HireVue in 2026: What It Records and How It Scores

HireVue's scoring models read what you said, not how you looked saying it. The company's 2025 AI Explainability Statement is explicit: "our AI relies only on what is said by the candidate and does not use any video analysis or other audio characteristics (meaning that we do not assess a candidate's facial expressions, body language, their background and surroundings, or tone-of-voice)" (HireVue 2025 AI Explainability Statement).

That single sentence resolves most of what candidates worry about before a HireVue interview, and it leaves the more interesting questions open. Your session is still recorded. Something is still watching the window you have in focus. And as of June 2026 the format itself has changed: HireVue now sells a two-way voice AI Interviewer alongside the recorded one-way interview most people picture. This guide covers what each HireVue surface captures, how the scoring works, what gets flagged, and how long any of it survives, based on HireVue's published documentation as of 1 September 2026.

The four HireVue surfaces you can meet

"A HireVue" is not one thing. Which product you were sent determines what is recorded and who scores it.

The AI Interviewer is the new one. HireVue announced it on 18 June 2026 as conducting "dynamic, two-way voice interviews with every applicant 24/7," delivering recruiters "an evaluated, ranked-for-review, and evidence-backed prioritization of candidates" (HireVue press release, 18 June 2026). It follows HireVue's 10 March 2026 acquisition of Hireguide's technology, which the company framed as moving hiring workflows "from simple automation to a smarter way to qualify candidate capability" (HireVue, 10 March 2026).

The practical consequence: an invitation that says "HireVue" no longer reliably means a silent recording of fixed prompts. It may mean a system that interrupts, probes, and asks you to expand on the answer you just gave.

How the transcript becomes a score

HireVue's explainability statement describes a three-stage pipeline. Speech is transcribed by a third-party speech-to-text system. A language model the company calls CUSTARD, built on RoBERTa, converts that text into numerical vectors. Those vectors are compared against Behaviorally-Anchored Rating Scales that HireVue reports were trained on 99,411 expert human ratings. The company also states it tests models for adverse impact using the 4/5ths rule and statistical measures including Cohen's d and Fisher's Exact.

These are vendor descriptions of vendor systems, and no candidate can verify them from the outside. But the architecture is consistent with a change HireVue made publicly. The company removed visual analysis from its algorithms in 2020, a decision reported alongside an independent audit by O'Neil Risk Consulting and Algorithmic Auditing published in January 2021. HireVue's chief data scientist said the content of a candidate's answers was the strongest predictor and that nonverbal data contributed roughly 0.25% of predictive power in most cases; the CEO's framing was that "it wasn't worth the incremental value" given bias concerns (Fortune, 19 January 2021).

Two things follow for a candidate. First, a poor microphone is a real scoring risk in a way that a plain background is not, because a transcription error is an error in the only input the model reads. Second, none of this removes the human viewer. HireVue's candidate FAQ describes its interviews and assessments as "first steps in the overall interviewing process," with top-group candidates moving "on to person-to-person interviews before hiring decisions are made by people" (HireVue candidate FAQ). Eye contact and framing still matter to whoever watches the recording later. They just are not what the model scores.

What is logged besides your answers

This is where the "it only records camera and audio" summary breaks down.

HireVue documents that its platform "periodically takes snapshots of candidates during interviews to confirm identity and flag suspicious activities" and "monitors if candidates switch tabs or windows during an interview" (HireVue on cheating mitigation). Those snapshots are an integrity signal rather than a biometric match: HireVue's FAQ states the company "never uses the video recording of your interview in order to identify you." If you want the wider picture of how ID checks in remote hiring differ from proctoring, our candidate guide to interview identity verification covers what platforms capture and keep.

CodeVue, the coding assessment, is instrumented more heavily than the video interview. HireVue documents using Measure of Software Similarity scoring to flag likely plagiarism, plus an activity timeline that records "time spent in another browser window," when content is pasted, and when code is tested successfully (HireVue on plagiarism in technical assessments). The company is unusually candid about the limits: "High code similarity is an indication of suspicious activity... it should not be used in isolation," and the timeline approach "is not entirely foolproof," since a developer legitimately working elsewhere produces the same trace. HireVue's suggested resolution is a follow-up on-demand video question asking the candidate to explain their own solution.

Two scope notes matter here. These are capabilities HireVue sells to employers, and an individual employer configures which are switched on; a flag is routed to a recruiter, not converted into an automatic rejection. And HireVue's public documentation describes focus and window signals, not full screen capture of your desktop — which is a statement about what is documented, not proof of what is technically absent. The general mechanics of what a browser-based assessment can and cannot see are covered in our explainer on whether interview websites detect tab switching.

