HireVue's video interview and assessment platform

What this models

This example runs on the Vendor-layer-class video interview assessment network: 10 components and 25 pathways between them. Every context in the Lab is a stylized model, never a reconstruction of any actual deployment, and each assumption behind it carries a provenance label.

Evidence base: 1 assumed · 9 published baseline. In the Lab, the shaded evidence band behind each headline readout draws its width from the least-established class below.

• baseline

  D48-derived new org (Phase 6, hiring-employment-screening). REGISTER FIRST, because it governs every value here: this is the VENDOR layer. Every governance event in this record happened at the party that builds the engine, while every deployment decision sat with the employers that bought it. Nothing on this diagram asserts anything about any individual client's pipeline, any client's reported benefits, or any client's hiring outcomes, and the shipped client-side hiring boards are a different subject entirely. The one historical event shared with the shipped video-assessment deployer board — the removal of the visual and facial channel — is told here from the side of the party that removed it.

• baseline

  TOPOLOGY. Ten nodes, all documented, none decorative. ONE model, carrying both of the platform's scored assessment families: the competency-based interview assessments and the game-based assessments are built from the same pooled store, placed on the same three-tier output, delivered to the same recruiter surface, written back to the same store and covered by the same annual bias audits, and the record's difference between them is that the interview family is the primary scored channel and the one the validity scholarship reaches. That is a difference of degree on one channel, so it is stated in the model's own description rather than drawn as a second sub-graph. TWO operator classes: the employers that bought the platform, and the vendor organization that builds the models and chooses the input set. The buyers are one class because the sources document one population holding one authority and one set of statutory duties, split by where it set a single dial — approximately 20 percent of customers using algorithmic scoring as of January 2021 and the rest using the platform for human review of recorded video — which is an adoption share, carried on the class and on its configuration pathway. THREE reviewers, because the record documents three separately constituted external channels that produced three different kinds of result: a commissioned audit that produced a narrow finding publicized broadly, an annual statutory bias audit that produced the file's only public numbers, and private litigation that produced the only price anyone has attached to anything here. THREE stores, because the pooled multi-employer training-and-audit record, the retained interview media and the published audit filings are three different things the record measures separately, each is the object of a different documented duty or instrument, and the wiring between them is the case. ONE input source, because the visual and facial feature channel is a real, named, measured and then removed input rather than a modelling convenience.

• baseline

  LEG RE-DERIVATION (2026-09-22, car LEG-3). This board was re-drawn at the coarsest granularity that still distinguishes every documented mechanism of the deployment, and it lands at ten nodes and twenty-five pathways, from twelve and thirty-six. Nothing documented was dropped and no width was changed: every surviving pathway keeps the rung it already carried. Two nodes were folded into survivors. The second scored engine folded into the one model, because the two families differ in degree rather than in wiring, and every fact its node carried — the auditor engagement naming both, the published filing reporting both, and that the validity scholarship reaches only the interview construct — is stated on the model. The second client class folded into one, because both populations hold the same authority and the same duties and are distinguished by an adoption share, and the share is stated on the class and on the configuration pathway, where the dial's two documented settings are named together. Eleven pathways folded into survivors carrying the same documented flow: the second engine's four pathways into their interview counterparts; the engines' return legs to their builders into the read of the pooled store the scores are written into; the employer's read of the filing it posts into the posting itself; the commissioned report's contribution to the public store into that same pathway, where the nondisclosure gate that is its whole documented content is stated; the employer's write into the pooled store into the accumulation pathway that carries the same applicant material; the human-review buyers' routing into the configuration pathway that now names both settings; and the settlement's reach back onto the recordings into the inbound litigation pathway, where the posture that is its whole documented content — preliminary approval, no admission, nothing paid — is stated. Three baseline-zero pathways remain and all three are source-documented absences: the removed visual input, the reconciliation of a filing against its source, and the independent read of the engine.

• baseline

  ABSENCES ARE DERIVED TOO, and three of them are load-bearing. There is NO enforcement node: the vendor emits a tier and the buyer decides what happens next, so no downstream action system is documented at this layer and drawing one would assert a mechanism the record refutes. There is NO external boundary and no egress pathway: the federal court let a third-party-disclosure claim proceed at the pleading stage, but that is an allegation held plausible and never a finding, and drawing an egress edge would put a disclosure on the diagram that no court, regulator or investigation has established. There is NO channel by which a screened-out person contests a score, and none is drawn, because none is documented and because candidates are boundary-only on every Lab diagram. There is also no worklist, no retriever and no guardrail: no queue, backlog, retrieval component or automated output screen appears anywhere in this record.

