PAN Lab example

CVS Health's Massachusetts applicant video-interview screen

Integrity video screening under a 1959 lie-detector statute: the employer's notice duty

A national retail employer routes applicants for its positions in one state into a recorded video interview run on a third-party platform. The questions are about integrity: what it means to the applicant, and an occasion when the applicant acted with it. As pleaded in a class complaint and recited in a federal court's published opinion, the recordings are uploaded onward to a further analysis firm whose models read facial expressions, eye contact, voice intonation and inflection, and the platform returns an employability score to the employer's recruiters. The contested judgement is therefore made two organizational hops from the party that carries the legal duty, by two companies that were never parties to the case. What was litigated was a missing sentence. This state has banned lie detector tests in employment since 1959, and since a 1985 amendment every employment application in it must carry, in clearly legible print, a one-sentence notice saying that requiring or administering such a test is unlawful and that an employer who does faces criminal penalties and civil liability. The provision then went roughly forty years without being enforced. The complaint alleged the application carried no notice, and in February 2024 the court held that the statute's private right of action reaches the notice provision and that being denied the information was a concrete injury: the notice would have specifically informed the applicant that the interview was a lie detector test, and he sat it without that warning. Both of the employer's motions were denied in full. The reason a polygraph-era statute reaches an inference model is a drafting choice. The banned instrument is defined by what it PURPORTS to do — any device, mechanism, instrument or written examination used for the purpose of purporting to assist in the detection of deception, the verification of truthfulness, or the rendering of a diagnostic opinion on honesty. So the platform's own marketing, claiming lie detection and the screening out of embellishers, is what brought the assessment inside a definition written for a machine with a cuff and a needle. The vendor's sales copy was the load-bearing evidence, and the vendor's engineering was never examined at all. Nothing was ever decided. The employer did not challenge the sufficiency of the counts alleging the screen itself was a lie detector, so the court accepted that characterization as plausibly pleaded rather than ruling on it. The employer then settled individually on confidential terms before any class-certification ruling, and the case was dismissed with prejudice in September 2024. What propagated instead was the pleading: more than twenty follow-on suits in about a year under the same statute, largely by one firm, most alleging only a missing notice and no screening device at all. Notice lines are now appearing on employment applications across the state — driven by per-application statutory damages rather than by any finding about any technology. Before you pick a target level: this board cannot be won under Service and Safety Targets or All Governance Targets, and a bigger budget buys nothing here. Take every instrument this employer could actually reach, set each one to full strength and ignore the budget entirely, at a total of thirty against the six you are given. Eight pathways are open at the end of that — one more than the floor, because one instrument on offer sends more of the decision through the screen. The floor is seven, and reaching it costs exactly the six you have. Four of the seven belong to the vendor chain: the score arriving at a recruiter, the features and the score written to a record the scored person has no access to, and the platform writing what it captures. The other three are the correction itself — the notice going onto the form, the statutory price landing on it, and the case being written down in public. Closing those three would mean the one thing that improved here never happening. Widening past what this employer held does move the number, and the two instruments that move it furthest are a sign-off over what the vendor writes and a rule about how much it keeps — neither of which a buyer holds over an engine it does not own. Explore and Service Targets Only can be won, and cheaply: one instrument, costing three of your six.

Stylized model of a documented deploymentHiring & employment screening AI

Open this example in PAN Lab v0.1 to apply pressures and levers and watch what the system does.

