What Is an AI Candidate Screening Algorithm? A Plain-Language HR Glossary

AI candidate screening is not a single technology — it is a stack of distinct algorithms, each solving a different problem in the hiring pipeline. HR and recruiting leaders who cannot distinguish between machine learning, natural language processing, and deep learning are flying blind when evaluating tools, auditing outputs, or explaining hiring decisions to legal counsel. This glossary defines every core term, explains what it actually does in a recruiting context, and flags where each technology introduces risk.

For the broader strategic framework governing how these technologies should be sequenced and governed, start with the HR AI strategy and ethical talent acquisition pillar. This glossary drills into the vocabulary that strategy depends on.

Artificial Intelligence (AI) in Recruiting

Artificial intelligence is the broad category: any system that performs tasks normally requiring human cognition — pattern recognition, language interpretation, decision-making — through software. In recruiting, AI is the umbrella term under which every other technology in this glossary lives.

The critical distinction HR leaders must make is between rules-based automation and AI. A workflow that automatically moves a candidate to the phone-screen stage when they pass a knockout question is automation — it follows a deterministic rule. A system that scores a candidate’s resume against a job description using learned patterns is AI. Both are valuable. They are not interchangeable, and conflating them leads to misplaced expectations in both directions.

Gartner identifies AI in HR as one of the fastest-growing technology categories in the enterprise, yet adoption consistently outpaces the data infrastructure needed to support accurate outputs. The result is AI on top of chaos — which is why automation must precede AI deployment in any mature talent acquisition strategy.

Key Components of AI in Candidate Screening

  • Perception: The system takes in raw input — text, audio, video, structured fields.
  • Processing: Algorithms extract meaning, classify information, or generate predictions from that input.
  • Output: The system returns a score, a ranked list, a recommendation, or a generated response.
  • Feedback loop: Outputs feed back into the model (explicitly or implicitly), adjusting future behavior.

Machine Learning (ML)

Machine learning is the subset of AI where systems learn from data rather than following explicitly programmed rules. An ML model is trained on a historical dataset, identifies statistical patterns, and applies those patterns to new inputs to generate predictions.

In candidate screening, ML models are most commonly used to:

  • Score resumes against job requirements based on patterns from past successful hires
  • Predict candidate quality-of-hire, time-to-productivity, or 90-day retention probability
  • Rank applicant pools by estimated fit without human review of every application
  • Flag anomalies — résumés with inconsistencies or patterns associated with high attrition in historical data

The bias risk is direct and serious. ML models trained on historical hiring data inherit every bias embedded in past decisions. If your organization historically hired fewer women into technical roles, an ML model trained on that history will learn to de-prioritize female candidates for technical roles — and it will do so at scale, invisibly, and consistently. Harvard Business Review has documented this amplification effect extensively. Bias detection strategies for AI resume parsing covers the auditing process in detail.

Related Terms

  • Supervised learning: The model is trained on labeled examples (e.g., “this candidate was a strong hire; this one was not”) and learns to replicate those labels on new data.
  • Unsupervised learning: The model identifies natural clusters in data without predefined labels — used in talent segmentation and skills taxonomy building.
  • Reinforcement learning: The model learns through trial and feedback — less common in recruiting but emerging in conversational AI applications.

Deep Learning

Deep learning is a subset of machine learning that uses artificial neural networks with many layers — the “depth” refers to the number of processing layers, not complexity in a general sense. These multi-layered networks can extract features from raw, unstructured data without manual feature engineering.

Deep learning excels where standard ML struggles: processing images, audio, video, and long-form natural language. In recruiting, this manifests as:

  • Advanced resume parsing that understands context — differentiating the same word used in different industries or roles
  • Video interview analysis — transcribing speech, analyzing tone and pacing, flagging language patterns
  • Semantic matching — understanding that “managed a team of engineers” and “led software development staff” are equivalent, even without shared keywords

Deep learning requires large training datasets to perform well. Organizations with thin historical hiring data will see poor deep learning outputs. It also requires rigorous bias auditing — the complexity of the model makes it harder to explain why a particular candidate received a particular score, creating “black box” compliance exposure.