• baseline

  WHERE THE LAB SHAPE DIVERGES FROM THE PAN SHAPE, and nothing is asserted here that the PAN file does not already record. Four divergences, and one former divergence that the re-derivation closed. First, PAN draws TWO models for the two scored assessment families; the Lab now draws ONE, because the two families share every pathway on this board and the record's difference between them is a difference of degree, which the model's own description carries. Second, PAN carries the removed visual channel as an attribute of its interview model; the Lab draws it as an input source with its own pathway, so the measurement and the removal are two distinguishable acts rather than one adjective. Third, PAN has no edge kind for a check at all. Two of the three check pathways here are PAN peer edges REDRAWN as checks — the auditor's recommendations reaching the model builders, and the consent claim reaching them — because a channel that improves the product is inhibiting in the Lab's vocabulary and reinforcing in PAN's, and their widths still come from PAN on the stated mapping; the third check, the annual statutory audit's read of the builders, has no PAN counterpart and is derived from the cited record. Fourth, PAN carries no operator-into-model edge, so the buyer's configuration dial is Lab-side. CLOSED: the Lab previously drew TWO client operator classes where PAN carries ONE with the adoption share in that class's attributes; the re-derivation draws one, on the same reading PAN already used.

• baseline

  BASELINES, and exactly how far the PAN org carries them. The PAN entry for this deployment holds twenty-seven edges. Eighteen of this network's twenty-five pathways have a one-to-one counterpart among them, and every one of those eighteen mirrors that edge's width on a single rung mapping (0.50 and above to 3, 0.30 to 0.49 to 2, 0.06 to 0.29 to 1, documented absent to 0) with no exceptions — including the two peer edges redrawn here as checks, whose widths still come from PAN. Where a Lab pathway now carries more than one PAN edge, it keeps the rung of the widest of them and its derivation comment names every edge it absorbs; no width was moved in either direction by the re-derivation. The remaining seven are derived from the cited record directly, and each says so on its own line: the retired visual channel's scoring pathway and the measurement that preceded its removal, the buyer's configuration dial, the annual statutory audit's check on the builders, the reconciliation of a filing against its source, the one-engine self-loop, and the independent read of the engine. Three contrasts are load-bearing. The pooled record reaches the engine at the top rung while the recording of any individual candidate reaches it one rung lower, which is the file's structural claim that the engine is made of the pool rather than of the person in front of it. The visual channel's measurement pathway runs at the low rung while its scoring pathway runs at zero, and that pair is the retirement drawn as two acts. And the aggregation from the pooled store into the published filings runs at the middle rung while the reconciliation back the other way runs at zero, which is the disclosure regime's inversion stated as a picture rather than as a complaint.

• baseline

  DEMAND 3 / CAPACITY 3. Demand 3 on the vendor's own time-stamped volume: more than 19 million video interviews and more than 700 customers as of January 2021, more than 33 million interviews and more than 800 customers as of January 2023, with a single published audit filing counting 20,060 male and 9,121 female applicants on one assessment for one job family pooled across employers over a two-year window, and an Illinois settlement class estimated at 91,305 people. Capacity 3 is the unusual half and rests on two things. The human counterfactual here is not hypothetical: the vendor reported that approximately 20 percent of its customers used the scoring feature as of January 2021 and that the rest used the platform for human review of recorded video, so the manual comparator is the majority configuration of the same platform doing the same job on the same recordings. And the independent validity scholarship puts human interviewer judgement in the criterion position: automated video-interview personality assessments trained on self-reports showed little evidence of reliability or validity, while models trained on interviewer reports did better with mixed cross-sample reliability. What the record does not contain is a head-to-head comparison of this deployment against its own human counterfactual, and a technology-press review of both audit regimes recorded that neither addressed whether the products improve hiring at all. The value rests on the documented majority configuration and the criterion position of human judgement, and on nothing measured about this deployment.