What this models

This example runs on the Employer-liability-class integrity video screen 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: 4 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
Every mechanical description on this board is an allegation of an amended complaint that a court accepted as plausible on a motion to dismiss. The employer did not challenge the sufficiency of the counts alleging that the screen itself was a lie detector, so the court accepted that characterization as plausibly pleaded rather than deciding it. There is no merits finding that this interview was a lie detector test, and there is no finding of any kind about the accuracy of anything drawn here.
baseline
The two inference layers are drawn as separate nodes one hop apart because the case is about the distance between them and the party that answered for them. The PAN entry for this deployment carries the same two components and states in its own comment that it lacks a model-to-model edge kind and carries the composition as a store hop instead; the Lab has the kind, so the hop is drawn where the record puts it.
assumed
The recorded answers are drawn as their own input rather than folded into the vendor-held record, which is where the PAN entry keeps them. The split is disclosed because it is a Lab-side refinement: in this deployment the capture and the derived record sit at different companies, so folding them would hide the organizational hop the whole case turns on.
baseline
The court record is drawn as a third store, which the PAN entry does not carry. The dossier states the reason in terms: the litigation record became the public description of the system, and the court's recitation of the pleaded mechanics is the authoritative public account because neither vendor's internal pipeline was ever put in evidence. A store that holds the only public description of a deployment belongs on the diagram of it.
baseline
One reviewer is drawn, and the count is derived rather than chosen. Enforcement of this statute is a civil action by any person aggrieved and nothing else: no regulator investigated this deployment, no agency acted, and no audit of the screen or of the vendor chain exists in the record. The PAN entry leaves its regulator and auditor slots deliberately empty for the same reason.
assumed
Applicants are outside the network, as screened people are on every Lab diagram. That places a strain on this board worth stating: the three deprivations the complaint pleads — the notice, the chance to decline, and a channel to challenge the assessment — are all channels to a person who has no node here. They are carried as the state of the application form and as the checks drawn at zero, and the pathway from a form to the person reading it is narrated rather than drawn.
baseline
Three widths are drawn at zero and each rests on a documented absence rather than on silence: no accuracy figure or independent evaluation of this deployment exists from any source; no employer-side review of the vendor chain's inferences or of the legality of running an integrity screen in this state appears in the record; and the public account of the screen cannot be reconciled against the pipeline it describes, because neither vendor's pipeline was ever put in evidence.
assumed
Every other width comes from the PAN entry for this deployment on one stated rung mapping, applied to the eighteen pathways that have a one-to-one counterpart there with no exceptions. Three further pathways that had such a counterpart were folded into pathways already drawing the same documented flow when this board was re-derived at a coarser granularity, and each fold is recorded in the entry below. Three of those are PAN peer edges redrawn as Lab checks, because a channel that improves practice is inhibiting in this vocabulary and reinforcing in PAN's; their widths still come from PAN. Every PAN edge for this deployment is marked estimated, so the drawn strengths are modelling choices in both files.
baseline
The standing workload reads high on the statute's own structure rather than on a volume figure. The duty attaches to every Massachusetts employment application, the pleaded class is everyone who applied for a position at this employer in the state, and each violation carries a separate statutory minimum. No application volume for this employer is in the verified record, and the widely-quoted hundred-thousand-dollar annual figure is a compliance-press illustration of the arithmetic, not a measured exposure for this or any employer.
assumed
The human comparator reads at the catalogue default, and the refusal to read it higher is deliberate. The score was advisory input to real recruiters and the opinion records no score-to-decision weighting; there is no head-to-head comparison of this screen against the interview it stood in front of anywhere in the record. The expert critique recorded at filing is about whether any automated instrument can detect deception, and it is not used as a measurement of a human process.
baseline
The platform vendor's timeline is carried alongside the pleaded mechanics rather than resolved against them. Visual analysis was removed from new assessment models in about March 2020 and the removal was announced in January 2021; the application in this case was made in about January 2021; whether affect analysis in fact ran on it was never adjudicated. Nothing on this board asserts that it did, and nothing asserts that it did not.
baseline
No liability or wrongdoing by either vendor is asserted anywhere on this board. Neither was a party, no claim was ever adjudicated against either, and the platform vendor disputes the deception-detection characterization on the record. The employer settled individually on confidential terms with no admission, so nothing here asserts liability by the employer either.
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 this deployment, and it lands at ten nodes and twenty-five pathways. No node changed: the two inference layers stay separate because the distance between them is the case, the three stores stay separate because a documented finding attaches to each one, and the three operator classes act on the error in three documented ways. Three pathways were folded into pathways that already carried the same documented flow, and every fact each one carried is stated on its survivor. First, the affect reading arriving at the platform is carried by the step that composes it, because the platform is the party that composes it. Second, the recruiter reading the score is carried by the score arriving at the recruiter, because the store is held on the vendor side and the opinion documents the employer consuming the output. Third, the application entering the screen is carried by the form creating the scored record, because both say the form is the entry condition for what the screen produces. Each fold joined two drawings that sat on the same rung, so no width and no intensity claim moved anywhere on this board.

