Algorithmic Bias Glossary: Key Terms for HR & Recruiting

Algorithmic bias is the defining compliance risk of AI-powered talent acquisition — and most HR teams cannot define the core terms they need to manage it. Before any automated screening tool touches a candidate record, every recruiter and HR director involved in that decision needs a working vocabulary: what bias is, where it enters the system, how fairness is measured, and what legal exposure looks like in practice. This glossary provides that foundation.

These definitions are designed to support the broader discipline of automating the end-to-end ATS workflow before layering in AI — because bias risk is compounded when AI is deployed on top of manual, inconsistent processes rather than clean, audited automation infrastructure.

Core Terms: Bias, Fairness, and Discrimination

These are the foundational definitions. Every HR professional using AI in any part of the hiring process should be able to define and operationalize all of them.

Algorithmic Bias

Algorithmic bias is a systematic, repeatable error in a computational system that produces unfair or discriminatory outcomes across demographic groups. In HR and recruiting, algorithmic bias most commonly manifests in resume screening, candidate ranking, and interview scheduling tools that have been trained on historical data reflecting past human biases.

The critical distinction: algorithmic bias does not require intent. A system trained on five years of hiring data from a company that historically hired predominantly from one demographic group will encode that pattern as a predictive signal — and then apply it to new candidates, at scale, faster than any human reviewer could catch.

McKinsey Global Institute research on AI deployment has consistently highlighted that the quality and representativeness of training data is the primary determinant of downstream model fairness, not the sophistication of the algorithm itself.

• Source: Training data that reflects historical human hiring decisions
• Mechanism: Model learns spurious correlations between demographic proxies and “success” labels
• Outcome: Systematic underscoring or rejection of candidates from underrepresented groups
• Key risk: Operates at scale and speed that makes manual oversight insufficient without structured auditing

Fairness (in AI Systems)

Fairness in AI is not a single definition — it is a family of competing mathematical criteria, each encoding different normative assumptions about what equitable treatment means. The practical implication for HR teams: you must choose which fairness definition applies to your context, because optimizing for one often trades off against another.

The three most relevant fairness definitions for HR applications are:

• Individual fairness: Similar candidates are treated similarly, regardless of group membership
• Group fairness: Outcomes are equitable across demographic groups at the aggregate level
• Counterfactual fairness: A decision would be the same if the candidate’s protected attributes were different, all else equal

Fairness Metrics

Fairness metrics are the quantitative operationalizations of fairness definitions — the specific numbers HR teams measure to determine whether an AI tool is producing equitable outcomes. Selecting the right metrics is a legal and strategic decision, not just a technical one.

The most widely applied fairness metrics in hiring contexts:

• Demographic parity (statistical parity): Selection rates are approximately equal across demographic groups. A model passes demographic parity if it advances roughly the same percentage of applicants from each group.
• Equal opportunity: True positive rates are equal across groups — qualified candidates are identified and advanced at equal rates regardless of group membership. This is generally more appropriate for hiring than demographic parity, because it conditions on actual qualification.
• Equalized odds: Both true positive rates and false positive rates are equal across groups. More stringent than equal opportunity alone.
• Calibration: The model’s confidence scores mean the same thing across groups — a 75% score predicts the same outcome probability for candidates regardless of their demographic.

SHRM guidance on AI in HR consistently emphasizes that fairness metric selection must be documented and defensible before deployment, not chosen retroactively to pass an audit.

Disparate Impact

Disparate impact is a legal doctrine established under Title VII of the Civil Rights Act and subsequent case law (Griggs v. Duke Power Co.) that holds employers liable when a facially neutral employment practice produces a disproportionately adverse effect on a protected group — regardless of discriminatory intent.

For AI tools in HR, disparate impact is the primary legal vector. An automated screening system that advances significantly fewer candidates from a protected racial group, even without any explicit reference to race in its logic, can constitute unlawful disparate impact discrimination.

The EEOC’s four-fifths rule (80% rule) is the most common practical test: if the selection rate for any group is less than 80% of the rate for the most-favored group, a presumption of adverse impact is triggered. HR teams should apply this threshold to every automated decision point in their workflow, not just final hire decisions.

• Legal basis: Title VII, Age Discrimination in Employment Act, Americans with Disabilities Act
• Practical test: EEOC four-fifths (80%) rule applied to selection rates by protected group
• Scope: Applies to any employment decision, including automated screening steps, not just offers
• Emerging regulation: NYC Local Law 144 (2023) requires bias audits of automated employment decision tools before use

Data Quality Terms: Where Bias Enters the System

Most algorithmic bias in HR originates not in the model architecture but in the data the model is trained on. Understanding these terms explains why data auditing must precede model deployment.