For a ground-level look at how these models perform in practice, see how to evaluate AI resume parser performance.

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 most directly relevant AI technology for the majority of recruiting workflows, because recruiting runs on text.

NLP is what powers:

  • Resume parsing: Extracting structured data — skills, job titles, dates, education, certifications — from free-form resume text
  • Job description analysis: Identifying biased language, vague requirements, or misaligned terminology that reduces application quality
  • AI recruiting chatbots: Interpreting candidate questions and generating contextually appropriate responses
  • Candidate communication analysis: Assessing written responses in application forms or asynchronous assessments for relevance and professionalism

NLP accuracy is directly dependent on input quality. Poorly structured job descriptions produce poor NLP matching results. Resumes with non-standard formatting, embedded tables, or graphics-heavy layouts defeat most NLP parsers. This is why optimizing job descriptions for AI candidate matching is not cosmetic work — it is foundational to NLP performance.

Key NLP Concepts in HR Tech

  • Named entity recognition (NER): Identifies and classifies entities — “Python,” “Seattle,” “5 years,” “MBA” — in unstructured text.
  • Semantic similarity: Measures how close two pieces of text are in meaning, not just in shared words.
  • Sentiment analysis: Classifies text as positive, negative, or neutral — used in candidate communication monitoring and employer brand analysis.
  • Text classification: Assigns documents or text segments to predefined categories — “qualified,” “over-qualified,” “missing required credential.”

For a deeper look at where NLP parsing creates real candidate insights versus false positives, see how AI resume parsing unlocks deeper candidate insights.

Computer Vision

Computer vision is the AI field that enables systems to interpret visual information — images and video — the way a human observer would. In recruiting, computer vision is primarily associated with video interview analysis tools that claim to assess candidate engagement, emotional state, or behavioral patterns from facial expressions and body language.

This is the highest-risk category in the HR AI stack. Computer vision applications in hiring face:

  • Active regulatory scrutiny in the EU AI Act and US state-level AI hiring laws (Illinois, New York City)
  • Documented concerns from researchers about accuracy variations across demographic groups
  • Significant candidate experience backlash when disclosed
  • Thin evidence base for predictive validity in hiring contexts

The practical guidance: do not deploy computer vision tools in candidate assessment without explicit legal review against applicable jurisdictions. The technology is not inherently unusable, but the compliance burden is high and the validated benefit is low for most organizations. See AI resume screening compliance and fairness for the full regulatory landscape.

Predictive Analytics

Predictive analytics applies statistical models — often ML-based — to historical data to forecast future outcomes. In talent acquisition, predictive analytics typically outputs a probability score attached to a candidate or a position.

Common predictive analytics applications in recruiting include:

  • Quality-of-hire prediction: Scoring applicants on estimated performance in role, based on patterns from high-performing current employees
  • Retention prediction: Flagging candidates whose profiles correlate with short tenure in historical data
  • Time-to-fill forecasting: Estimating how long a requisition will take to fill based on role, location, and market data
  • Candidate drop-off prediction: Identifying which candidates in the pipeline are likely to disengage before offer

The integrity of every predictive model depends on the integrity of the historical data used to train it. Asana’s Anatomy of Work research consistently documents how much time recruiting teams spend on work that is never captured in structured data — which means the training datasets for most predictive models are systematically incomplete. Garbage in, garbage prediction out.

For the metrics that validate whether predictive analytics are actually delivering, see KPIs for measuring AI talent acquisition success.

Applicant Tracking System (ATS) and AI Integration

An applicant tracking system is the database layer that manages candidate records, requisitions, pipeline stages, and communication logs. It is not inherently an AI system — it is the data infrastructure that AI algorithms operate on top of.

