold, what a fairer bar would have cost, and why you set it where you did, because the threshold is often the least discriminatory alternative a plaintiff will point to.
A clean audit expires
Applicant pools shift and vendors retrain models, so a passing audit can quietly fail; that is why the law wants an audit within the prior year. Run governance as a cadence with re-audit triggers and continuous monitoring, and keep the record current in your dossier so "we audited it" is always answerable with a recent date.
Disability is the class a standard audit can miss
The ADA adds accommodation duties and screen-out limits, and disability is often invisible to a group-by-group audit, so a real, specific accommodation route in the candidate notice is frequently the primary protection, not an afterthought.
Passing the four-fifths rule is not the finish line
The four-fifths rule is a rule of thumb. Courts also weigh statistical significance, so a small gap in a large applicant pool can still be actionable while a scary ratio in a tiny pool may be noise. Report both the four-fifths result and whether the disparity is significant for your sample; never let one stand in for the other.
Validity is the defense, and "accurate" is not validity
Once adverse impact is shown, the burden shifts to you to prove the tool is job-related and consistent with business necessity, which means a validation study showing it predicts success at this job. The HireVue facial-analysis feature was dropped because it added almost nothing to predicting performance: adverse-impact risk with no validity is indefensible. If your vendor cannot supply a validity study, treat that gap as a live liability now, not paperwork for later.
Know the shape of the case before it is filed
Disparate impact runs in three ordered steps: they show the gap, you show the tool works, they show a fairer tool would have worked too. Governing hiring AI is pre-loading the document that answers each step. Ask, on an ordinary day, "which document answers step two, and do we have it today," and the whole job becomes concrete.
This record persists
The bias-audit-and-notice record you build here is not a one-off; it feeds your conformity file (see Topic 5.6) and your evidence annex (see Topic 10.6), so when the regulator's letter arrives (see Topic 5.7) you produce documents, not assurances.
Listen to it
The same lesson, talked through by two hosts as episode 37 of the podcast.
Real cases
Example 1: Mobley v. Workday (United States, 2023 to present), the agent theory made real. This is the anchor. Derek Mobley, an applicant who is Black, over forty, and has a disability, applied to more than one hundred jobs whose employers used Workday's applicant-screening platform and was rejected every time, sometimes within an hour. His suit targeted the vendor, not only the employers. Judge Rita Lin of the Northern District of California held that Workday's software could be liable as the employers' agent because it participates in hiring decisions rather than passively recording them, and in May 2025 she certified a nationwide ADEA collective action reaching applicants forty and older screened out since 24 September 2020, drawing roughly fourteen thousand opt-ins. In March 2026 she denied Workday's motion to dismiss. The case is unresolved on the merits (emerging), but the doctrine, that a hiring-AI vendor can be an agent and that a 1967 statute reaches a modern algorithm, is established on this record. Source: Holland & Knight, "Federal Court Allows Collective Action Lawsuit Over Alleged AI Hiring Bias" (2025); Seyfarth Shaw, "Mobley v. Workday: Court Holds AI Service Providers Could Be Directly Liable" (2024).
Example 2: New York City Local Law 144 (United States, effective 2023), the first hiring-AI-specific rule. New York City became the first United States jurisdiction to require, by law, an independent bias audit of hiring algorithms, a published summary of the selection rates and impact ratios, and advance notice to candidates. Enforcement began 5 July 2023. The instructive part is what happened next: a New York State Comptroller audit in 2025 concluded that enforcement had been thin and compliance uneven. The lesson for your organization is not "the rule has no teeth"; it is that a live legal obligation exists whether or not the city is aggressively policing it, and a plaintiff's lawyer or a European regulator will read your missing audit very differently from a busy city agency. Source: NYC Department of Consumer and Worker Protection, Local Law 144 pages; Office of the New York State Comptroller, "Enforcement of Local Law 144" (2025).
Example 3: Illinois House Bill 3773 (United States, effective 1 January 2026), discrimination without intent, and the ZIP-code trap. Illinois amended its Human Rights Act to make discriminatory use of AI in employment a civil-rights violation regardless of intent, and it specifically banned using a ZIP code as a proxy for a protected class. That ZIP-code clause is a tell that legislators now understand proxy discrimination: a model that never sees race can still discriminate by race if it leans on a variable tightly correlated with it. The unsettled notice rules (temporarily withdrawn in 2026) show something else worth teaching: even where the operational detail is still being written, the underlying prohibition is live, and "the rules were not final" is not a defense to using a discriminatory tool. Source: Illinois General Assembly, HB 3773; Seyfarth Shaw and Ogletree Deakins analyses (2024 to 2026).
