calhire
Capability

AI text interview

Two decisions define this feature: it is text rather than video, and the model never receives anything that identifies the person answering.

CalHire’s AI interview is a structured, text-based conversation that asks every candidate for a role the same core questions and scores the answers against a fixed rubric. It is deliberately not a video interview: no facial analysis, no tone or speech scoring, no recorded likeness. No personally identifying information is sent to the model at any point, and the resulting score is one weighted input into a ranking a person reviews.

Last reviewed

The short version

  • Text only. Video interviews on the platform are conducted by humans, never scored by a model.
  • No personal data reaches the model. That is a hard architectural boundary, not a privacy setting.
  • Every candidate for a role gets the same core questions, which is what makes the answers comparable.
  • The interview score is one weighted component of a composite that a person acts on.

Why text, when everyone else records video

Video interview scoring had a bad decade and deserved it. Systems that rated candidates on facial expression and vocal tone were sold widely, then withdrawn or restricted after regulators and researchers looked closely. The core problem was never the accuracy of the model. It was that appearance and speech carry disability, ethnicity, age and neurotype, and a model trained on past hiring outcomes learns to reward whatever the past rewarded.

Text sidesteps the whole category. What is being read is what the candidate wrote about the work. A candidate with a stammer, a strong accent, a facial difference, or a webcam pointed at a shared kitchen is on identical footing with everyone else, and no adjustment or accommodation request is needed to get there.

It is also a better instrument for most of what an early-stage interview is trying to find out. Written reasoning is legible, quotable and re-readable by a second reviewer. A hiring manager can look at exactly what was said rather than at a recording of how it looked to say it.

How an interview runs

  1. 1

    You define what the role is testing for

    Competencies and the questions that probe them. The platform can draft these from the job, and you edit them. The set is fixed before the first candidate starts, which is the entire point of a structured interview.

  2. 2

    Every candidate answers the same core questions

    In text, in their own time within the window you set. Follow-ups may vary where a candidate needs clarification, but the core set does not move between candidates.

  3. 3

    Answers are scored against the rubric

    Each competency is scored on the criteria you defined, with the reasoning attached to the score. A score with no stated reason is not something a reviewer can check or a candidate can appeal.

  4. 4

    The integrity layer reads the answers, not the room

    Content-authenticity signals look for machine-generated text, cross-candidate similarity, style inconsistency and reasoning that falls apart under a clarifying question. These raise flags with a confidence level attached.

  5. 5

    The result feeds the composite

    Weighted at 35 percent by default, alongside the verified test score and role fit. A person reads the ranked list and decides what happens next.

The boundary the model never crosses

No personally identifying information is sent to any language model. Not the candidate’s name, not their email, not their employment history, not anything that could be used to work out who they are. The model sees the question, the answer and the rubric.

This is worth stating precisely because it is the kind of claim vendors make loosely. It is enforced at the architecture level: the interview service composes prompts from a bounded, non-identifying set of fields, and the boundary is one of the platform’s standing invariants rather than an option a tenant can switch off.

It matters for a reason beyond privacy. A model that cannot see who is answering cannot condition its scoring on who is answering. Removing identity from the model’s input is the same protection that anonymous screening gives a human reviewer, applied to the component that scales fastest.

What structure actually buys you

  • Comparability

    Two candidates asked different questions cannot be compared, however carefully you compare them. Same questions, same rubric, same scale is the minimum condition for a meaningful ranking.

  • A defensible record

    Question, answer, rubric, score and reasoning are all retained. If a decision is ever challenged, the record is the thing that answers it.

  • No scheduling

    Candidates answer within a window rather than in a booked slot. Anyone in another timezone, on shift work, or in a job they have not left yet is not quietly excluded by your calendar.

  • Predictive validity

    Structured interviews outperform unstructured ones as predictors of job performance. This is one of the most consistent findings in a century of personnel research, and it is free to act on.

What this does not do

It does not replace a human interview. It is an early structured stage that narrows a field fairly. Judgement about a person, a team and a specific job is human work, and the platform expects you to do it before an offer.

It cannot assess what writing cannot carry. Bedside manner, physical craft, live negotiation under pressure and stagecraft are all real requirements that a text interview measures badly or not at all. For those, use a work sample or a human conversation, both of which the platform supports.

Text advantages fluent writers. This is a real limitation and worth naming rather than glossing. Rubrics score substance rather than polish, and the model is instructed accordingly, but a role where written communication is genuinely irrelevant should weight this component down or skip it.

Integrity signals on interview answers are probabilistic. Machine-generated text detection produces both false positives and false negatives, and the platform treats every such signal as evidence for a human to weigh rather than a finding to act on. Nobody is declined on a detector’s say-so.

Questions people actually ask

Does the AI decide who advances?
No. It scores answers against a rubric and contributes a weighted component to a ranking. Advancing and declining are actions taken by a person, recorded against their name in the audit trail.
Is there a video interview at all?
Yes, later in the process, and it is conducted by your team with scheduling support. It is a human conversation. No model scores video, analyses a face, or rates a voice at any stage.
What personal data does the model receive?
None. Prompts are composed from the question, the candidate’s answer and the rubric. Identity is held separately by the identity broker and is not available to the interview service.
Can candidates use AI to write their answers?
They can try, and the content-authenticity signals are built for exactly that: machine-likeness, similarity across candidates, style inconsistency, and answers that collapse when asked a clarifying question. Flags are reviewed by a person, because a wrong accusation here is costly and unfair.
How long does a candidate get?
You set the window. Longer windows help people fitting an interview around a current job; shorter windows reduce the time available to seek outside help. Most teams settle somewhere between two and seven days.
Which model is used?
Model choice is configuration rather than code, and routes are provider-agnostic by design so a model can be changed or replaced without a release. What does not change is the boundary: whichever model runs, it does not receive personal data.

Where this connects to the rest of the platform.

  • Composite scoring and blind ranking

    A composite score you set the weights for, defaulting to 40 percent test, 35 percent interview and 25 percent role fit. It orders a list. It never makes a decision.

  • Assessment integrity without surveillance

    Four families of signal, each with a confidence level and PII-free evidence. No webcam, no verdicts, and environmental flags are excluded from scoring where accommodations apply.

  • AI governance, bias auditing and the audit trail

    Adverse-impact analysis on the four-fifths rule, a hash-chained audit trail, candidate appeals, and DSAR handling. The evidence exists before anyone asks for it.

Read the reasoning

The evidence and the argument behind what is on this page.

Hiring Playbook8 min read

Structured interviews: the highest-return change most teams never make

Same questions, same order, same rubric, scored independently. Structured interviews are unglamorous, uncomfortable to adopt, and the best value in hiring.

Read
Candidate Guide8 min read

How to prepare for an AI interview

AI interviews reward specificity, structure and reasoning you make visible. What they actually assess, how to prepare, and what your rights are.

Read
Integrity8 min read

Candidates are using AI in your assessments. Now what?

AI-text detectors are unreliable and disproportionately flag non-native speakers. What to do instead: assessment design, behavioural signals, and human review.

Read

Browse all topics on the blog

See a verified pipeline for one of your roles

Post a role free and review anonymous, skill-ranked candidates. No card, no sales call to get started.