What this example does not show

Concluded as to CVS: individual confidential settlement before any class-certification ruling; voluntarily dismissed with prejudice September 20, 2024. No merits adjudication that the screen WAS a lie detector, and no classwide relief. The Baker ruling seeded an ongoing statewide litigation wave under the same statute against other employers.
Everything mechanical here is pleaded, not found. The upload to the affect-analysis firm, the affect features, the employability score and the honesty inference are allegations of an amended complaint accepted as true on a motion to dismiss. The employer did not challenge the sufficiency of the lie-detector counts at that stage, so the court accepted that characterization as plausibly pleaded rather than deciding it.
The platform vendor's public timeline runs the other way and is carried alongside rather than resolved: visual analysis was removed from new assessment models in about March 2020, announced in January 2021, and the vendor told a newspaper that visual and audio analysis had since been eliminated. The application in this case was made in about January 2021. Whether affect analysis ran on it was never adjudicated, and nothing here asserts that it did.
The vendors were never parties. No claim was adjudicated against the interview platform or the analysis firm, and the platform disputes the deception-detection characterization on the record, saying its assessments measure work competencies statistically linked to job success using validated industrial and organizational psychology. Nothing on this board asserts liability or wrongdoing by either.
No number here measures the screen. There is no published error rate, no accuracy figure and no independent evaluation of this deployment from any source. The 500-dollar figure is a statutory minimum per violation, not a payment: the case settled confidentially and no class member received anything. The hundred-thousand-dollar annual exposure figure that circulates in compliance writing is an illustration of the arithmetic — two hundred applications at five hundred dollars — and is not a measured exposure for this or any employer.
This board is the DEPLOYER's liability under a state employment statute. The vendor-layer board for the same interview platform is a different deployment with different evidence — its own audits, its own regulator complaint and its own consent litigation — and nothing about that party's control surface, customer base or product governance is imported here.
The people this screen sorted are outside the dynamics, as screened people are on every Lab diagram. No hiring outcome, score or rejection for any person is computed from anything drawn here. The three deprivations the complaint pleads — the notice, the chance to decline, and a channel to challenge the assessment — are channels to people who have no node on this board, and they are carried as the state of a form and as pathways drawn at zero.