Example 4: The EU AI Act's Annex III, point 4 (European Union), hiring AI named as high-risk by category. The Act does not wait to see whether a given screener discriminates; it classifies recruitment and selection AI as high-risk by its function (Regulation (EU) 2024/1689, Annex III, point 4). That is a different regulatory philosophy from the United States' outcome-first, litigate-after model: the EU front-loads the duty. The high-risk obligations for stand-alone Annex III systems were deferred to 2 December 2027 by the 2026 Digital Omnibus, so an organization has a real window to prepare, but the classification itself is settled and the direction is clear. Source: EUR-Lex, Regulation (EU) 2024/1689; Council of the EU, Digital Omnibus adoption (29 June 2026); European Commission draft guidelines on high-risk AI in employment (2026).
Example 5: HireVue drops facial analysis (United States, 2019 to 2021), the interview stage, and validity as the real test. HireVue, a widely used video-interview vendor, sold a product that scored candidates partly on facial expressions, and it marketed claims about measuring traits like "cognitive ability" and "emotional intelligence" from a recorded interview. In November 2019 the Electronic Privacy Information Center (EPIC) filed a complaint with the United States Federal Trade Commission arguing the practice was unfair and deceptive. HireVue commissioned an independent algorithmic audit from O'Neil Risk Consulting and Algorithmic Auditing (ORCAA), and in January 2021 it announced it would stop using facial analysis in its assessments. The detail governance teams should carry is why it dropped the feature: reporting indicated the facial-analysis component added only a tiny sliver to predicting job performance. Read against Section 3H, that is the whole lesson in one fact. A feature carrying obvious adverse-impact and privacy risk that also cannot show it predicts job success is indefensible: it fails the job-relatedness step with nothing to trade for the risk. Note the honest boundary of the story: HireVue kept analyzing other signals such as speech and intonation, so "dropped facial analysis" is not "solved bias." The teachable point is the test, not a clean ending: a hiring-AI feature that cannot prove validity is a feature you cannot defend. Source: EPIC, "HireVue, Facing FTC Complaint From EPIC, Halts Use of Facial Recognition" (2021); Fortune, "HireVue stops using facial expressions to assess job candidates amid audit of its A.I. algorithms" (19 January 2021).
A note on globally varied practice. These five examples span a city, a state, a federal system, a supranational regulation, and a vendor forced to change by public and regulatory pressure, and they represent two governance philosophies: the United States tends to prohibit outcomes and enforce after the fact through litigation and agencies, while the European Union front-loads duties on a named high-risk category before deployment. The practical contrast is worth holding as a pair of questions each philosophy asks of your tool:
- The United States, outcome-first, asks after the fact: did this tool produce a worse outcome for a protected group, and if so, can you prove it was job-related and that no fairer alternative existed? You may deploy first, but you carry the burden the day a claim lands.
- The European Union, duty-first, asks before deployment: is this tool in a named high-risk category, and have you completed the risk management, documentation, oversight, and conformity assessment the category demands? The duty attaches to the category, before any harm is shown.
A tool used in both worlds owes both answers at once, which is why your record must satisfy the stricter reading of each. Other jurisdictions are moving too, and this layer changes quickly; the discipline is to re-verify your own operating jurisdictions each hiring cycle rather than assume the map you learned last year still holds (see Topic 6.1 for the wider United States mosaic and other regions).
Where people go wrong
Mistake 1: "We do not use AI to hire." Most organizations that say this are wrong. They picture a robot reading resumes and miss the targeted-ad tool that decides who sees the job, the applicant-tracking system that auto-ranks, and the video-interview product that scores candidates. The fix: map the whole funnel from Section 3B, stage by stage, against your systems inventory, before you answer the question.
Mistake 2: "The vendor built the model, so the liability is the vendor's." Mobley v. Workday cuts this off in both directions. The employer delegated a hiring function and carries the liability of that delegation; the vendor's software participates in the decision and can be pulled in as an agent. "It is the vendor's model" is not a defense; it is a description of shared exposure. The fix: treat vendor hiring AI as your system, with your audit rights and your documentation.
Mistake 3: "The model never sees race, so it cannot discriminate." Proxy discrimination is the whole reason Illinois banned ZIP code as a stand-in for a protected class. A model can reproduce a protected trait through correlated variables it is allowed to see. Facial neutrality of inputs does not equal fairness of outcomes. The fix: audit outcomes (selection rates and impact ratios), not just inputs.