• baseline

  EVIDENCE STATUS, labelled where it is used, because this record mixes tiers more than most. VENDOR-REPORTED and not independently audited: every scale figure (19 and 33 million interviews, 700 and 800 customers, 200 million chat engagements), the approximately 20 percent scoring-adoption share, and the 0.25 percent and 4 percent feature-contribution measurements that justified the removal. VENDOR CHARACTERIZATION: the phrase that the assessments work as advertised with regard to fairness and bias, which describes a one-use-case perimeter. PRIMARY COURT RECORD: the federal memorandum order of 26 February 2024 and the state settlement notice of June 2026. PRIMARY AUDIT FILING: the July 2023 bias-audit summary with every impact-ratio table and applicant count. PEER-REVIEWED: the study of all 116 public filings, the validity investigation of the construct class, and the analysis of vendor claims and checking practices. ADVOCACY FILING: the November 2019 complaint to the federal consumer regulator, which produced no public enforcement action of any kind. No parameter on this diagram is scaled by a vendor figure without that figure being named as the vendor's on the line that uses it.

• baseline

  THE STORE'S DOCUMENTED BLIND SPOT, stated as the structural fact it is rather than as a measurement. A rejected candidate generates no outcome data anywhere in this loop, so the pooled record that builds the models and computes the mandated audits cannot see the population the models screened out. That is why the audits measure selection-rate ratios over the applicants they can classify and never accuracy, and it is why no error rate for this engine has ever been published by anyone. Nothing on this diagram computes a harm to a screened-out person, and nothing could: the absence is a property of the loop, and it is drawn as the reconciliation pathway that runs at zero rather than as an outcome anybody has counted.

• assumed

  Screened people are not in the dynamics. No hiring decision, tier placement, rejection, or employment outcome for any person is computed from anything drawn here, and no score over any person is authored anywhere in this network. The impact ratios in the published filing are recorded external observations from a mandated disclosure — ratios of selection rates over the applicants an audit could classify — and they are neither error rates nor findings of discrimination; the peer-reviewed study of the whole filing regime found this auditor's comparator and aggregation method sometimes yields ratios above one. The estimated 91,305 Illinois class members and the 3.75 million dollar fund are recorded litigation facts about a settlement that is preliminarily approved and not final.

