Ch 4 — AI for Recruiting & Talent

The technical mechanics of recruiting AI — parsing, matching, bias encoding, auditing, and compliance
Regulatory content current as of November 2026. Verify before acting on legal guidance
Under the Hood
database
Data
arrow_forward
tune
Features
arrow_forward
model_training
Model
arrow_forward
score
Score
arrow_forward
fact_check
Audit
arrow_forward
verified
Comply
-
Click play or press Space to begin the deep dive...
Step- / 8
description
How Resume Parsing Actually Works
NLP tokenization, named entity recognition, and why formatting breaks everything
The Parsing Pipeline
When a resume hits your ATS, the parser runs through a multi-step NLP pipeline:

1. Document conversion: PDF/DOCX is converted to raw text. Tables, columns, headers, and images are often lost or mangled at this step.
2. Tokenization: The text is broken into individual words and phrases (“tokens”). “Senior Software Engineer” might become three tokens or one, depending on the parser.
3. Named Entity Recognition (NER): The model identifies which tokens are names, companies, job titles, dates, skills, degrees, and institutions.
4. Relationship extraction: Which skills belong to which job? Which dates correspond to which employer?
5. Normalization: “Sr.” becomes “Senior.” “MIT” becomes “Massachusetts Institute of Technology.”
Why Formatting Matters
PARSES WELL: Simple, single-column layout Standard section headers (Experience, Education) Consistent date formatting (Jan 2022 – Present) Plain text or simple PDF PARSES POORLY: Multi-column layouts (left sidebar + main body) Tables for formatting Graphics, icons, or progress bars for skills Non-standard section headers ("My Journey") Scanned PDFs (image, not text) International date formats (27/03/2026)
Non-Standard Resumes
Career changers: NER models expect linear career paths. A teacher-to-developer transition confuses parsers that weight “years in role” heavily.
Employment gaps: Most parsers flag gaps but can’t understand context (caregiving, illness, travel).
International formats: CV styles from India, Germany, or Japan differ dramatically from US norms. Parsers trained on US resumes struggle with these.
Technical reality: No parser is perfect. Vendors publish their own accuracy figures and independent benchmarks are scarce, so treat those numbers with caution: across the dozens of fields on a resume, even a small error rate per field means a meaningful share of resumes will have at least one field parsed wrongly. If that field is used for screening, qualified candidates get eliminated for formatting, not qualifications.
join_inner
Matching Algorithms Explained
TF-IDF, embeddings, LLMs, and why “managed a team” should match “led direct reports”
Three Generations of Matching
Generation 1: TF-IDF (Term Frequency–Inverse Document Frequency)
Counts how often words appear in a resume vs. how common they are overall. If “Kubernetes” appears in your resume but is rare across all resumes, it gets a high score. Problem: it’s still just word counting. “Managed a team” and “led direct reports” score completely differently.

Generation 2: Embeddings
Words and phrases are converted into numerical vectors (lists of numbers) that capture meaning. “Managed a team” and “led direct reports” produce similar vectors because they mean similar things. This is the breakthrough that made semantic matching possible.

Generation 3: LLM-Based Matching
Large language models understand context, nuance, and inference. They can reason that “reduced customer churn by 15%” is relevant to a “customer success manager” role even if the exact title never appears.
Cosine Similarity in Plain Language
The Core Idea: Every piece of text gets converted into a list of numbers (a "vector"). Think of it as coordinates in a vast space of meaning. Cosine Similarity: How much do two vectors point in the same direction? For text, scores run from about 0 (unrelated) to 1 (identical meaning). Example: "managed a team of 12" "led direct reports" Cosine similarity: ~0.87 (very similar) "managed a team of 12" "proficient in Excel" Cosine similarity: ~0.14 (unrelated) // This is why semantic matching finds // candidates that keyword matching misses. // The math captures meaning, not just words.
Vendor question: “Does your matching use keyword matching, embeddings, or LLM-based reasoning?” The answer tells you how sophisticated (and expensive) their AI actually is. Some vendors claim “AI matching” but still rely on keyword scoring under the hood.
warning
How Bias Gets Encoded
From training data to proxy variables — a concrete walkthrough of bias in the pipeline
The Bias Pipeline
Bias doesn’t enter at one point — it compounds through the entire pipeline. Here’s a concrete example of how it works in a resume screening model:

Step 1 — Training data: You feed the model 5 years of hiring decisions. Your company historically hired 70% men for engineering roles. The model learns: “male-associated features predict hire.”