Training Data Bias

Training data bias is the condition in which the dataset used to train an AI model contains patterns that do not accurately represent the target population or that encode historical discrimination as a predictive signal.

In HR, the most common form is historical outcome bias: the model is trained on who was hired (and succeeded) in the past, and learns that past selection patterns are valid predictors of future performance. If past selection patterns reflected biased human decisions, the model learns and automates that bias.

Harvard Business Review research on algorithmic hiring has noted that this mechanism is particularly insidious because the model appears to be performing well — it accurately predicts past outcomes — while systematically replicating the discrimination embedded in those outcomes.

Data Imbalance

Data imbalance is the condition in which a training dataset contains significantly more examples from some demographic groups than others, causing the model to perform better for majority-group members and worse for underrepresented groups.

A model trained on 10,000 resumes from men and 1,000 resumes from women will generally perform less accurately for women — not because it is intentionally discriminatory, but because it has had far fewer examples from which to learn what qualified female candidates look like.

Mitigation strategies include oversampling of underrepresented groups, synthetic data generation, reweighting techniques, and — most fundamentally — expanding the diversity of the historical data collected before training begins.

Proxy Variable

A proxy variable is a data input that is not itself a protected characteristic but correlates strongly enough with a protected characteristic to function as a substitute for it in a model’s decision logic.

Proxy variables are the primary reason that removing protected attributes from training data does not eliminate algorithmic bias. Classic examples in HR contexts:

• Zip code: Correlates with race, ethnicity, and income in most U.S. markets
• University name and tier: Correlates with socioeconomic status, race, and first-generation status
• Employment gap length: Correlates with gender (caregiving leave) and disability status
• Applicant name: Even stylistic features of names correlate with perceived race and ethnicity in audit studies
• Industry jargon patterns: Certain terminology clusters correlate with professional networks that are themselves demographically homogeneous

Identifying and addressing proxy variables requires correlation analysis between model input features and demographic outcomes — a step that must be performed by someone with statistical fluency before any model is deployed in a hiring decision.

Label Bias

Label bias occurs when the outcome variable used to train a model is itself the product of biased human decisions. In hiring, the most common label is “hired” or “successful” — but if the historical hiring decisions that generated those labels were biased, the model is trained to replicate biased judgments, not objective quality assessments.

Label bias is particularly difficult to detect because it is invisible in standard model performance metrics. A model with high accuracy against biased historical labels may be doing an excellent job of replicating past discrimination.

Transparency and Explainability

Transparency and explainability are not the same thing, and conflating them creates compliance gaps. Both are required for defensible AI deployment in HR.

AI Transparency

AI transparency is the degree to which the design, data sources, decision logic, and performance characteristics of an AI system are visible and auditable by human stakeholders. Transparency operates at the system level — it answers the question: “How does this tool work overall?”

For HR compliance, transparency requires being able to document: what data the model was trained on, what variables it uses as inputs, what fairness testing was performed at deployment, and how performance is monitored over time.

Gartner has identified AI transparency as a top-three governance priority for HR technology buyers, and that priority has only increased as regulatory scrutiny of automated employment decision tools has intensified.

Explainability (Explainable AI / XAI)

Explainability is the narrower ability to articulate why a specific decision was made for a specific individual at a specific moment. Where transparency answers “how does the system work,” explainability answers “why did this candidate receive this outcome.”

Explainability is a legal requirement in several jurisdictions. Under NYC Local Law 144, candidates must be notified when automated decision tools are used and provided access to bias audit results. Under emerging EU AI Act provisions (high-risk AI category includes employment decisions), candidates may have rights to human review of automated decisions.

Technical methods for achieving explainability include:

• SHAP (Shapley Additive Explanations): Assigns each feature a contribution score for a specific prediction
• LIME (Local Interpretable Model-Agnostic Explanations): Builds a locally interpretable model around each individual prediction
• Decision trees and rule-based overlays: Constrain model behavior to human-readable logic at the decision boundary

Black Box Model

A black box model is an AI system whose internal decision-making process is not interpretable by humans — inputs go in, outputs come out, but the mapping between them is opaque. Deep neural networks are the most common black box architecture in commercial AI applications.

Black box models in HR create significant audit risk: if a candidate is rejected by a system that cannot explain its reasoning, the employer has no defensible basis for that decision under adverse impact scrutiny or candidate challenge.

The practical guidance: treat black box status as a pass/fail criterion in AI vendor selection for any system that touches hiring decisions, not as a tradeoff to be managed.

Mitigation and Governance Terms

These terms describe the actions HR teams take to reduce bias and maintain accountability in AI-powered hiring processes.