What this example does not show

• LITIGATION AND REGULATORY POSTURE, verbatim from the evidence dossier and load-bearing. Operating vendor at scale. Facial-analysis input retired (2020, announced 2021). BIPA class settlement preliminarily approved June 25, 2026 for $3.75M, final approval hearing October 28, 2026 — not yet final, no payment made, no admission of liability (HireVue denies collecting biometrics subject to BIPA). No public FTC enforcement action on the EPIC complaint is on the record. 2025 Intuit/HireVue civil-rights charges pending; HireVue disputes them. Annual DCI-conducted LL144 bias audits published via client employers.
• Two courts, two case numbers, and this scenario states both without asserting a connection between them. The federal action in the Northern District of Illinois (No. 1:22-cv-01284, Judge Jeremy C. Daniel) produced the 26 February 2024 ruling on the motion to dismiss: claims under three subsections proceeded, the profit claim was dismissed, and the court held the state's video-interview statute does not preclude the biometric statute. The settlement is being approved in the Circuit Court of Lake County, Illinois (No. 2026LA00000141, Hon. Daniel L. Jasica). The record consulted does not document the procedural mechanism of the venue shift, so none is asserted. A claim that survives a motion to dismiss is an allegation held plausible; no court has ever found this vendor violated the biometric statute or anything else.
• The complaint filed with the Federal Trade Commission in November 2019 was an advocacy filing, and the Commission never announced an enforcement action or any finding. The honest statement is that the vendor was the subject of a complaint, and this scenario says nothing stronger. The sequence of complaint, then quiet removal of the visual channel in about March 2020, then public announcement with the audit in January 2021 is stated in that order because that is the order it happened in; causation by the complaint is an inference the record does not make, and the vendor's own stated reasons were the low measured contribution and rising public concern.
• The phrase that the assessments work as advertised with regard to fairness and bias is the vendor's characterization of the commissioned audit, and it travels with the audit's perimeter wherever it appears: one representative pre-built early-career assessment use case, the tool's technical design and training data outside the examination, structured stakeholder interviews as the method, and a report published on the vendor's own site behind a nondisclosure agreement. Recorded criticisms are part of the same record — its depth was publicly contrasted with a contemporaneous source-code audit of a competitor, an auditor paid by the audited party carries a conflict risk, and neither audit addressed whether the products improve hiring.
• Every scale figure and both feature-contribution figures are vendor-reported and none is independently audited: more than 19 million interviews and more than 700 customers as of January 2021, more than 33 million interviews, 200 million chat-based engagements and more than 800 customers as of January 2023, the approximately 20 percent scoring-adoption share, and the roughly 0.25 percent and 4 percent nonverbal contribution measurements. This scenario names them as the vendor's wherever it uses them, and no value on the diagram is treated as an audited quantity.
• The impact ratios in the published bias audits are computed on POOLED multi-employer data with a comparator and aggregation method that sometimes yields ratios above one, and the peer-reviewed study of all 116 public filings could not determine the cause of every set of identical results it identified — data pooling, shared underlying audits and coincidence are all left open by its authors. Those findings are observations about a measurement regime. They are not findings of discrimination against anyone, and they are not findings of audit fraud.
• The Illinois video-interview statute places its duties on employers, not on the vendor, and its text states no enforcement mechanism and no private right of action; the New York City ordinance likewise binds employers. Nothing here describes either instrument as regulating the vendor directly. What the vendor is documented doing is supplying the compliance machinery its buyers need and arguing publicly that vendors should share the audit responsibility — and, in litigation, arguing that the video-interview statute displaced the biometric one, which the federal court rejected.
• The March 2025 civil-rights charges filed with a state civil-rights division and the federal employment agency over an automated video interview are pending allegations, and both companies dispute them; the vendor's chief executive called the complaint entirely without merit and stated that the employer did not use one of its assessments at all. Nothing on this diagram rests on them, and no stressor, baseline or lever is derived from them.
• This board is the VENDOR layer and asserts nothing about any buyer's pipeline. The shipped client-side hiring boards in this catalogue are different deployments with different evidence, and no benefit figure, time-to-hire claim or diversity figure reported by any deploying employer appears anywhere here. The single historical event this record shares with the shipped video-assessment deployer board is the removal of the visual and facial channel, told here from the side of the party that removed it.
• Screened people are not modeled. No hiring decision, tier placement, rejection or employment outcome for any person is computed from anything on this diagram, and the applicant counts, impact ratios and class-size figures are recorded external observations from a mandated filing and a court record rather than quantities this network derives. The structural fact the record does establish — that a rejected candidate generates no outcome data anywhere in the loop — is drawn as a pathway that runs at zero, never as a harm anyone has counted.

Sources and evidence

What this example rests on, claim by claim. Every entry resolves to the same ledger the Evidence Registry publishes.

• HireVue's video-assessment platform is a vendor-layer deployment: one scoring engine behind hundreds of separate employers' hiring pipelines. Candidates record answers to a structured question set, or play game-based assessments; per-assessment models score verbal and paraverbal features of the responses into competency scores that place each applicant in a Bottom, Middle, or Top tier, and the client employer chooses where to cut — one deploying employer's published bias audit records it evaluating at Top-plus-Middle against Bottom. The client also chooses whether to use algorithmic scoring at all: as of January 2021 the vendor reported that approximately 20 percent of its customers used the predictive-analytics feature and that the rest used the platform for human review of recorded video. Every scale figure is vendor-reported and none is independently audited: more than 19 million video interviews and more than 700 customers as of January 2021, and more than 33 million interviews, 200 million chat-based candidate engagements, and more than 800 customers as of January 2023. The models are built on historical applicant data pooled across employer implementations, and the mandated bias audits are computed on that same pooled store. A rejected candidate generates no outcome data anywhere in that loop, so the store that trains and audits the models cannot observe the population the models screened out.