Sources and evidence

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

Around January 2021 a Milton, Massachusetts resident applied for a CVS supply chain position, sat a HireVue video interview, and was not hired. As recited in the court's published opinion from the amended complaint, the interview asked integrity-framed questions — what integrity means to the applicant, and a time the applicant acted with integrity — and HireVue uploaded the recordings to Affectiva, a Boston affect-analysis firm spun out of the MIT Media Lab, whose AI 'analyzes candidates' facial expressions, eye contact, voice intonation, and inflection' to draw conclusions about the applicant's degree of cultural fit; HireVue then provided CVS with employability scores. The amended complaint lists the affect features as smiles, surprise, contempt, disgust, and smirks, and describes the score as rating traits including conscientiousness and responsibility and an innate sense of integrity and honor. Every mechanical element of that description is an allegation accepted as true for the purpose of a motion to dismiss; none of it was ever adjudicated, and no error rate, accuracy figure, or independent evaluation of this screen exists from any source.
empirical
- Government Baker v. CVS Health Corporation, Civil Action No. 23-11483 (D. Mass. 16 February 2024) (Saris, J.), memorandum and order on motions to dismiss https://caselaw.findlaw.com/court/us-dis-crt-d-mas/115835556.html
- Trade press HR Dive (Crist, C.) (2024, July 24). CVS settles lawsuit alleging it used AI 'lie detector' https://www.hrdive.com/news/cvs-settles-lawsuit-over-using-ai-based-lie-detector/722249/
- Investigative The Boston Globe (Johnston, K.) (2023, May 22). A Milton resident's lawsuit against CVS raises questions about the use of AI lie detectors in hiring https://www.boston.com/news/the-boston-globe/2023/05/22/milton-residents-lawsuit-cvs-ai-lie-detectors/
Massachusetts General Laws chapter 149, section 19B has banned lie detector tests in employment since 1959, and a 1985 amendment added both a private civil action and a mandatory notice. The statute defines the banned instrument by PURPORTED function: it covers 'any test utilizing a polygraph or any other device, mechanism, instrument or written examination' used 'for the purpose of purporting to assist in or enable the detection of deception, the verification of truthfulness, or the rendering of a diagnostic opinion regarding the honesty of an individual.' Every Massachusetts employment application must carry, in clearly legible print, one sentence: 'It is unlawful in Massachusetts to require or administer a lie detector test as a condition of employment or continued employment. An employer who violates this law shall be subject to criminal penalties and civil liability.' Remedies run to a civil action within three years by any person aggrieved, not less than 500 dollars in damages per violation, treble damages for lost wages or benefits, costs, and reasonable attorney fees, with criminal penalties of 300 to 1,000 dollars for a first violation and up to 1,500 dollars and 90 days thereafter. The notice provision then went roughly forty years without being enforced, until filings began in 2023. The 500-dollar figure is a statutory minimum, not a payment made to anyone in this case.
empirical
- Government Massachusetts General Laws, Part I, Title XXI, Chapter 149, Section 19B (lie detector tests; employment), enacted 1959, notice and private action added 1985 https://malegislature.gov/Laws/GeneralLaws/PartI/TitleXXI/Chapter149/Section19B
- Trade press Agency Checklists (2025, June 2). Mass. Insurance Employers: A Hidden Trap That Could Cost You $500 Per Applicant https://agencychecklists.com/2025/06/02/mass-insurance-employers-a-hidden-trap-that-could-cost-you-75990/
- Trade press Morgan Lewis LawFlash (Engelman, K. L., Schuster, S., & Perocchi, A. K.) (2025, May 20). Class Action Litigation Over Massachusetts Lie Detector Statute Surges: What Employers Need to Know https://www.morganlewis.com/pubs/2025/05/class-action-litigation-over-massachusetts-lie-detector-statute-surges-what-employers-need-to-know
On 16 February 2024 Judge Patti B. Saris denied both of CVS's motions in their entirety in Baker v. CVS Health Corporation, No. 1:23-cv-11483 (D. Mass.) — the failure-to-state-a-claim motion aimed at the notice count and the separate Article III standing motion. No count and no defendant was dismissed. The court held that the statute's private right of action for any person aggrieved reaches notice violations, and that denial of information to which a plaintiff has a legal right can be a concrete injury in fact: the required notice 'would have specifically informed Baker that the HireVue Interview was a lie detector test,' and he participated without that warning. CVS did not challenge the sufficiency of the counts alleging that the screen itself violated the lie-detector prohibition, so the court accepted that characterization as plausibly pleaded rather than deciding it. This is a pleading-stage holding. There is no merits finding that the HireVue interview was a lie detector test, and no finding of any kind about the screen's accuracy.
empirical
- Government Baker v. CVS Health Corporation, Civil Action No. 23-11483 (D. Mass. 16 February 2024) (Saris, J.), memorandum and order on motions to dismiss https://caselaw.findlaw.com/court/us-dis-crt-d-mas/115835556.html
- Trade press ArentFox Schiff (Davidson, S.) (2024, May 20). Using New Technology to Vet Job Applicants Can Violate the Massachusetts Lie Detector Statute; and New England Biz Law Update (2024, February 23). Federal judge allows lawsuit against CVS to proceed over alleged lie detector use https://www.afslaw.com/perspectives/ai-law-blog/using-new-technology-vet-job-applicants-can-violate-the-massachusetts-lie
Liability in this case landed on the employer while the two companies that built and ran the contested inference sat outside it. The defendants were CVS Health Corporation and CVS Pharmacy, Inc. only; HireVue and Affectiva were never parties and no claim was ever adjudicated against either. On the pleaded account the inference executed two organizational hops from the party bearing the statutory duty: the recordings were captured on the interview platform, uploaded onward to the affect-analysis firm, composed into an employability score at the platform, and returned to CVS recruiters. The statutory duties of section 19B bind the employer, not the vendors. The applicant-facing surface was the recorded interview alone: the complaint pleads three deprivations — no notice, no opt-out, and no ability to challenge the assessment results — and the record documents no applicant access to the recording, the affect analysis, or the score, and no feedback of hiring outcomes into retraining. Because neither vendor's internal pipeline was ever put in evidence, the court's recitation of the pleaded mechanics is the authoritative public account of how this system worked.