Step 2 — Feature selection: You remove gender from the data. But the model finds proxy variables: university names (all-women’s colleges), extracurriculars (“women’s rugby”), even name patterns.

Step 3 — Threshold effects: You set the screening cutoff at 75%. Due to biased training, women score 6 points lower on average. The cutoff eliminates 62% of female applicants but only 38% of male applicants.
The Bias Pipeline in Code
STAGE 1: TRAINING DATA Historical hires: 70% male, 30% female // Model learns: male patterns = "good hire" STAGE 2: FEATURE SELECTION Removed: gender, race, age Still present as proxies: ├ University name (→ gender, class) ├ Zip code (→ race, income) ├ Name patterns (→ ethnicity) ├ Employment gaps (→ gender) ├ Extracurriculars (→ gender, culture) STAGE 3: MODEL OUTPUT Avg score, male applicants: 78 Avg score, female applicants: 72 Screening threshold: 75 STAGE 4: RESULT Male pass rate: 62% Female pass rate: 38% Four-fifths test: 38/62 = 61% — FAIL // Disparate impact even though gender // was never an explicit input.
The core problem: Removing protected attributes from the input does not remove bias from the output. Proxy variables carry the same information through different channels. This is why bias audits test outcomes, not inputs.
calculate
Adverse Impact Analysis
The four-fifths rule with real numbers, selection rates, and what triggers scrutiny
Calculating Selection Rates
Adverse impact analysis compares selection rates across demographic groups at each stage of the hiring funnel. A selection rate is simply: (number selected) / (number who applied) for each group.

The four-fifths (80%) rule: If any group’s selection rate is less than 80% of the highest group’s rate, that is a flag for possible adverse impact, not a legal finding. It is a rule of thumb: sample sizes and statistical significance matter too. If the disparity holds up, the employer has to show the selection criterion is job-related and consistent with business necessity.

What triggers scrutiny: A candidate complaint or lawsuit, a state civil rights investigation, or a legally required audit. The EEOC announced in June 2026 that it will no longer pursue disparate impact cases, so this math is now tested mainly by private plaintiffs and state agencies. NYC Local Law 144 requires annual bias audits for covered automated employment decision tools used on New York City candidates.
Worked Example
AI SCREENING: 500 APPLICANTS Group Applied Passed Rate White 200 120 60% Black 120 48 40% Hispanic 90 45 50% Asian 90 54 60% Highest rate: 60% (White & Asian) 80% threshold: 48% (60% × 0.80) Four-fifths test: White: 60/60 = 100% — PASS Black: 40/60 = 67% — FAIL Hispanic: 50/60 = 83% — PASS Asian: 60/60 = 100% — PASS Adverse impact detected for Black applicants. // This analysis must be run at EVERY stage // where AI filters candidates: screening, // assessment, interview selection, offer.
Who enforces now: The iTutorGroup case ($365K, 2023) targeted age-based algorithmic rejection. In June 2026 the EEOC said it will stop pursuing disparate impact cases, but Title VII and the ADEA have not changed, and private plaintiffs and state agencies still bring those claims. In Mobley v. Workday, a federal court has let an age-bias collective action proceed against an AI screening vendor. The Eightfold AI class action (Jan 2026) adds another front: alleged secret applicant scoring dossiers may violate FCRA. The law doesn’t care whether a human or an algorithm made the discriminatory decision.
visibility
Explainability Requirements
GDPR Article 22, NYC Local 144, and what “explainable AI” actually means in practice
Can the AI Explain Why?
When an AI rejects a candidate, can it explain the reasoning? This isn’t just an ethical question — it’s increasingly a legal requirement.

GDPR Article 22: People in the EU have the right to not be subject to decisions based solely on automated processing that significantly affect them. Employment decisions qualify. When automated decisions are made, individuals have the right to obtain an explanation and to contest the decision.

NYC Local Law 144: Employers using automated employment decision tools (AEDTs) in New York City must conduct annual bias audits by an independent auditor, publish audit results on their website, and notify candidates that an AEDT is being used.