  empirical
  • Trade press Fortune (2021, January 19). HireVue stops using facial expressions to assess job candidates amid audit of its A.I. algorithms https://fortune.com/2021/01/19/hirevue-drops-facial-monitoring-amid-a-i-algorithm-audit/
  • Trade press Maurer, R. (2021). HireVue Discontinues Facial Analysis Screening. SHRM; with HireVue and ORCAA audit announcements (2021). https://www.shrm.org/topics-tools/news/talent-acquisition/hirevue-discontinues-facial-analysis-screening
  • Government DCI Consulting Group (2023). New York City Local Law 144 bias audit for HireVue, summary produced 5 July 2023, published by Pfizer as a deploying employer https://cdn.pfizer.com/pfizercom/CareersEmploymentDocs/HireVue_2023_Bias_Report_14AUG2023.pdf
  • Vendor HireVue (2021, January 12). Industry Leadership: New Audit Results and Decision on Visual Analysis; and HireVue (2023, January). HireVue engages external auditor DCI Consulting Group for external bias audit of algorithms (press release) https://www.hirevue.com/blog/hiring/industry-leadership-new-audit-results-and-decision-on-visual-analysis

• The vendor removed visual and facial analysis from new assessment models in early 2020 — the Society for Human Resource Management reports the discontinuation as March 2020 — and announced the decision publicly on 12 January 2021, together with the results of an algorithmic audit it had commissioned from O'Neil Risk Consulting and Algorithmic Auditing. Its stated reason was that advances in language analysis had left visual features contributing little: internal research put the nonverbal visual contribution at about 0.25 percent of the model's predictive power in most job models and about 4 percent for high-customer-contact roles, figures given by the vendor's chief data scientist and never independently audited. Its chief executive said it was not worth the concern it was causing people. The sequence is complaint in November 2019, removal in about March 2020, public announcement with the audit in January 2021; causation by the complaint is an inference the record does not make, and the vendor's own stated reasons were the low measured contribution and rising public concern. The removal is a rare documented instance of an input dropped as its measured value approached zero while the cost of scrutiny rose, and it was applied platform-wide by a party no single client employer could have compelled.

  empirical
  • Trade press Maurer, R. (2021). HireVue Discontinues Facial Analysis Screening. SHRM; with HireVue and ORCAA audit announcements (2021). https://www.shrm.org/topics-tools/news/talent-acquisition/hirevue-discontinues-facial-analysis-screening
  • Trade press Fortune (2021, January 19). HireVue stops using facial expressions to assess job candidates amid audit of its A.I. algorithms https://fortune.com/2021/01/19/hirevue-drops-facial-monitoring-amid-a-i-algorithm-audit/
  • Vendor HireVue (2021, January 12). Industry Leadership: New Audit Results and Decision on Visual Analysis; and HireVue (2023, January). HireVue engages external auditor DCI Consulting Group for external bias audit of algorithms (press release) https://www.hirevue.com/blog/hiring/industry-leadership-new-audit-results-and-decision-on-visual-analysis
  • Investigative Schellmann, H. (2021, February 11). Auditors are testing hiring algorithms for bias, but there's no easy fix. MIT Technology Review https://www.technologyreview.com/2021/02/11/1017955/auditors-testing-ai-hiring-algorithms-bias-big-questions-remain/

• Two audit regimes reach this deployment and their perimeters are documented. The 2020 commissioned audit, announced 12 January 2021, examined one representative pre-built early-career assessment use case, did not examine the tool's technical design or its training data, proceeded largely by structured stakeholder interviews, and its report is published on the vendor's own site only behind a nondisclosure agreement; the phrase that the assessments work as advertised with regard to fairness and bias is the vendor's characterization of exactly that scope, and the recorded criticisms are that its depth compared unfavourably with a contemporaneous source-code audit of a competitor, that an auditor paid by the audited party carries a conflict risk, and that neither audit addressed whether the products improve hiring at all. The audit did produce recommendations, on investigating accent bias and on the flagging of candidates who give brief answers. From January 2023 the vendor engaged DCI Consulting Group for the annual bias audits New York City Local Law 144 requires, covering competency-based and game-based algorithms across race, gender, and intersectional groups. A summary produced 5 July 2023 and published by Pfizer as a deploying employer reports the mechanics: nationwide applicant data from January 2021 to December 2022, pooled across employer implementations, analysed per implementation, and then aggregated; on the Communication assessment for intern and new-college-graduate jobs, 20,060 male against 9,121 female applicants, selection rates of 0.66 and 0.65 for Top-plus-Middle against Bottom, a gender impact ratio of 0.98, race and ethnicity ratios from 0.87 to 0.96 and intersectional ratios down to 0.82; on the Adaptability assessment for the same population, 7,161 male against 3,884 female applicants, a gender ratio of 0.95 and intersectional ratios from 0.79 to 1.01. A peer-reviewed study of all 116 publicly available Local Law 144 bias audits published between July 2023 and November 2024 found that DCI conducted 20 percent of them, that 54 percent of audits carried at least one impact ratio above 1 — the majority of those being DCI audits of HireVue tools deployed by JetBlue, Citizens, Pfizer, or Burlington, an artifact of the aggregation and comparator-group method — and that every identified 'silent duplicate', meaning identical quantitative results republished across or within reports, appeared in DCI-conducted audits, all but one describing HireVue tools. Its authors could not determine the cause of all the duplicates. These are observations about a measurement regime; they are not findings of discrimination and not findings of audit fraud. An impact ratio is a ratio of selection rates over the applicants an audit could classify, and it is not an error rate — no error rate for this engine has ever been published by anyone.

