AI & ML Glossary: Essential Concepts for HR and Recruiting

AI terminology moves fast, and the gap between what vendors promise and what HR teams actually understand is where failed implementations live. This glossary defines the core AI and machine learning concepts that recruiting professionals and HR leaders encounter daily — not as abstract computer science, but as operational vocabulary tied to real hiring decisions. It is a companion reference to the HR AI strategy roadmap for ethical talent acquisition, and every term below maps to a decision or risk you will encounter when building or auditing your talent acquisition stack.

Use this glossary to evaluate vendor claims, structure internal training, and build the shared language your team needs before purchasing or configuring any AI-adjacent tool.

Artificial Intelligence (AI)

Artificial Intelligence is the broad capability of a computer system to perform tasks that would otherwise require human judgment. The term is an umbrella, not a specification. It covers everything from a rule-based chatbot that routes candidate questions to a sophisticated model that predicts 90-day offer acceptance probability.

In HR and recruiting, AI manifests in several layers:

  • Task automation with intelligence: Scheduling interviews, routing applications, triggering follow-up messages based on candidate stage.
  • Pattern recognition: Identifying which resume signals correlate with successful hires in a specific role family.
  • Generative output: Drafting job descriptions, outreach messages, or interview question sets from a brief prompt.
  • Prediction: Forecasting time-to-fill, offer acceptance rate, or 12-month retention likelihood.

The critical distinction HR leaders must internalize: AI is not a monolith. When a vendor says their platform “uses AI,” that claim is nearly meaningless without knowing which layer — and what data it runs on. See the common misconceptions about AI resume parsing for a practical breakdown of how this confusion costs teams time and budget.

Why It Matters

Gartner identifies AI in HR as one of the highest-impact technology investments available to CHROs — and simultaneously one of the highest-risk when deployed without a structured data and process foundation beneath it.

Machine Learning (ML)

Machine Learning is a subset of AI in which a system improves its performance on a task by learning from data rather than following explicitly programmed rules. Instead of a developer writing every decision path, an ML model is trained on historical examples and learns to generalize patterns to new inputs.

Two ML modes dominate HR applications:

Supervised Learning

The model trains on labeled historical data — resumes tagged as “progressed to interview” or “screened out,” employees tagged as “high performer” or “churned within 12 months.” It learns to predict labels for new, unlabeled inputs. Most AI resume ranking tools, performance prediction models, and attrition risk scores use supervised learning.

Unsupervised Learning

The model finds structure in unlabeled data — clustering employees into cohorts by behavior patterns, or grouping job descriptions by required competency profiles without pre-defined categories. Used in workforce segmentation, talent pool mapping, and compensation benchmarking analytics.

Why It Matters

Every ML model is only as good as its training data. McKinsey research identifies poor data quality as the primary reason AI initiatives underperform. In recruiting, this means years of biased or incomplete ATS records directly degrade any model trained on them. This is why the recruitment AI readiness assessment begins with a data audit, not a vendor evaluation.

Natural Language Processing (NLP)

Natural Language Processing is the branch of AI that enables computers to read, interpret, and generate human language. It is the reason a resume parser can extract “led cross-functional teams of 12” as a leadership signal rather than just a string of characters.

NLP capabilities in recruiting include:

  • Resume parsing: Converting unstructured candidate documents into structured data fields — skills, job titles, tenure, education level.
  • Synonym mapping: Recognizing that “software engineer,” “developer,” and “SWE” describe the same role family.
  • Sentiment analysis: Classifying candidate feedback, exit interview text, or employee pulse survey responses as positive, negative, or neutral — and detecting emotional nuance beneath the surface classification.
  • Job description analysis: Identifying gendered language, unrealistic requirement stacking, or keyword gaps that reduce application rates from qualified candidates.

For a detailed look at how NLP quality separates high-performing parsers from commodity tools, see the guide on evaluating AI resume parser performance.

Why It Matters

NLP quality is the single largest differentiator in resume parsing accuracy. A parser that relies on keyword matching without NLP context will screen out a qualified candidate who uses a non-standard title — a failure mode that compounds across thousands of applications per quarter.

Generative AI

Generative AI is a category of AI models that produce new content — text, images, audio, or structured data — by learning patterns from large training datasets. Unlike discriminative models that classify or rank existing inputs, generative models create outputs that did not previously exist.

In HR and recruiting, generative AI is most commonly applied to:

  • Job description drafting: Producing role-specific JDs from a brief prompt or competency list, with iteration based on feedback.
  • Candidate outreach personalization: Generating individualized InMail or email sequences at volume without manual writing for each record.
  • Interview question generation: Creating structured, role-relevant interview guides that map to specific competencies.
  • Candidate summary synthesis: Condensing a recruiter’s notes, resume data, and assessment scores into a single briefing document for a hiring manager.

