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.”

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.

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.

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 40% of female applicants but only 25% of male applicants.

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, there is prima facie evidence of adverse impact. This isn’t automatic proof of discrimination — but it shifts the burden to the employer to demonstrate the selection criterion is job-related and consistent with business necessity.

What triggers an audit: A candidate complaint, a pattern in EEOC charges, an OFCCP compliance review (for federal contractors), or a state-level investigation. NYC Local Law 144 now requires annual bias audits for all automated employment decision tools.

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: EU residents 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.

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.

Calibration across groups: Adjusting score thresholds so that each demographic group has equivalent selection rates. Effective for outcome fairness but controversial — some argue it constitutes differential treatment.

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).

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.

EEOC (Title VII, ADA, ADEA): Existing anti-discrimination laws apply to AI decisions. The EEOC’s 2023 guidance on AI in hiring was removed from its website following the rescission of the Biden-era AI executive order — but this does not reduce employer liability. Title VII, ADA, and ADEA still apply in full. If AI produces disparate impact, the employer is liable regardless of whether federal guidance documents exist. The four-fifths rule applies regardless of whether a human or algorithm makes the selection.

OFCCP: Federal contractors face additional scrutiny. OFCCP requires affirmative action plans and can audit your selection procedures, including AI tools, at any time. If your AI screening creates adverse impact, you must demonstrate job-relatedness and business necessity.

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 — a landmark case for AI hiring liability.