  empirical
  • Investigative Schellmann, H. (2021, February 11). Auditors are testing hiring algorithms for bias, but there's no easy fix. MIT Technology Review https://www.technologyreview.com/2021/02/11/1017955/auditors-testing-ai-hiring-algorithms-bias-big-questions-remain/
  • Trade press Maurer, R. (2021). HireVue Discontinues Facial Analysis Screening. SHRM; with HireVue and ORCAA audit announcements (2021). https://orcaarisk.com/in-the-news/2021/1/12/orcaas-audit-of-hirevue-is-live
  • Government DCI Consulting Group (2023). New York City Local Law 144 bias audit for HireVue, summary produced 5 July 2023, published by Pfizer as a deploying employer https://cdn.pfizer.com/pfizercom/CareersEmploymentDocs/HireVue_2023_Bias_Report_14AUG2023.pdf
  • Academic Gerchick, M., Encarnacion, A., Tanigawa-Lau, C., Armstrong, L., Gutierrez, A., & Metaxa, D. (2025). Auditing the Audits: Lessons for Algorithmic Accountability from Local Law 144's Bias Audits. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT '25) https://facctconference.org/static/docs/facct2025-206archivalpdfs/facct2025-final14-acmpaginated.pdf

• Three external channels reached the vendor and produced three different kinds of result. On 6 November 2019 the Electronic Privacy Information Center filed a complaint with the Federal Trade Commission alleging that the AI-based candidate assessments constituted unfair and deceptive practices under Section 5, that the company falsely denied using facial recognition, and that its results were biased, unprovable, and not replicable; no public FTC enforcement action against the company is on the record as of August 2026, so what the record supports is that the vendor was the subject of an advocacy complaint and nothing stronger. Six Illinois residents filed Deyerler v. HireVue, Inc. on 27 January 2022 in the Northern District of Illinois (No. 1:22-cv-01284), alleging that the software collected facial geometry and voice data during virtual job interviews without the disclosures and written consent the Illinois Biometric Information Privacy Act requires; on 26 February 2024 Judge Jeremy C. Daniel granted in part and denied in part the motion to dismiss, letting claims under sections 15(a), (b) and (d) proceed, dismissing the section 15(c) profit claim on the ground that selling software is not selling biometric identifiers, and rejecting the argument that the Illinois Artificial Intelligence Video Interview Act precludes BIPA claims — holding the two statutes impose different but concurrent obligations. A claim that survives a motion to dismiss is an allegation held plausible and never a finding of violation. On 25 June 2026 the Circuit Court of Lake County, Illinois (No. 2026LA00000141, Hon. Daniel L. Jasica) granted PRELIMINARY approval of a 3,750,000 dollar class settlement covering an estimated 91,305 people who completed an interview involving the challenged voice and facial biometrics technology while in Illinois between 27 January 2017 and 25 June 2026, with an estimated 150 dollars per valid claimant subject to pro rata reduction, a claims deadline of 13 October 2026, and a final approval hearing on 28 October 2026. Nothing has been paid, the class was conditionally certified for settlement purposes only, the settlement is expressly no admission of wrongdoing, and the vendor denies that it collected or possessed biometrics or any other information subject to BIPA. On 19 March 2025 the ACLU of Colorado filed charges with the Colorado Civil Rights Division and the EEOC on behalf of a Deaf, Indigenous employee, alleging that an automated video interview relying on automated speech recognition disadvantaged her and that a request for human-generated captioning was denied; both companies dispute the charges and the vendor's chief executive called the complaint entirely without merit and stated that the employer did not use one of its AI-based assessments. No court and no regulator has ever found this vendor violated the biometric statute, the FTC Act, or any discrimination law.

