An AI Screen Buys Throughput and Costs You the Evidence

What Does an AI First-Round Screen Actually Buy?

Slots filled without a coordinator, the same questions asked in the same order every time, a transcript for every conversation, and a queue that never runs out of attention at four in the afternoon. Those are real gains. Three of the four have nothing to do with a machine asking the questions, and consistency in particular is a property of the script rather than of the interviewer.

That matters for the buying decision, because the pitch bundles them. Scheduling automation, recording, transcription and structured note capture are available without handing over the conversation, and they are where most of the recovered hours actually sit. A recruiter reading the same two questions from the same list delivers the consistency the demo attributes to the model.

What is genuinely new is asynchronous capacity: hundreds of candidates can complete a round overnight with nobody awake. That is a real capability and it is the one worth pricing honestly, because everything it costs is downstream of the same property. Nobody was there, so nothing was observed, and what remains is a recording and whatever the tool wrote about it. Whether a one-way video round is still worth keeping is the version of this question most teams face first.

Why Does the Evidence Get Worse?

Because the record thins at exactly the point it will be needed. Automatic transcripts are not verbatim: an audit of one widely used speech-to-text model found roughly 1% of transcription segments contained entire phrases or sentences that appear nowhere in the audio, and 38% of those carried explicit harms such as invented violence or false authority ¹. A fabricated sentence reads exactly like a real one in the file a reviewer opens.

The same audit found the hallucinations fell disproportionately on speakers with longer non-vocal stretches, a common symptom of aphasia. That is one model version audited against a research corpus rather than against interviews, and four other commercial services showed no comparable behavior on the same segments, so it is a finding about a tool rather than about transcription in general. Which model does the transcribing is a fair question to put to any vendor before the first candidate uses it.

The format changes the candidate too. A 2026 preprint on asynchronous AI interviewers, combining forum analysis, seventeen applicant interviews and a 180-participant test of a research prototype, found that expectations shaped by familiarity with these models went unmet against the employer's framing, which damaged the applicants' sense of agency and trust and pushed them toward workarounds ². That describes what an unaccountable one-way format provokes, not dishonest candidates. It has not been peer-reviewed and the interview sample is small, so read it as a direction. What an AI notetaker's summary may put in a hiring file is the same evidence problem one stage later.

Keep the Asking and the Deciding Human

Keep a person on the two acts that carry judgment: asking the questions, and deciding who moves on. Public objection lands on who decides, not on whether software sits in the process anywhere. In Pew's survey of 11,004 US adults, fielded in late 2022, 71% opposed AI making a final hiring decision against 7% in favor, while opinion on AI merely reviewing applications split 41% opposed, 28% in favor and 30% unsure ³.

Disclosure carries a cost in applicant willingness, and one number gets quoted for it more than any other. In the same survey, 66% said they would not want to apply for a job with an employer that uses AI to help make hiring decisions, against 32% who would ⁴. That is stated intention from a general sample rather than a measured drop-off in any real funnel, and nobody in it was standing in front of a job they wanted. Read it as a design constraint on how the use is described, not as a forecast.

The practical split is clean. Machine: invitations, reminders, slot booking, recording, transcription, filling the same fields for every candidate, and flagging where a required question went unanswered. Human: asking, following up, and deciding. The follow-up is the load-bearing half, because a scripted question with no follow-up is a survey, and the difference between a prepared answer and a real one usually appears two exchanges in.

Set These Four Rules Before You Switch It On

Disclose it in the posting, obtain consent where the law requires it, keep the transcript and the questions with the requisition, and require a named person to make every advance and every rejection. Four rules, none of which slows the tool down, and together they are the difference between a defensible round and one that has to be reconstructed from memory a year later.

Consent is not optional everywhere. Illinois has required since 2020 that an employer using AI to analyse recorded video interviews for a position in the state notify the applicant beforehand, explain what general types of characteristics the system evaluates, and obtain consent, and it bars evaluating an applicant who has not consented ⁵. The statute names no penalty and no private right of action, so the practical duty is the notice and the alternative you offer to anyone who declines. Two other jurisdictions reach an automated round on their own terms: New York City has required since January 2023 a bias audit from within the prior year and notice 10 business days before an automated employment decision tool is used on its candidates ⁷, and California's discrimination regulations, effective October 1, 2025, name analysis of facial expression, word choice and voice in online interviews as an automated-decision system ⁸⁹. Which of these binds a given requisition is a question for counsel, not for a vendor's compliance page.

The fourth rule is the one that pays for itself. In a vignette experiment with 921 working-age adults in Austria, a rejection from an AI with no explanation scored lowest on all four measures taken, while an AI rejection that included an explanation drew the same ratings as a human rejection without one ⁶. Every condition in it rated below the midpoint of the scale, so what it ranks is unhappy outcomes and the gaps are fractions of a point. The explanation still does most of the work, and an explanation only exists if a person can describe what happened.

So write the rejection language before the first candidate arrives, and check that somebody could actually say it truthfully. If the honest sentence is "a rating you cannot see placed you below a threshold nobody set deliberately", the round is not ready. What is owed to a candidate who asks why the AI screened them out is worth settling before the question arrives, and the two AI questions a human screen can answer in fifteen minutes is what the automated round is competing against.

See a sample report >