empirical
- Government Baker v. CVS Health Corporation et al, No. 1:23-cv-11483 (D. Mass.), public docket record via PacerMonitor https://www.pacermonitor.com/public/case/49427403/Baker_v_CVS_Health_Corporation_et_al
- Government Baker v. CVS Health Corporation, Civil Action No. 23-11483 (D. Mass. 16 February 2024) (Saris, J.), memorandum and order on motions to dismiss https://caselaw.findlaw.com/court/us-dis-crt-d-mas/115835556.html
- Trade press HR Dive (Crist, C.) (2024, July 24). CVS settles lawsuit alleging it used AI 'lie detector' https://www.hrdive.com/news/cvs-settles-lawsuit-over-using-ai-based-lie-detector/722249/
The case ended without deciding anything about the technology. A settlement notice was filed 17 July 2024; the public docket shows the case terminated 22 July 2024; a stipulation of voluntary dismissal with prejudice followed on 20 September 2024. The settlement was individual and confidential and was reached before any class-certification ruling: no monetary terms, no practice changes, and no admission of liability were disclosed, and no class member received anything. The named claim was extinguished and the pleaded class went unrepresented. Both docket dates are carried here rather than a single 'settled' date. The mirrored opinion does not recite an original filing court or removal path, so the filing history is stated only as reported by the Boston Globe on 22 May 2023 and on the federal docket from 30 June 2023.
empirical
- Trade press HR Dive (Crist, C.) (2024, July 24). CVS settles lawsuit alleging it used AI 'lie detector' https://www.hrdive.com/news/cvs-settles-lawsuit-over-using-ai-based-lie-detector/722249/
- Government Baker v. CVS Health Corporation et al, No. 1:23-cv-11483 (D. Mass.), public docket record via PacerMonitor https://www.pacermonitor.com/public/case/49427403/Baker_v_CVS_Health_Corporation_et_al
- Investigative The Boston Globe (Johnston, K.) (2023, May 22). A Milton resident's lawsuit against CVS raises questions about the use of AI lie detectors in hiring https://www.boston.com/news/the-boston-globe/2023/05/22/milton-residents-lawsuit-cvs-ai-lie-detectors/
After the February 2024 ruling, litigation under section 19B surged: more than twenty class actions in roughly a year, largely by a single New York firm, often with the same individuals suing multiple employers, including Procter & Gamble. Most allege ONLY that an employment application lacked the statutory notice, with no actual screening device involved — so what propagated was the notice theory rather than any finding about AI screening. Compliance advisories now direct every Massachusetts employer to print the statutory notice in the application itself rather than in a policy filed elsewhere, and illustrate the exposure arithmetically: an employer processing 200 Massachusetts applications a year faces roughly 100,000 dollars of annual notice-violation exposure at the 500-dollar statutory minimum. That figure is an illustration of the statutory structure, not a measured exposure for CVS or any other employer, and no CVS Massachusetts application volume or class size appears in the verified record. The system-level correction now visible — notice lines appearing on Massachusetts employment applications — is driven by per-application statutory damages and copycat litigation pressure rather than by any adjudicated finding about the technology.
empirical
- Trade press Morgan Lewis LawFlash (Engelman, K. L., Schuster, S., & Perocchi, A. K.) (2025, May 20). Class Action Litigation Over Massachusetts Lie Detector Statute Surges: What Employers Need to Know https://www.morganlewis.com/pubs/2025/05/class-action-litigation-over-massachusetts-lie-detector-statute-surges-what-employers-need-to-know
- Trade press Agency Checklists (2025, June 2). Mass. Insurance Employers: A Hidden Trap That Could Cost You $500 Per Applicant https://agencychecklists.com/2025/06/02/mass-insurance-employers-a-hidden-trap-that-could-cost-you-75990/
Two things in this record cut against the pleaded account and both are carried rather than resolved. HireVue removed facial analysis from new assessments in March 2020 and announced the change in January 2021 alongside a commissioned algorithmic audit, retaining speech-and-language analysis; it told the Boston Globe that visual and audio analysis 'have since been eliminated,' and its chief data scientist rejected the deception-detection characterization, saying the assessments measure work competencies 'statistically linked' to job success using 'validated industrial organizational psychology.' The application in this case was made around January 2021, and whether Affectiva affect analysis in fact ran on it, or on any class member's interview, was never adjudicated. Separately, the scientific question the 1959 statute was answering was never reached either: Brandeis psychologist Leonard Saxe told the Globe that there is no neurological signal of deception and 'no way for an automated system to distinguish a falsehood from the truth.' That is an expert view recorded in a newspaper, not a finding, and it is a statement about the construct rather than a measurement of this or any deployment.
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
- Investigative The Boston Globe (Johnston, K.) (2023, May 22). A Milton resident's lawsuit against CVS raises questions about the use of AI lie detectors in hiring https://www.boston.com/news/the-boston-globe/2023/05/22/milton-residents-lawsuit-cvs-ai-lie-detectors/

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

Mark AI-written records — Provenance labeling
Gate vendor updates — Vendor quality gate
Keep prompts neutral — Framing and mirroring reduction
Review on schedule — Oversight cadence & retrospectives
Peer sharing rules — Peer-edge governance
Understand the system — Understand the system
Check copied records — Reconcile copied records