  empirical
  • Advocacy Electronic Privacy Information Center (2019, November 6). In re HireVue: complaint and request for investigation filed with the Federal Trade Commission; complainant docket page returned HTTP 403 on 2026-08-28 and was not read, with the filing date and allegations corroborated by The National Law Review (2020) and MIT Technology Review (2019, November 7) https://epic.org/documents/in-re-hirevue/
  • Government Memorandum Order, Deyerler v. HireVue, Inc., No. 22 CV 1284 (N.D. Ill. 26 February 2024) (Daniel, J.), Dkt. 56 https://www.govinfo.gov/content/pkg/USCOURTS-ilnd-1_22-cv-01284/pdf/USCOURTS-ilnd-1_22-cv-01284-0.pdf
  • Government Notice of Proposed Class Action Settlement, Deyerler et al. v. HireVue, Inc., No. 2026LA00000141 (Circuit Court of Lake County, Illinois, Hon. Daniel L. Jasica; preliminary approval 25 June 2026), settlement administration by Simpluris https://www.classaction.org/media/hirevue-long-notice.pdf
  • Trade press ACLU, ACLU of Colorado, Public Justice & Eisenberg & Baum LLP (2025, March). Civil-rights charges over an automated HireVue video interview at Intuit (Deaf and Indigenous applicant); via HR Dive and Fisher Phillips coverage. https://www.hrdive.com/news/ai-intuit-hirevue-deaf-indigenous-employee-discrimination-aclu/743273/

• Both statutes that reach this deployment place their duties on EMPLOYERS rather than on the vendor, and that placement is the structural finding of the case. The Illinois Artificial Intelligence Video Interview Act (820 ILCS 42, effective 1 January 2020) requires employers using AI analysis of video interviews to notify applicants before the interview, explain how the AI works and what general types of characteristics it uses to evaluate them, and obtain consent to be evaluated; it restricts sharing of the video, requires deletion within 30 days of an applicant's request, and from 2022 requires employers relying solely on AI analysis to report applicant race and ethnicity data annually. Its text states no express enforcement mechanism and no private right of action — which is why the Deyerler plaintiffs sued under the biometric statute instead, and why the February 2024 holding that the video-interview statute does not displace the biometric one mattered: it kept open the only channel in this record with teeth. New York City Local Law 144 likewise places its annual independent-bias-audit and publication duty on employers. The vendor holds the pooled data and engages the auditor; the employers hold the duty and publish the result — which is how four named employers came to publish the same vendor-level numbers as their own audits. The vendor commissioned the audits voluntarily and argued publicly, in announcing the January 2023 engagement, that vendors should also bear audit responsibility. Nothing in this record describes either statute as regulating the vendor directly.

  empirical
  • Government Artificial Intelligence Video Interview Act, 820 ILCS 42 (Illinois General Assembly, effective 1 January 2020) https://www.ilga.gov/Legislation/ILCS/Articles?ActID=4015&ChapterID=68&Print=True
  • Government Memorandum Order, Deyerler v. HireVue, Inc., No. 22 CV 1284 (N.D. Ill. 26 February 2024) (Daniel, J.), Dkt. 56 https://www.govinfo.gov/content/pkg/USCOURTS-ilnd-1_22-cv-01284/pdf/USCOURTS-ilnd-1_22-cv-01284-0.pdf
  • Academic Gerchick, M., Encarnacion, A., Tanigawa-Lau, C., Armstrong, L., Gutierrez, A., & Metaxa, D. (2025). Auditing the Audits: Lessons for Algorithmic Accountability from Local Law 144's Bias Audits. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT '25) https://facctconference.org/static/docs/facct2025-206archivalpdfs/facct2025-final14-acmpaginated.pdf
  • Vendor HireVue (2021, January 12). Industry Leadership: New Audit Results and Decision on Visual Analysis; and HireVue (2023, January). HireVue engages external auditor DCI Consulting Group for external bias audit of algorithms (press release) https://www.hirevue.com/blog/hiring/industry-leadership-new-audit-results-and-decision-on-visual-analysis

