Because in recruitment AI, the same word can mean very different things, and those differences determine whether a tool holds up under legal scrutiny or fails at the first audit.

Vendors, analysts, and practitioners often use the same words to mean different things, and different words to mean the same thing. "AI-powered" appears on almost every product in the market. "Fair" and "explainable" are claimed by tools with fundamentally different architectures. "Compliant" can mean anything from a data-processing agreement to a full audit trail.

And the consequences are real. When buyers can't distinguish between approaches, procurement decisions get made on demos and pricing rather than on what actually determines whether a tool will hold up under legal scrutiny, produce consistent results, or treat candidates fairly at scale. For recruiters responsible for shortlisting thousands of candidates, the wrong tool (or even the right tool misunderstood) creates regulatory and reputational exposure.

The terms below are organized to be practically useful: what the concept means, why it matters in a hiring context, and what to ask when a vendor uses it.

These four terms define how AI systems are built and understanding them tells you whether a hiring tool can actually score candidates fairly and consistently, or whether it's guessing.

Large language model (LLM): A type of AI trained on vast amounts of text data to generate human-like responses. LLMs power many conversational tools, including general-purpose chatbots. In hiring, an LLM can conduct a natural-sounding interview conversation. The limitation: LLMs are probabilistic which means the same input can produce different outputs at different times. This makes them poorly suited to consistent, auditable candidate scoring without an additional assessment layer on top.

Deterministic model: An AI model in which the same input always produces the same output. Deterministic scoring is essential for legally defensible hiring: if two candidates give identical answers, they receive identical scores – every time, and with a full explanation of why. This is the architectural difference between a tool that can be audited and one that cannot. When evaluating any AI screening product, ask specifically whether its scoring layer is deterministic or probabilistic.

Generative AI: AI that creates new content, such as text, images, audio, in response to a prompt. Generative AI is useful for drafting interview questions, writing job descriptions, or powering conversational interfaces. It is not, by itself, a method for assessing candidates. Conflating generative capability with assessment accuracy is one of the most common sources of confusion in vendor pitches.

Natural language processing (NLP): The branch of AI that enables machines to understand, interpret, and respond to human language. NLP underpins both conversational interview tools and response analysis. Strong NLP is what allows a screening interview to feel natural across 30+ languages. But it is not the same as a strong scoring methodology.

These five terms describe how AI evaluates candidates, and knowing them helps you tell apart tools built on rigorous methodology from those that borrow the language without the substance.

Structured interview: An interview in which every candidate is asked the same questions, in the same order, assessed against the same criteria. Decades of research in organizational psychology support structured interviewing as significantly more predictive of job performance than unstructured interviews. AI-powered structured interviews automate this consistency at scale, meaning every applicant gets the same experience regardless of who reviewed their CV first or when they applied.

Competency-based assessment: Evaluation of candidates against specific, defined competencies relevant to the role, rather than CV credentials or gut feel. Competency-based screening measures what candidates can actually do, not just what they claim or what their educational background signals. Combined with structured questions, it is the basis for predicting hiring success with significantly greater accuracy than traditional methods.

Skills-based hiring: A hiring philosophy that prioritizes demonstrated ability over proxies like degree requirements or prior job titles. AI screening tools that use competency-based assessment support skills-based hiring by surfacing candidates who might not pass traditional CV filters. NSS Group, for example, saw a 50% increase in hires from candidates who would never have passed traditional CV screening after deploying Hubert’s structured AI interviews.

Knockout criteria: A threshold requirement that, if not met, automatically disqualifies a candidate from progressing. Knockouts are legitimate in hiring when the criterion is genuinely role-relevant and clearly defined, for example, a legal right-to-work requirement. They become problematic when they encode bias (using a degree as a proxy for capability) or when candidates are not told why they were screened out. Any AI tool applying knockouts should be able to name the specific criterion, not return a generic rejection.

Shortlist: The reduced set of candidates passed forward for human review after initial screening. In AI-assisted hiring, a shortlist should be scored, auditable, and accompanied by a clear explanation of how each candidate was assessed. "Shortlist" is sometimes used loosely to describe any ranked output, but the meaningful question is whether the shortlist is legally defensible: can the recruiter explain every inclusion and exclusion if challenged?

Fairness in AI hiring isn't just a value. It's a technical property that can be measured, audited, and required by law.

Adverse impact: A statistically significant difference in selection rates between protected groups (by gender, ethnicity, age, etc.) that disadvantages one group over another. Adverse impact can occur in AI hiring tools when training data reflects historical biases in hiring decisions. Recruiters evaluating AI tools should ask whether adverse impact testing has been conducted, and on what data, in what contexts.

Algorithmic bias: Systematic error in an AI system that produces unfair outcomes for specific groups. Bias can enter at the data collection stage, the model training stage, or the deployment stage. It is not inherently avoided by using AI. The safeguard is regular auditing, transparent methodology, and human oversight of final decisions.