Key Limitation

Large language models — the technology underlying most generative AI tools — do not retrieve facts. They generate statistically plausible text. Every generative output in a hiring context requires human review for factual accuracy, compliance with EEO language standards, and alignment with actual role requirements.

Why It Matters

Harvard Business Review research on generative AI in professional services consistently finds that the productivity gains are real — and that unchecked generative output introduces new error categories that did not exist in manual workflows. The efficiency gain and the new review requirement must both be staffed.

Predictive Analytics

Predictive analytics applies statistical models and machine learning to historical data in order to forecast future outcomes. In talent acquisition, it shifts decision-making from reactive to anticipatory.

Common predictive analytics applications in HR:

  • Time-to-fill forecasting: Predicting how long a requisition will take based on role type, location, market conditions, and historical pipeline velocity.
  • Offer acceptance probability: Estimating the likelihood a candidate will accept an offer given compensation, commute, and competing offer signals.
  • Attrition risk scoring: Identifying employees at elevated flight risk based on engagement signals, tenure, manager changes, and compensation trajectory.
  • Quality-of-hire prediction: Correlating pre-hire assessment data with post-hire performance ratings to refine future screening criteria.

For the KPI framework that makes predictive analytics measurable and accountable, see the guide to KPIs for AI talent acquisition success.

Why It Matters

Deloitte’s Global Human Capital Trends research identifies predictive workforce analytics as a top-five HR capability gap. Organizations that deploy predictive models without first establishing clean baseline data produce predictions that are worse than informed human judgment — because the model has learned from flawed history.

Algorithmic Bias

Algorithmic bias is the systematic production of skewed or discriminatory outputs by an AI model, typically traceable to patterns embedded in its training data. It is not a software bug — it is a data and design problem.

Bias enters hiring AI through several pathways:

  • Historical hiring data: If past hiring decisions over-represented certain demographic groups in specific roles, a model trained on those decisions learns to replicate the pattern.
  • Proxy variables: Zip code, educational institution, gap years, and extracurricular affiliations can all correlate with protected characteristics without being explicitly included as inputs. A model may weight them heavily and produce disparate impact without ever “seeing” race or gender.
  • Label bias: If “successful hire” is defined as “stayed 12+ months and received a high manager rating,” and those ratings are themselves biased, the model learns a biased definition of success.

Understanding the mechanics of algorithmic bias is the prerequisite for meaningful audit practice. The detailed guide on bias detection and mitigation strategies for AI resume screening covers adverse impact analysis, disparate impact testing, and audit cadence in depth.

Why It Matters

SHRM and SIGCHI research both document that AI bias in hiring is not hypothetical — it has been demonstrated in resume screening, video interview scoring, and candidate ranking tools at scale. Regulatory frameworks are catching up: auditing requirements are now law in several jurisdictions and expanding.

Explainable AI (XAI)

Explainable AI refers to AI systems designed to produce outputs accompanied by human-readable rationales — explanations that describe why the model reached a specific conclusion.

In HR, XAI is both an ethical standard and an emerging legal requirement. When a candidate is screened out by an automated system, the organization must be able to articulate the basis for that decision in terms that:

  • Do not reference protected characteristics, even indirectly.
  • Are consistent across candidates evaluated under similar criteria.
  • Can be reviewed by a human decision-maker before the screening result is acted upon.

“Black box” models — those that produce a ranking or score without any interpretable rationale — are increasingly incompatible with regulated hiring environments. Several U.S. jurisdictions now require or are moving toward mandatory explainability disclosures for automated employment decision tools.

Why It Matters

Forrester research on AI governance identifies explainability as the capability most frequently absent from enterprise AI deployments — and the one most frequently cited in regulatory enforcement actions. Purchasing a high-performing model that cannot explain itself is a compliance liability.

For the compliance framework that governs explainability in practice, see the AI resume screening compliance guide.

Large Language Model (LLM)

A large language model is a type of generative AI trained on massive text corpora to predict the next token in a sequence — which, at scale, produces coherent, contextually aware language generation.

LLMs power the generative AI tools entering HR stacks at speed: job description drafters, candidate summarizers, interview question generators, and AI-assisted offer letter composers. They are not databases — they do not retrieve facts from a stored record. They generate text that is statistically consistent with their training data.

Practical implications for HR teams using LLM-powered tools:

  • LLM outputs must be reviewed for factual accuracy before use in any candidate-facing communication.
  • LLMs can generate confident-sounding but legally non-compliant language — EEO review remains mandatory.
  • LLMs trained on internet-scale data may reflect cultural or demographic biases present in that training corpus.
  • Prompt quality determines output quality — teams using LLM tools need prompt engineering literacy, not just access to the interface.

Automation vs. AI: The Distinction That Determines Sequence

Automation executes deterministic rules. AI applies probabilistic judgment. They are not synonyms, and treating them as such is the most common cause of failed HR technology implementations.