• Independent scholarship bounds what automated video assessment can validly claim, and it measures the construct class rather than any one product. Hickman and colleagues, in the Journal of Applied Psychology in 2022, investigated automated video-interview personality assessments across a development sample of 1,073 and a retest sample of 99: models trained on self-reports showed little evidence of reliability or validity, while models trained on interviewer reports performed better but with mixed cross-sample reliability, and the authors cautioned vendors and adopting organizations accordingly. Raghavan and colleagues, at ACM FAT* in 2020, analysed what algorithmic pre-employment vendors publicly claim about validation and bias mitigation and found those claims largely unverifiable from what vendors disclose. Against that, a technology-press review of both audit regimes reaching this deployment recorded that neither addressed whether the products improve hiring at all. So the record contains no measurement of this engine's accuracy from any source: the mandated audits report selection-rate ratios, the commissioned audit did not examine the technical design or the training data, and no error rate has ever been published.

  empirical
  • Academic Hickman, L., Bosch, N., Ng, V., Saef, R., Tay, L., & Woo, S. E. (2022). Automated Video Interview Personality Assessments: Reliability, Validity, and Generalizability Investigations. Journal of Applied Psychology, 107(8), 1323-1351 https://doi.org/10.1037/apl0000695
  • Academic Raghavan, M., Barocas, S., Kleinberg, J., & Levy, K. (2020). Mitigating Bias in Algorithmic Hiring: Evaluating Claims and Practices. In Proceedings of FAT* '20, 469-481. https://doi.org/10.1145/3351095.3372828 https://arxiv.org/abs/1906.09208
  • Investigative Schellmann, H. (2021, February 11). Auditors are testing hiring algorithms for bias, but there's no easy fix. MIT Technology Review https://www.technologyreview.com/2021/02/11/1017955/auditors-testing-ai-hiring-algorithms-bias-big-questions-remain/

Where this connects

Institutional pressures in this domain

• Workload surge — Demand outruns staffing; per-case attention shrinks and review becomes triage.
• Vendor opacity — The deploying institution cannot inspect the model, data, or update pipeline it is accountable for.
• Compliance over substance — Paper controls (sign-offs, checklists) satisfy audits while the behavior they describe erodes.
• Data & policy drift — The world, the intake process, and the rules change under a system trained on how things used to be — two mechanisms with different remedies: the statistical properties of what the system processes move (concept drift), or the mixture of inputs arriving in deployment differs from the mixture it was trained on (covariate shift).
• Reviewer bottleneck — One fixed-capacity checking stage sits between AI output and consequence; everything queues behind it.

All of them in context on the Hiring & employment screening AI domain page.

Levers available here and the patterns behind them

• Store less data — Data minimization
• Upgrade model — Improve the model
• Gate vendor updates — Vendor quality gate
• Understand the system — Understand the system
• Check with a second model — Cross-model verification
• Check copied records — Reconcile copied records
• Mark AI-written records — Provenance labeling
• Gate record entries — Human-in-the-loop write gating
• Review on schedule — Oversight cadence & retrospectives
• Pause AI on alarms — Deployment circuit-breaker
• Vet connections — Connection authorization

Documented case histories

• HireVue video assessment (vendor layer)
• A resume screener that learned the past's bias
• Vendor screening across thousands of employers (litigation live)
• Graduate-hiring AI with its audits on the record
• The 1959 statute and the integrity video screen (Baker v. CVS Health)
• An internal promotion, a recorded screen, and a captioning request (D.K. charges against Intuit and HireVue)
• Aon pre-hire assessment suite (vendor's own tables)
• The cooperative audit: a paid source-code examination, and what happened to its verdict
• McHire and the 64-million-record custody exposure
• SiriusXM's iCIMS applicant screening
• Checkr gig-economy background screening
• The rule with no number to disclose
• The account goes dark at nine; the reason arrives on day twenty-six
• iTutorGroup Tutor Application Screen
• Meta Job-Ad Delivery: the guardrail and the layer below