AI integration with an ATS typically works in one of three patterns:

  • Native AI: The ATS vendor has built AI capabilities directly into the platform — scoring, ranking, or parsing happen inside the system.
  • Integrated AI layer: A third-party AI tool connects to the ATS via API, processes data, and writes results back to candidate records.
  • Automation middleware: A workflow automation platform connects the ATS to AI tools and other systems, orchestrating data flow between them.

The quality of AI output in any of these patterns is capped by the quality of ATS data. Inconsistent field mapping, free-text notes in structured fields, and incomplete candidate records all degrade algorithmic performance. This is why data hygiene and process standardization precede AI deployment in every effective implementation. See boosting ATS performance with AI resume parsing integration for the implementation sequence.

Common Misconceptions About AI Screening Algorithms

Misconception 1: AI screening is objective. AI outputs are only as objective as the data they were trained on. Models trained on biased historical decisions produce biased predictions — with the added problem that the bias is harder to see and challenge than a human recruiter’s gut feeling. Objectivity requires audited training data, not just an algorithm. The AI resume parsing myths and facts satellite addresses this in detail.

Misconception 2: More AI means less bias. More AI means more scale — bias included. McKinsey Global Institute research on AI adoption consistently finds that organizations that deploy AI without governance frameworks amplify existing process problems rather than eliminating them.

Misconception 3: AI replaces recruiter judgment. The correct framing is AI handles volume so recruiters can apply judgment where it matters. AI screening is not a replacement for human assessment — it is a filter that narrows the field to where human attention is most productive. Forrester’s research on talent acquisition technology consistently positions AI as augmenting recruiter capacity, not substituting for it.

Misconception 4: Any AI tool works out of the box. No AI tool performs well on a new organization’s data without a configuration and calibration period. The tool’s training came from other organizations’ data. Your hiring patterns, your job descriptions, your candidate populations — these require adaptation. Budget for it.

Related Terms Quick Reference

  • Algorithm: A set of rules or instructions a computer follows to accomplish a task or make a calculation. Every AI model is implemented as one or more algorithms.
  • Training data: The historical dataset used to teach an ML model what patterns to recognize. Its quality and representativeness determine model accuracy.
  • Model: The output of training an algorithm on data — a mathematical representation of the patterns the algorithm identified, ready to apply to new inputs.
  • Feature: An individual measurable property or characteristic used as input to an ML model (e.g., years of experience, skills listed, degree type).
  • Inference: The process of applying a trained model to new data to generate a prediction or classification.
  • Bias audit: A structured review of an AI model’s outputs across demographic groups to identify disproportionate impact. Required before deployment in most regulated hiring contexts.
  • Black box: An AI model whose internal logic cannot be easily inspected or explained — a compliance and legal risk in hiring where decision rationale must be articulable.
  • Explainability: The degree to which a model’s outputs can be traced back to specific inputs and logic — a critical requirement for hiring tools under equal employment opportunity obligations.

Why This Vocabulary Matters Operationally

SHRM research documents that HR technology adoption failures trace most often to misaligned expectations between buyers and vendors — not to the technology itself. HR leaders who cannot interrogate vendor claims about algorithm type, training data, bias auditing, and explainability will accept whatever they are told. That is a compliance risk and a budget risk simultaneously.

The vocabulary in this glossary gives recruiting leaders the specific questions to ask:

  • What type of algorithm powers your candidate scoring feature?
  • What dataset was the model trained on, and how recent is it?
  • Has the model been audited for disparate impact across protected classes, and can you share the results?
  • If a candidate challenges a screening decision, what explanation can I provide?
  • How does the model perform on our specific job categories, and how do we recalibrate if it doesn’t?

These are not technical questions — they are procurement and governance questions that any vendor of a legitimate AI screening tool should answer without hesitation. The ones who cannot or will not tell you everything you need to know about the risk profile of what you are buying.

For the cost implications of getting these decisions wrong — and the ROI of getting them right — see the hidden costs of manual screening versus AI analysis.

This glossary is part of the HR AI strategy and ethical talent acquisition content series. The next step after building this vocabulary is assessing your organization’s readiness to actually deploy these technologies — covered in the recruitment AI readiness assessment guide.