In practice: “Explainable AI” means the system can identify which factors contributed to a score — not just that the candidate scored 72, but why: missing skill X, insufficient years in role Y, low semantic match on competency Z.
Explainability Spectrum
BLACK BOX (Not Explainable) "The candidate scored 42 out of 100." // No insight into why. Cannot audit. // Cannot defend in legal challenge. PARTIAL EXPLAINABILITY "Top factors: skills match (low), experience years (medium), education match (high)." // Better. Shows which categories matter. // Still doesn't show how they interact. FULL EXPLAINABILITY "Score: 42. Key factors: - Missing 3 of 5 required skills (-22) - 2 yrs experience vs 5 required (-14) - Strong education match (+8) - Semantic match on leadership: 0.83 Confidence: 78%. Similar candidates who were hired: 12% at this score." // This is what you should demand. // Auditable, defensible, actionable.
Operational standard: If your AI screening tool cannot produce a factor-level explanation for every score, it is far harder to diagnose bias, defend in litigation, or review meaningfully as a human. Outcome audits can still be run without it, but explanations are increasingly expected, and in some jurisdictions required.
balance
De-biasing Techniques
What works, what doesn’t, and what’s just marketing
Techniques That Have Evidence
Removing protected attributes: Necessary but not sufficient. As we showed in Step 3, proxy variables carry the same signal. It’s a starting point, not a solution.

Adversarial de-biasing: Training a second model whose job is to detect whether the primary model’s outputs correlate with protected attributes. If the “adversary” can guess the candidate’s demographic group from the score, the primary model is biased and gets penalized. This has real research backing but requires significant technical investment.

Group-specific cutoffs (do not use): Setting a different score threshold for each demographic group would equalize selection rates, but Title VII has prohibited adjusting employment test scores or cutoffs by race, color, religion, sex, or national origin since 1991. The lawful route is to change the model, the inputs, or the cutoff for everyone until the gap closes.

Blind resume review: Removing names, addresses, universities, and dates before AI screening. Reduces some proxy bias but doesn’t eliminate it (writing style, vocabulary, and career patterns still carry signal).
What Works vs. Marketing
Mostly Marketing
“Our AI is bias-free.” No AI is bias-free — this is a red flag, not a selling point. “We removed race and gender from the data.” Necessary but not sufficient. “Our diverse training data ensures fairness.” Diverse data helps but doesn’t guarantee fair outcomes. Demand audit results, not claims.
Backed by Evidence
“We conduct quarterly adverse impact analyses.” “We use adversarial de-biasing and can share the methodology.” “We publish our bias audit results and model card.” “We support blind screening and configurable feature exclusion.” These are verifiable, auditable claims.
The honest truth: No de-biasing technique eliminates all bias. The goal is monitoring, measurement, and continuous improvement — the same approach you use for any operational risk. Vendors who promise zero bias are either naive or dishonest.
assessment
Audit Frameworks
How to structure a bias audit: methodology, independence, cadence, and intersectionality
Audit Structure
1. Data analysis: Collect demographic data on all candidates processed by the AI at each funnel stage. Calculate selection rates by group.

2. Disparate impact testing: Apply the four-fifths rule at each stage. Identify which groups are disproportionately affected and at which pipeline stages.

3. Intersectional analysis: Test not just single categories (race, gender) but intersections (Black women, older Hispanic men). Bias often hides at intersections that single-axis analysis misses.

4. Feature importance review: Examine which input features most influence the model’s scores. Flag any feature that correlates strongly with a protected attribute.

5. Ongoing monitoring: Set up dashboards that track selection rates in real time, with alerts when any group’s rate drops below the four-fifths threshold.
Who, When, and How
WHO SHOULD AUDIT Independent third party — not the vendor, not your internal team. Look for: ├ I/O psychologists ├ AI fairness consultants ├ Employment law firms with AI expertise ├ Firms compliant with NYC LL144 standards CADENCE Annual: Full comprehensive audit Quarterly: Selection rate monitoring Continuous: Real-time dashboard alerts Trigger-based: After model updates, new features, or complaints DOCUMENTATION REQUIREMENTS [ ] Audit methodology described [ ] Data sources identified [ ] Statistical methods documented [ ] Results by group reported [ ] Intersectional analysis included [ ] Remediation actions listed [ ] Auditor independence attested
NYC LL144 standard: The law requires an “independent auditor” but does not define qualifications. Best practice: use a firm with I/O psychology expertise and demonstrated experience in EEOC-style adverse impact analyses. The audit must be published on your website — it will be scrutinized.
gavel
The Compliance Stack
Federal, state, local, and international laws governing AI in recruiting
Federal Framework
Title VII, ADA, ADEA: Existing anti-discrimination laws apply to AI decisions. Federal agency enforcement has narrowed: the EEOC removed its 2023 AI guidance, its June 2026 enforcement plan says it will not pursue disparate impact cases, and a June 2026 Justice Department opinion called the EEOC’s disparate impact guidelines (the source of the four-fifths rule) unconstitutional. None of this changes the statutes or binds the courts. Private plaintiffs and state agencies can still bring disparate impact claims, and the four-fifths ratio remains the standard yardstick in bias audits and litigation.