The operational distinction:

  • Automation: If a candidate completes a phone screen, trigger a scheduling link. Always. No exceptions. No model needed.
  • AI: Given this candidate’s profile, experience signals, and assessment results, what is the probability they will perform in the top quartile at 12 months? A fixed rule cannot answer this — a trained model can approximate it.

The strategic implication is sequencing. Automation must underpin AI — clean, consistent, rules-based processes produce the reliable data that AI models require to function accurately. Deploying AI directly onto a broken or manual process produces what practitioners call “AI on top of chaos”: sophisticated outputs built on unreliable inputs.

For a practical look at how automation and AI complement each other across the recruiting workflow, see the overview of ways AI and automation boost HR efficiency.

Key Components of an AI-Ready HR Stack

Understanding individual terms is necessary but insufficient. The following components must work together for any AI deployment to produce reliable, compliant, and defensible outputs:

  • Structured data layer: ATS, HRIS, and assessment data must be consistently formatted, complete, and historically clean before any model is trained on it.
  • Automation spine: Deterministic workflows that handle routing, scheduling, follow-up, and data sync — reducing the noise and variability that degrades AI inputs.
  • Model governance: Documented training data sources, audit cadence, bias testing methodology, and explainability standards.
  • Human review checkpoints: Defined moments in the pipeline where human judgment reviews, overrides, or confirms AI outputs — particularly for screening decisions that affect protected groups.
  • Feedback loop: Post-hire outcome data flowing back into the model to enable continuous improvement and bias correction over time.

Related Terms

Applicant Tracking System (ATS)
The database and workflow platform that manages the recruiting pipeline from application to hire. The primary source of training data for most HR AI models — which makes ATS data quality a top-tier strategic priority.
Adverse Impact Analysis
A statistical assessment that compares hiring outcomes across demographic groups to identify whether a selection procedure disproportionately disadvantages a protected class. Required by EEOC guidelines and triggered by automated screening tools in several jurisdictions.
Structured Interview
An interview format in which every candidate for a role receives the same questions in the same order, with responses evaluated against pre-defined scoring criteria. Structured interviews produce more reliable data for ML models and are more legally defensible than unstructured conversations.
Skills-Based Hiring
A talent acquisition philosophy that prioritizes demonstrated competencies over credential proxies (degree, institution, title). AI skills-matching tools are the primary enabler of skills-based hiring at scale.
Semantic Search
A search capability that interprets the meaning and intent of a query rather than matching exact keywords. In recruiting, semantic search allows a recruiter to surface candidates who match a role’s competency requirements even when their resume language differs from the job description’s phrasing.
Model Drift
The gradual degradation of a model’s predictive accuracy as the real-world environment changes and diverges from the conditions represented in the training data. HR AI models require periodic retraining as labor markets, role requirements, and workforce demographics shift.
Prompt Engineering
The practice of designing and refining input instructions to generative AI systems to produce more accurate, relevant, and compliant outputs. An emerging competency requirement for HR professionals using LLM-powered tools.

Common Misconceptions

Misconception 1: “AI is objective because it removes human bias.”

AI removes explicit human bias from individual decisions. It does not remove bias — it encodes the aggregate bias present in historical training data and applies it at scale. An AI system can produce systematically discriminatory outputs without any individual human intending discrimination. The solution is structured bias auditing, not the assumption of objectivity.

Misconception 2: “More data always produces a better model.”

Volume does not correct for quality. A model trained on ten years of biased, incomplete, or inconsistently labeled hiring records will produce a biased, unreliable model — regardless of how many records are included. The APQC data quality research confirms that organizations systematically overestimate the quality of their own data. A pre-implementation data audit is not optional.

Misconception 3: “AI can replace recruiter judgment.”

AI can augment recruiter judgment at scale — handling the screening volume that would otherwise require weeks of manual review. It cannot replace the relationship intelligence, contextual reading, and ethical accountability that qualified recruiters bring to offer negotiations, candidate experience, and hiring manager alignment. The evidence-based framing from Harvard Business Review: AI performs best as a decision support system, not a decision replacement system.

Misconception 4: “Automation and AI are the same investment.”

They are not. Automation is deterministic, relatively low-cost, and immediately measurable. AI is probabilistic, requires data infrastructure investment, and produces value on a longer time horizon. Conflating them leads teams to either under-invest in automation (assuming AI will handle everything) or over-invest in AI (before the process foundation is ready). Both failure modes are common.

Putting the Vocabulary to Work

This glossary is a starting point, not an endpoint. The real test of AI fluency in HR is the ability to evaluate a vendor’s claims, audit a model’s outputs, and make sequencing decisions about where automation ends and AI begins.

The HR AI strategy roadmap provides the decision framework that connects these definitions to implementation priorities. For teams ready to assess their current state before committing to any AI tool, the recruitment AI readiness assessment is the logical next step.

Vocabulary shapes strategy. Teams that cannot define these terms precisely cannot evaluate the tools claiming to use them — and cannot protect themselves when those tools produce outputs that require defense.