How often the flags actually fire

HireVue published its own numbers in a September 2025 white paper, Do Cheaters Prosper?, drawing on internal studies from 2023 and 2024 (HireVue white paper, September 2025):

• In a sample of 1,000 transcribed candidate responses, under 1% were flagged for high similarity — and the flagged candidates "performed poorly on the assessment anyway and were not selected."
• A 2023 impersonation study found 1.9% of completions were both in the fastest 10% of completion times and achieved a top score.
• Tested against HireVue's Virtual Job Tryouts, ChatGPT-3.5 scored 1 to 2 on a 5-point scale and took over an hour on a 20-minute assessment. On AI-scored interviews it landed between 2 and 5, mostly average.
• Aggregate scores showed no drift after large language models became widely available, which HireVue reads as evidence of no meaningful advantage at scale.

Treat these as vendor-reported findings from a party with an interest in the answer, and note the model tested: ChatGPT-3.5 was already two generations behind by the time the paper was published in 2025. The stable-score finding is the more durable claim, because score drift is measured across an entire population rather than inferred from a detector. The honest summary is that HireVue's detection layer catches a small fraction of a small population, and the company's own data suggests that generic generated answers score near the middle of the distribution rather than at the top — which is a weaker argument for using one than most people assume.

Who sees it, and how long it lives

Retention is the employer's decision, not HireVue's. The candidate FAQ states: "Hirevue stores your personal data based on the retention period (for the amount of time) chosen by the company that you're interviewing with. We automatically purge personal data when that company's retention period expires." Candidates can "request that your personal data be deleted at any point in the process" through the recruiter or the company's application page.

What that looks like in practice is visible in employers' own notices. EY's published privacy notice for its HireVue digital interview and CodeVue tools lists the data collected — name, email, "your voice; your image; your video images; test results showing your talents and cognitive abilities" — names the internal and vendor parties with access, and sets retention at "three (3) years of absent data user activity" (EY privacy notice for HireVue tools). Three years is a long tail for a forty-minute recording, and it is a term you would only find by reading the notice attached to that specific application.

Illinois gives candidates a firmer floor. The Artificial Intelligence Video Interview Act requires an employer using AI to analyze video interviews to notify the applicant beforehand, explain how the AI works and "what general types of characteristics it uses to evaluate applicants," and obtain consent before evaluating them. Videos may only be shared with "persons whose expertise or technology is necessary in order to evaluate an applicant's fitness for a position," and on request the employer must delete the interview and instruct anyone holding copies to delete theirs, including backups, within 30 days (820 ILCS 42). If you are trying to work out which disclosure rules bind an employer interviewing you elsewhere, our guide to AI hiring rules and candidate disclosure in 2026 maps what survived the EU's timeline changes and what the US state patchwork already requires.

What HireVue does not publish

Worth naming, because the gaps shape how much weight the rest deserves:

• The scoring rubric for your specific role. Explainability is described at the level of method, not of the competency model applied to your recording.
• Which integrity signals a given employer enabled. The features are documented; the configuration is not disclosed to candidates.
• What the AI Interviewer captures and retains relative to the older on-demand product. The June 2026 launch materials describe the interview experience and the rubric, not the data lifecycle.
• Any per-question threshold or score you would need to advance, which is set by the employer.

Preparing for the one you were actually sent

Read the invitation for the format, not the brand. A recorded one-way interview rewards a rehearsed, structured, spoken answer delivered once. A two-way AI Interviewer rewards the ability to be interrupted and to extend an answer without losing the thread. A CodeVue challenge is scored on working code and a submission timeline that you should assume is reviewable.

Then take the three steps that reliably change outcomes: test your microphone on the actual device, since the transcript is the model's only input; ask the recruiter what the AI does and whether accommodations are available, which HireVue's FAQ says most employers now support, including advance access to questions; and rehearse out loud against a timer rather than reading a script, because a read answer is audible in the transcript.

On the question people arrive with: a HireVue on-demand interview or assessment is a closed evaluation unless the employer explicitly says otherwise, and it is instrumented with focus monitoring and identity snapshots. Running a live answer generator into it is a poor trade against HireVue's own reported score distributions. Our broader guide on how to prepare for AI and one-way video interviews sets out the format check and evidence plan in full.

Where a desktop assistant belongs in this process is before the recording, not during it. Control is a Windows and macOS desktop app that transcribes speech in real time and answers from screenshots, which makes it usable as a rehearsal partner for spoken answers and a study tool for the coding round. New accounts include 5 messages and 2 minutes of voice at no cost, which is enough to see whether a spoken-answer drill helps you before you pay for anything. Use it on the practice run, and let the recording capture your own words.