HireVue scored job candidates partly by their faces, then announced in 2021 it had dropped the practice

A hiring score built partly from a face

HireVue sold asynchronous video interviews. A job candidate answered an employer's questions on camera, alone. Software scored the recording on word choice, tone of voice, and, until HireVue removed it, facial movements, then ranked the candidate against others by an "employability" score.¹ It collected what the company called "tens of thousands of data points" from each interview and said the score revealed a candidate's cognitive ability, psychological traits, emotional intelligence, and social aptitude.²

By late 2019 the assessment had reached more than 100 companies and over a million applicants.¹ HireVue counted more than 700 corporate customers overall, among them Hilton and Unilever.²

EPIC brings it to the FTC

On November 6, 2019, the Electronic Privacy Information Center filed a complaint with the Federal Trade Commission, asking it to investigate HireVue for unfair and deceptive practices under Section 5 of the FTC Act.² The deceptive-practices claim turned on a question and answer HireVue published for candidates. Asked whether the product was facial recognition, the company answered: "No. HireVue does not use facial recognition technology or track facial features for identity recognition purposes."² EPIC argued that answer had no reasonable basis. The FTC's own 2012 definition of facial recognition covered technology that analyzes facial geometry to predict expression or emotion, and by EPIC's account HireVue's system did that.²

How much the face counted was no secret in 2019. HireVue's own executives had said so. Loren Larsen, the chief technology officer, put 10 to 30 percent of a candidate's score on facial expressions. Nathan Mondragon, the chief industrial-organizational psychologist, said "facial action units," the individual facial-muscle movements a standard coding system catalogs, could make up as much as 29 percent.² The public answer denied tracking facial features while the company's own leaders described the face as up to roughly a third of the score.

EPIC's unfairness claim was broader: secret, unproven algorithms scoring biometric data caused a substantial injury that job candidates could not reasonably avoid.² The law was only starting to reach any of this. Illinois had passed the Artificial Intelligence Video Interview Act, Public Act 101-0260, now at 820 ILCS 42, signed in August 2019 and effective January 1, 2020.³ It requires an employer to tell an applicant that AI may analyze the interview, to explain in plain terms how the AI works and what characteristics it uses, and to get consent. It also limits who may see the video and requires deletion within 30 days on request.³ What it does not require is that the AI be valid, accurate, or free of bias. It governs notice and consent, not whether the measurement means anything.³

What a face cannot report

A score can measure something real and still not measure what it claims. That gap has a name: construct validity, the question of whether a measurement captures the trait it is supposed to rather than some other thing that happens to move alongside it. HireVue's score plainly measured something. It read muscle movements, vocal features, and word patterns. The claim that turned it into a hiring tool was that those signals track competence and character.

The smaller version of the claim is whether a face reveals an emotion at all. In 2019 five researchers led by the psychologist Lisa Feldman Barrett published the most systematic review of that question to date. They examined how people move their faces and how observers read them across many cultures, in infants and children, and in people born blind, for six basic emotion categories.⁴

People do move their faces in emotion-linked ways more often than chance would give, but a given configuration, a scowl or a smile, is too variable and too dependent on context to be read back as one specific inner state.⁴

HireVue was making a larger claim than the one the review tested. Barrett and her colleagues asked whether a face reveals an emotion. HireVue said a face and a voice reveal traits and the competence to do a job, which sit a step further from the muscles than an emotion does.² The review runs no hiring study, and it does not need to. It removes the thing HireVue's claim stood on, that facial movements reliably and specifically reveal an inner state. Take that away and there is no established path from a candidate's expressions to a valid measure of how they would perform.⁴

An invalid measure still changes lives once people act on it. People tend to trust a number from a machine more than it has earned, a tendency called automation bias. HireVue's ranking was built to be acted on: it sorted applicants so employers could decide whom to carry forward.² A score with no established link to job performance stood in for a judgment about a person.

HireVue calls the removal a value decision

HireVue announced on January 12, 2021 that it had removed the visual-analysis component from its new assessments early in 2020, releasing an outside audit the same day.⁵ The company framed the removal as its own decision, taken because advances in natural-language processing had raised the predictive power of language analysis, so that, in its words, "visual analysis no longer significantly added value to assessments."⁵ The language-based scoring stayed.

By 2021 the company's account of the face's value had changed. Kevin Parker, then chairman and chief executive, told Fortune the visual signal had contributed "about 0.25%" of a model's predictive power in most cases, and up to 4 percent for customer-facing roles. In his words, "it wasn't worth the incremental value we might have been getting from it." About 20 percent of HireVue's customers had used the predictive scoring at all.⁶

Both accounts can hold at once, because they do not measure the same thing. A component can be a large fraction of a score and still add little once the other inputs are in place.⁶ The change in framing does not settle the question EPIC and the science raised. HireVue never conceded that facial scoring was baseless or biased. It presented the removal as a matter of value, not of validity.⁵

The audit released alongside the announcement did not fill that gap. Conducted by the firm ORCAA, it examined one product: the assessment many companies use to screen recent college graduates. As MIT Technology Review reported, ORCAA "didn't evaluate the technical design of the tool itself." It interviewed stakeholders about potential harms and gave HireVue recommendations.⁷ The full report sits on HireVue's website behind a nondisclosure agreement.⁷ ORCAA's own public notice of the audit points readers to that report rather than stating its findings.⁸ The audit tested a fairness process. It did not test whether the scoring measures what it claims. HireVue still summarized it as finding its assessments "work as advertised with regard to fairness and bias issues," a broader conclusion than a review of one assessment's stakeholder impacts can carry.⁵

At the time of the announcement HireVue reported cumulative totals of more than 18 million video interviews and 114 million chat-based engagements for its more than 700 customers.⁵ That is a platform-wide count, not the reach of the AI video scoring, which a smaller share of customers used.⁶ What would settle whether the scoring ever measured job competence is the one thing neither the removal nor the audit provides: an independent test of the scores against real job performance. The audit did not attempt one.⁷

The same inference, still for sale

The specific product changed, but the inference behind it did not go out of use. Barrett's review named that inference as a live commercial application: the belief that a facial configuration maps onto an inner state, it wrote, "continues to fuel commercial applications in industry and government," and it singled out automated detection of emotions from faces.⁴

The tell is always the same leap, from something a sensor can record to something inside a person that it cannot. A camera can measure a brow. It cannot read the state behind it, and the science says no fixed reading exists.⁴ A hiring tool, a proctoring service, or a workplace monitor that claims to score attention, sincerity, or aptitude from a face or a voice faces the question HireVue's score could not answer: what evidence links the signal it measures to the thing it says it found?