REFERENCE — GLOSSARY

The AI interviewing
glossary

20 terms that come up whenever a hiring team evaluates AI interviewing — defined plainly, and defined for the field rather than for us. Where Lantern is relevant, it says so and stops there.

Adaptive questioning

Adaptive questioning is an interviewing technique in which each follow-up is chosen in response to the answer just given, rather than read from a fixed script. The same opening question branches differently depending on what the previous answer left open — who else was involved, what the candidate would have done unaided, what the decision cost. It is the property that separates an interview from a questionnaire.

In an automated interview, adaptivity is also what makes the record worth auditing: the follow-up chosen, and the reason it was chosen, belong in the interview log alongside the answer. Lantern's interview engine is adaptive in this sense — the same question with three different answers leads to three different follow-ups.

See how a branching interview works →

AI interviewer

An AI interviewer is software that conducts a live, two-way interview with a candidate — usually by voice — and scores the answers against a rubric the hiring team defined in advance. It is not a recording tool and not a resume parser: it listens to each answer, asks the follow-up that answer earned, and returns a structured score with the transcript evidence behind it. If a system cannot change its next question based on the answer it just heard, it is not interviewing.

Lantern is an AI interviewer. It runs live, adaptive voice interviews in 40+ languages at any hour, scores every answer against the hiring team's own interview criteria with transcript-linked evidence, and hands the team a ranked shortlist by morning. It never rejects anyone — humans make every final decision.

A working definition for hiring teams →

AI proctoring

AI proctoring is the automated supervision of a remote test or interview — monitoring video, audio, screen activity, and browser state — to detect impersonation, unauthorised assistance, or other integrity breaches. It answers a different question from interviewing: not whether someone can do the work, but whether it is really them sitting there.

Proctoring earns its cost where selection is scarce and the incentive to cheat is high — admissions, licensing, certification. AdmissionGuard, built by Lantern's team for admissions, combines AI-proctored aptitude tests under secure browser lockdown, written assessments checked for generated text and plagiarism, and structured video interviews with behavioural analysis.

AdmissionGuard, for admissions offices →

Applicant tracking system (ATS)

An applicant tracking system (ATS) is the system of record for hiring: it stores every applicant, tracks their stage in each open requisition, and holds the notes, scores, and correspondence attached to them. Because the workflow lives there, the ATS is the integration point any screening tool has to meet — a tool that does not write back creates a second, divergent record.

Lantern plugs into 30+ ATS and HRIS systems — including Greenhouse, Workday, Lever, Ashby, iCIMS, and SmartRecruiters — typically in one day, and syncs shortlists, transcripts, and scores back automatically. Anything else connects over open API, webhook, or a scheduled export.

Audit trail

An audit trail is a complete, exportable record of how a decision was reached — every input, step, and output preserved in a form a third party can review after the fact. In hiring, it is what lets a regulator, a works council, or an internal reviewer reconstruct why one candidate advanced and another did not, months later and without the original interviewer in the room.

For a high-risk AI system under the EU AI Act, logging is an obligation rather than a design choice. In Lantern, every question, answer, and score is exportable for regulators and internal review.

What sits behind every number →

Bias drift monitoring

Bias drift monitoring is the continuous measurement of a scoring system's outcomes across demographic groups, to detect disparities that appear or widen after deployment. It treats fairness as an ongoing measurement problem rather than a one-time pre-launch check, because the applicant population, the roles, and the calibration all keep moving once a system is live.

The reason to measure continuously is so that drift gets flagged rather than absorbed. Lantern measures and reports scoring drift across demographic groups on an ongoing basis.

How bias monitoring works →

Candidate satisfaction score

A candidate satisfaction score is the share of candidates who report a positive experience of a hiring process, usually collected by survey right after an interview or at the point of decision. It is the counterweight to efficiency metrics: a process can get faster and cheaper while quietly getting worse for the people going through it, and only this number shows that.

97% of Lantern candidates finish satisfied with how the interview went. Candidates consistently describe it as less stressful than a live panel — they can interview at any hour, in the language they think in, and hear back quickly.

Candidate experience →

Code-switching

Code-switching is the practice of alternating between two or more languages or dialects within a single conversation, often mid-sentence. It is ordinary bilingual speech rather than error or imprecision, and in multilingual hiring markets it is simply how many candidates talk about technical work.

Interview tools that force one language per session penalise it, and the penalty falls on exactly the candidates the market is shortest of. Lantern supports code-switching: a candidate can open in their own language, drop into English for the technical part, and come back — Lantern follows.

SEE ALSO Language drift · AI interviewer

Multilingual interviews and code-switching →

Deepfake detection

Deepfake detection is the identification of synthetic or manipulated media — a generated face, a cloned voice, an answer scripted by a language model — presented as a live human. In remote interviewing it is an integrity control, not a hiring judgment: it establishes who is speaking, never whether they are any good.

The design question that matters is what happens after a flag. Lantern detects deepfakes, cloned voices, and AI-generated scripts in real time, and a flag is never an auto-rejection: the interview goes to the hiring team with the evidence attached, and a human makes the call.

SEE ALSO AI proctoring · Human-in-the-loop

Integrity safeguards in full →

EEO-aware scoring

EEO-aware scoring is candidate evaluation designed around equal employment opportunity law: it scores job-related evidence and deliberately keeps protected characteristics and their proxies — names, accents, photographs, schools — out of the scoring path. Aware is the operative word; the attributes are known and excluded by design rather than assumed absent because nobody looked.