Explainability: The degree to which an AI system can account for how it reached a particular output. In a hiring context, explainability means a recruiter can see, for any candidate score, exactly which response contributed to which score. No black box, no opaque ranking. Explainability is both a candidate right and a legal requirement under emerging regulation. It is also the foundation of any meaningful appeal or audit process.

Human oversight: The principle that final hiring decisions must remain with a human being, not delegated entirely to an automated system. Under the EU AI Act, high-risk AI systems (which includes employment-related AI) must support meaningful human review. "Meaningful" is the operative word: oversight that exists on paper but is bypassed in practice does not satisfy the requirement. Well-designed AI hiring tools augment recruiters; they do not replace the decision.

These four terms are the ones most likely to appear in a procurement process, a vendor contract, or a regulatory inspection – and confusing them creates real exposure.

EU AI Act: The European Union's comprehensive regulatory framework for AI systems, which came into force in 2024. Employment-related AI, including tools used for recruitment, CV screening, and candidate assessment, is classified as high-risk under the Act. High-risk systems must meet requirements including transparency, human oversight, data governance, and audit logging. Tools built for EU AI Act compliance from the ground up are structurally different from those retrofitting compliance onto an existing product.

Legally defensible: A hiring process or decision that can withstand challenge from a rejected candidate, a regulator, or an employment tribunal because it is consistent, documented, and grounded in objective criteria. "Legally compliant" means meeting minimum requirements; "legally defensible" means the process holds up under scrutiny. The distinction matters: compliance is a floor, not a standard.

Audit trail: A complete, time-stamped record of decisions made in a process, including the inputs, the criteria applied, and the outputs. In AI hiring, an audit trail means being able to reconstruct exactly how any candidate was scored, by what model, on what version, at what point in time. Audit trails are essential for regulatory inspections, internal reviews, and candidate appeals.

Data minimization: The principle that only data strictly necessary for a defined purpose should be collected and retained. In recruitment AI, this means not storing candidate responses beyond what is needed for the hiring decision and giving candidates control over their own data. Data minimization is both a GDPR obligation and, for many candidates, a trust signal.

The following four terms describe how candidates actually experience AI hiring and they're increasingly the metrics that determine whether a tool is fit for enterprise use.

Candidate experience: The overall quality of a candidate's interaction with an employer's hiring process, from first contact through to offer or rejection. In high-volume hiring, candidate experience frequently suffers because speed pressures lead to generic communications, long waits, and no feedback. AI screening, when designed correctly, can improve candidate experience at scale: every applicant gets a structured interview, a clear process, and meaningful feedback regardless of outcome.

Completion rate: The percentage of candidates who begin an AI screening interview and complete it. Completion rate is a direct signal of how candidates experience the process: a low rate suggests the interview is too long, too confusing, or feels unfair. Across Hubert deployments, the average completion rate is 96%, a figure that reflects both interview design and the degree to which candidates feel the process respects their time.

Asynchronous interview: An interview conducted at a time chosen by the candidate, rather than scheduled in real time with a recruiter. Asynchronous AI interviews remove one of the most significant access barriers in hiring: availability. Candidates in different time zones, working multiple jobs, or with caring responsibilities can complete a screening interview when it suits them. Across ManpowerGroup deployments, 60% or more of interviews are completed outside traditional office hours.

Candidate feedback: Information returned to a candidate about how they performed or why a decision was made. Feedback is consistently one of the most requested elements of the hiring process by candidates and one of the most commonly withheld. Platforms committed to fair hiring treat feedback not as an optional extra but as a design requirement. The absence of feedback after an AI-screened rejection is not a neutral outcome; it is a fairness failure.

Knowing these terms changes the questions you ask, and better questions lead to better procurement decisions, stronger internal conversations, and more defensible hiring processes.

"Is your scoring explainable?" is a better question than "Is your AI fair?" because explainability is measurable and fair is not. "Is the assessment layer deterministic?" separates tools that can be audited from those that cannot. "What does your audit trail cover?" tells you whether a vendor has built for compliance or is retrofitting it.

The terminology also helps inside your organization. When briefing legal, finance, or the board on AI hiring adoption, precision matters. "We're using an AI that produces scored, auditable shortlists assessed by deterministic models" is a different conversation from "we're automating screening" and it is a more defensible one.

Hubert's approach is built on structured interviewing science, deterministic assessment, and full explainability. Every score is tied to a specific response, every shortlist auditable, and every deployment designed to meet EU AI Act requirements for high-risk employment AI. Trusted by ManpowerGroup, Securitas, Coop, and other leading enterprises, Hubert predicts hiring success with 5x greater accuracy than traditional methods and delivers 80% faster time-to-hire across 30+ languages and ATS integrations.

If you're evaluating AI hiring tools and want to see how these principles apply in practice, book a demo with the Hubert team.