Bias Audit

A bias audit is a structured evaluation of an AI system’s outputs across demographic groups to identify disparate outcomes and their potential causes. Bias audits are now legally required in some jurisdictions (NYC Local Law 144 mandates annual independent bias audits for covered automated employment decision tools) and represent best practice everywhere else.

A rigorous bias audit includes:

• Disaggregated performance analysis by race, gender, age, and other protected characteristics
• Application of multiple fairness metrics (not just the one the vendor prefers)
• Examination of proxy variable correlations in input features
• Review of training data composition and labeling methodology
• Documentation sufficient to withstand regulatory or legal challenge

Blind Screening

Blind screening is the practice of removing identifying information — names, photos, graduation years, addresses — from candidate records before automated or human review, to prevent conscious or unconscious bias from influencing evaluation.

Blind screening reduces bias at the human decision point but does not eliminate algorithmic bias caused by proxy variables in the remaining data. It is a necessary component of a fair screening process, not a sufficient one. For detailed implementation guidance, see our how-to on automated blind screening to reduce hiring bias.

Human-in-the-Loop (HITL)

Human-in-the-loop is a system design principle in which a human reviewer is required to approve, reject, or modify automated decisions before they are finalized — rather than allowing the AI to execute decisions autonomously. HITL is one of the primary governance mechanisms for managing bias risk in AI-powered hiring.

RAND Corporation research on AI governance has highlighted that HITL requirements are most effective when the human reviewer has access to the model’s reasoning (explainability) and is not subject to automation bias — the tendency to defer to algorithmic recommendations without independent evaluation.

Disparate Treatment

Disparate treatment is intentional discrimination — treating a candidate differently because of their membership in a protected class. It is legally distinct from disparate impact (unintentional adverse effect) and requires proof of intent. AI systems are less commonly implicated in disparate treatment claims, but intentional design choices (such as training a model specifically to exclude candidates from certain groups) could qualify.

Algorithmic Accountability

Algorithmic accountability is the organizational principle that humans — not algorithms — remain responsible for the outcomes of automated decisions. For HR teams, this means that deploying an AI screening tool does not transfer legal or ethical responsibility for hiring decisions to the vendor. The employer retains full accountability for every automated decision the tool makes on their behalf.

SIGCHI research on human-computer interaction in organizational decision-making has consistently found that accountability structures must be explicit and documented before AI deployment, not assumed from general organizational norms.

Related Terms Snapshot

Common Misconceptions

“Removing protected attributes from the model eliminates bias.”

This is the most persistent misconception in AI-powered HR. Removing race, gender, and age from the input data does not prevent those characteristics from influencing model outputs if proxy variables remain. A model that knows your zip code, your university, and the length of your employment gaps has enough information to infer protected characteristics with significant accuracy. Bias elimination requires proxy variable analysis, not just attribute removal.

“Our AI vendor has already handled bias — we don’t need to audit.”

Vendor bias testing is conducted on the vendor’s benchmark datasets, not your historical hiring data. Your applicant pool, job descriptions, and historical outcomes are unique. A model that passes fairness audits on generic benchmarks can still exhibit significant disparate impact on your specific population. You are the employer of record — vendor certification does not transfer your legal accountability.

“Bias only matters at the resume screening stage.”

Automated decisions with potential disparate impact occur at every stage: job description optimization (which candidates self-select to apply), screening, interview scheduling prioritization, assessment scoring, and offer generation. A bias audit must cover the full automated workflow, not just the most visible step.

“Higher accuracy means less bias.”

Overall model accuracy and fairness are independent dimensions. A model can achieve 90% accuracy on your historical dataset while producing significantly worse outcomes for underrepresented candidates — because the historical dataset itself contains underrepresentation. Disaggregated performance metrics by demographic group are required; aggregate accuracy metrics are insufficient.

Practical Implications for HR Teams

These definitions are not academic — they map directly to operational decisions every HR team using AI needs to make.

For teams exploring AI-powered screening, the sequence matters: automated candidate screening designed to reduce bias starts with clean data infrastructure, not with model selection. For teams evaluating different AI approaches, understanding the tradeoffs between AI parsing vs. Boolean search strategy for your ATS requires understanding what each approach optimizes for — and what bias vectors each introduces.

For teams looking to extend AI capability beyond basic screening, the six ways AI transforms your existing ATS beyond resume parsing framework provides a roadmap — but each transformation point introduces new fairness considerations that require metric selection and audit planning before deployment.

The foundational principle that governs all of this is the one established in the parent pillar on building the automation spine your ATS needs before AI deployment: deterministic automation — routing, scheduling, data capture — should be locked down first, because clean, consistent data infrastructure is the prerequisite for any AI system that performs equitably. AI deployed on chaotic manual data inherits that chaos and amplifies it at scale.