Exclusion only holds if it can be checked. Lantern scores against EEO-aware rubrics that read evidence, not accents, names, or where someone went to school, and every score is auditable line by line.

Fairness and bias monitoring →

EU AI Act high-risk system

An EU AI Act high-risk system is an AI system placed in a regulated category by the Act — a list that includes systems used for recruitment, candidate selection, and evaluation — and which therefore carries obligations for risk management, data governance, technical documentation, logging, human oversight, accuracy, and transparency. The classification follows from where a system is used, not how sophisticated it is: software that screens job applicants is high-risk by definition.

The practical consequence is that compliance is a documentation and oversight discipline rather than a feature. Lantern is classified, documented, and logged as a high-risk AI system, with SOC 2 Type II, ISO 42001, and the NIST AI RMF underneath.

How we approach the EU AI Act →

High-volume hiring

High-volume hiring is a recruiting pattern in which one role or role family draws far more applicants than the team can interview — common in hospitality, retail, logistics, and construction. The binding constraint is interviewer hours, which means the process breaks at the screening stage rather than at the decision.

Adding reviewers scales cost linearly with applicants, so the usual outcome is a shallower screen instead of a bigger one. Gammon Construction reached shortlist 40% faster with Lantern, screening thousands of applicants without adding headcount.

Running interviews 24/7 →

Human-in-the-loop

Human-in-the-loop describes an automated system in which a person reviews, approves, or overrides the machine's output before it takes effect. In hiring it means the software may rank, score, and recommend, but a human makes every decision that changes a candidate's outcome — which is what turns an automated score into a defensible one.

The test of a human-in-the-loop design is what it does at the edges, not in the easy middle. Lantern never rejects anyone; uncertainty, complaints, accommodation requests, and anything the rubric does not cleanly cover escalate to the client's team with the full transcript and Lantern's reasoning attached.

Human oversight, in detail →

Interview rubric

An interview rubric is a written scoring guide that names the criteria a candidate is assessed on, the levels of performance within each criterion, and the observable evidence that separates one level from the next. It is what converts an interviewer's impression into a comparable score, and it has to be fixed before interviews begin to be worth anything at all.

A rubric an AI can score consistently is more specific than one written for humans: each level states the evidence that earns it, and the probes are written alongside the levels. Lantern applies the same interview criteria to every candidate, every time.

How to write a rubric an AI can score →

Language drift

Language drift is a systematic difference in how a scoring system rates the same quality of answer depending on the language it was given in. It is a specific fairness failure of multilingual assessment: the candidate loses points for the language they interviewed in rather than for what they actually said.

The remedy is one set of criteria across every language, monitored for the difference — not a separate standard per market. Lantern interviews in 40+ languages and dialects, spoken natively rather than translated, and monitors scoring for language drift so nobody scores lower for interviewing in the language they think in.

The languages Lantern speaks →

One-way video interview

A one-way video interview — also called an asynchronous or recorded video interview — asks a candidate to record answers to a fixed list of prompts, with no interviewer present and no follow-up questions. It solves the scheduling problem without solving the interviewing one: because the questions cannot respond to the answers, a thin answer and a substantive one both simply end.

It also moves the work rather than removing it — someone still has to watch the recordings. The dividing line between this format and an adaptive AI interview is whether the next question depends on the last answer.

How the two formats differ →

Shortlist

A shortlist is the ranked set of candidates a screening stage passes to the people who make hiring decisions. It is a recommendation, not a verdict — and its value lies less in the ranking than in the evidence attached to each name, which is what allows a hiring manager to disagree with the order and say why.

Lantern hands the hiring team a ranked shortlist by morning with the evidence behind every score. Recruiters read, decide, and hire; Lantern never rejects a candidate itself.

What a shortlist looks like →

Structured interview

A structured interview asks every candidate about the same predefined criteria and scores their answers against a common rubric, so differences in score reflect differences in answers rather than differences in interviewer. Structure means consistency of criteria, not rigidity of wording: a structured interview can still probe and follow up, as long as it probes toward the same criteria for everyone.

The failure mode it corrects is the unstructured chat, where each interviewer improvises criteria and the record is a set of impressions no one can compare. Lantern applies the same rubric to every candidate, every time — no good-mood interviews and no Friday-afternoon interviews.

Writing the criteria down →

Time-to-fill

Time-to-fill is the number of days between a requisition opening and an offer being accepted. It is the standard measure of recruiting speed, and because most of the elapsed time is spent waiting — for a slot, for a reviewer, for a reply — it is usually a queueing problem rather than a working-harder problem.

Removing a queue is worth more than compressing the work inside it. Hyatt reached 70% faster time-to-fill with Lantern, at a 7x ROI on high-volume hospitality roles.

Client outcomes →

Transcript-linked evidence

Transcript-linked evidence is a scoring practice in which every score points back to the exact moment in the interview transcript that produced it, so a number can always be checked against the words that earned it. It is the difference between an auditable score and a black-box verdict: without the link, a score is only an assertion.

In Lantern, every score links back to the exact moment in the transcript that earned it. Click a number and you see the evidence — a specific answer, at a specific timestamp.

Every score has a why →

“Lantern gave us a structured, consistent way to assess every candidate against the same criteria. That’s been a real step forward for fairness.”

CATALO

Vocabulary is the easy part. Seeing an adaptive interview score a real answer against a real rubric is the part that settles it.

Book a demo

The AI interviewingglossary