Federal contractors: Executive Order 11246, the basis for OFCCP’s race and sex affirmative action audits, was revoked in January 2025 and its regulations were rescinded in 2026. Obligations for veterans (VEVRAA) and people with disabilities (Section 503) remain, in narrower form.

Key principle: The employer is responsible for the AI’s decisions, even if the vendor built the tool. “The vendor told us it was fair” is not a legal defense. The Eightfold AI class action (January 2026) alleges the platform generated secret applicant scoring dossiers without disclosure, potentially violating the Fair Credit Reporting Act and California consumer reporting laws. The case is still at the motion-to-dismiss stage, but it is the first major test of FCRA claims against an AI hiring vendor.
Compliance Checklist by Jurisdiction
FEDERAL (15+ EMPLOYEES; ADEA 20+) [ ] Title VII / ADA / ADEA compliance // EEOC stepped back from disparate impact // (June 2026); private and state suits continue [ ] VEVRAA / Section 503 (if fed contractor) [ ] FCRA compliance for AI scoring tools // See Eightfold AI class action (Jan 2026) STATE LAWS [ ] California CRD Regulations (Oct 2025): ADS covered by bias law; 4-yr records; testing and oversight support a defense [ ] California SB 947 (July 2027): No AI-only discipline or termination [ ] California CCPA ADMT rules (Jan 2027): Pre-use notice, opt-out, access rights // Risk assessments phased in from 2026 [ ] Colorado SB 26-189 (Jan 2027): Notice, explanation, human review // Replaced the repealed SB 24-205 [ ] Texas TRAIGA / HB 149 (Jan 2026): No AI use with intent to discriminate // Rejects disparate impact as standalone [ ] Illinois HB 3773 (Jan 2026): No AI with discriminatory effect; notice required [ ] Illinois AIVIA: Video interview AI consent [ ] Connecticut PA 26-15 (Oct 2027): Written notice before AI-assisted decisions [ ] New Jersey disparate impact rules (Dec 2025): cover automated tools [ ] Maryland: Facial recognition needs consent LOCAL LAWS [ ] NYC Local Law 144: Annual bias audit + published results + notice INTERNATIONAL [ ] EU AI Act: Employment AI classified as "high-risk" — requires conformity assessment, transparency, human oversight [ ] GDPR Art. 22: Right to not be subject to solely automated decisions WATCH [ ] No federal AI law has passed (Nov 2026); Washington is pushing to preempt state rules. Do not plan around harmonization
The landscape is moving fast. The rescission of the Biden-era AI executive order and removal of EEOC guidance have not changed the statutes. What changed is who enforces them: the EEOC has stepped back from disparate impact cases, while private class actions and state agencies have stepped in. Meanwhile, states are filling the federal gap: California (CRD regulations, Oct 2025; SB 947 and the CCPA automated-decision rules in 2027), Illinois (HB 3773, Jan 2026), Texas (HB 149, Jan 2026), and Colorado and Connecticut (new notice laws in 2027). No federal AI law has passed, and Washington is moving to preempt state AI rules (Executive Order 14365, Dec 2025, set up a DOJ task force to challenge them), so expect federal-vs-state fights, not quick harmonization. Build your AI recruiting governance to the highest standard you face.
Sources & Further Reading
• EEOC, “Select Issues: Assessing Adverse Impact in Software, Algorithms, and AI” (guidance removed 2025; underlying Title VII framework unchanged)
• NYC Local Law 144 (2023) — Automated Employment Decision Tools
• California Civil Rights Department, Regulations on Automated Decision Systems (effective Oct 2025)
• EEOC National Enforcement Plan, FY 2025–2029 (June 2026); DOJ Office of Legal Counsel opinion on disparate impact liability (June 2026)
• Colorado SB 26-189 (signed May 2026, effective Jan 2027; replaced SB 24-205, which never took effect)
• California SB 947 (effective July 2027); CCPA automated decisionmaking regulations (employer compliance Jan 2027)
• Texas HB 149, Texas Responsible AI Governance Act (effective Jan 2026)
• Illinois HB 3773 (effective Jan 2026); Illinois Artificial Intelligence Video Interview Act (2020)
• Connecticut Public Act 26-15; New Jersey N.J.A.C. 13:16 (Dec 2025)
• Mobley v. Workday (N.D. Cal.); Eightfold AI class action (Jan 2026): FCRA and California consumer reporting claims
• EU Artificial Intelligence Act, Regulation 2024/1689

Regulatory content current as of November 2026. Verify before acting on legal guidance.