Mistake 4: "The audit passed, so we are fine." An audit that reports a checkmark can be hiding a failing impact ratio, an omitted small group, or an intersectional failure it never computed. A "pass" on race and a "pass" on sex can coexist with a severe failure for older women. The fix: read the actual numbers, demand the intersectional and full-funnel cuts, and treat a bare "pass" as unaudited.
Mistake 5: "New York City barely enforces Local Law 144, so we can skip the audit." Weak city enforcement is a real finding, but it protects you only from the city. It does nothing against a private plaintiff, a class action, or an EU regulator, all of whom will read your missing audit as a smoking gun. The fix: comply to the standard, not to the current enforcement temperature.
Mistake 6: "Human oversight means a person clicks approve." A recruiter who rubber-stamps a ranked list without being able to see or reverse the tool's rejections is not oversight; the tool is deciding. That is the configuration the agent theory targets. The fix: give the human the ability to see why the tool decided and the authority to override it, and log when they do.
Mistake 7: "Notice is a legal formality we can put in the privacy policy." A buried, vague notice is not a notice. The rules want a specific tool named, in plain words, before it runs, with a real alternative offered. The fix: post a plain, specific notice in advance; it is the cheapest obligation you have and the easiest to prove.
Mistake 8: "High-risk under the EU Act means it is banned, so we must stop." High-risk is not prohibited; it is heavily governed, and the stand-alone obligations are deferred to December 2027. Panic-pulling a useful tool can create its own harm. The fix: govern it to the high-risk standard on a timeline, do not confuse a duty with a ban (see Topic 5.3).
Mistake 9: "It passes the four-fifths rule, so it is legally clean." The four-fifths rule is a rule of thumb, not the ceiling of the law. With a large applicant pool, a gap too small to fail the four-fifths rule can still be statistically significant and actionable; courts look at standard deviations, not only the eighty-percent line. The fix: report both the four-fifths result and whether the disparity is statistically significant for your sample size, and never treat a four-fifths pass as the end of the analysis.
Mistake 10: "The tool is accurate, so it is defensible." Accuracy and validity are not the same thing, and validity is what the law asks for. A model can be internally accurate at reproducing whatever it was trained on, including past biased decisions, while having no evidence that it predicts actual job success. When adverse impact appears, the defense is job-relatedness (criterion, content, or construct validity), and "it is accurate" is not that evidence. The fix: hold a validation study that shows the tool predicts performance on this job, and if the vendor cannot supply one, treat that gap as a live liability, not a formality.
Mistake 11: "We can use the vendor's own bias audit to satisfy our obligation." New York City's Local Law 144 requires an independent bias audit, and "independent" is doing real work in that sentence: an audit the vendor commissioned of its own general product, on its own applicant pool, is not automatically independent of your specific use, your specific applicant pool, and your specific configuration of the tool (your threshold, your feature set). A vendor's marketing-adjacent audit of "the product" is a different document from an audit of the tool as you actually deploy it. The fix: confirm the audit was performed by an auditor independent of the vendor, on data that reflects your own applicant pool and configuration, not just accept whatever PDF the vendor already had on file.
Questions people ask
- What is automated employment decision tool (AEDT)?
- The term New York City Local Law 144 uses for a computational tool that substantially assists or replaces a hiring or promotion decision. If a tool meets this definition for a New York City role, it triggers the bias-audit, publication, and notice duties. More on Automated employment decision tool (AEDT)
- What is adverse impact (disparate impact)?
- A selection procedure that is neutral on its face but produces a worse outcome for a protected group. Under United States federal law it can be unlawful even without intent to discriminate, unless the employer shows the procedure is job-related and consistent with business necessity.
- What is agent theory (agent liability)?
- The legal idea, accepted in Mobley v. Workday, that a hiring-AI vendor can be liable under anti-discrimination statutes as the employer's agent, because the software carries out an employment function rather than passively recording the employer's own decision.
- What is Annex III, point 4 (EU AI Act)?
- The part of Regulation (EU) 2024/1689 that classifies AI used for recruitment, selection, and employment-relationship decisions (promotion, termination, task allocation, performance monitoring) as high-risk.
- What is article 6(3) exemption (EU AI Act)?
- A narrow carve-out that can drop an Annex III system out of high-risk if it performs only a narrow procedural task and does not materially influence the outcome. It never applies to a system that profiles people, which a candidate screener does by design.
Keep going
This lesson builds AI in hiring and employment decisions, and that page shows the roles that hire for it. Every Certified AI Governance Professional (CAIGP) lesson.
Watch the lesson and prove it
Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.
Watch the lesson and prove it →