Domain Atlas / Hiring & employment screening AI
An internal promotion, a recorded screen, and a captioning request (D.K. charges against Intuit and HireVue)
Explore this deployment in the PAN Lab ↗
In the PAN Lab, the readouts of this case's model organization carry a shaded evidence band whose width follows the least-established class among the modeling inputs the readings rest on.
The least-established input behind this case's model organization's readings is an assumption, not a measurement. Evidence base: 5 assumed · 8 published baseline.
On 19 March 2025 the ACLU, the ACLU of Colorado, Public Justice and Eisenberg & Baum LLP filed a Complaint of Discrimination with the Colorado Civil Rights Division and the U.S. Equal Employment Opportunity Commission on behalf of D.K., a Deaf Pawnee woman who communicates in English with a deaf accent and in ASL, against BOTH Intuit, Inc. and HireVue, Inc. It alleges violations of the Colorado Anti-Discrimination Act (Colo. Rev. Stat. § 24-34-402), the Americans with Disabilities Act and Title VII in the denial of her promotion to Seasonal Manager, and pleads HireVue as an employment agency, an agent of the employer, an indirect employer and an aider and abettor under state law. None of those theories has been tested. Everything the charge asserts is an allegation and no probable-cause determination, dismissal, right-to-sue notice, court filing or settlement has been made public.[3]
What happened
Start with what the employer already knew. Per the charge and her own published account, D.K. was hired as a seasonal Tax Associate in late 2019. By 2021 she was a Tax Expert Lead, supporting a team of roughly four hundred Tax Associates, and she held that role for three seasons. She earned a bonus every year. Her supervisors praised her communication. In 2023 she joined Intuit's Accessibility Team. When she applied for Seasonal Manager in spring 2024, one of the people on the hiring team was her own manager, who had encouraged her to apply. Five seasons of direct, in-house measurement of exactly the qualities a promotion decision is about, held by the party making the decision.
On 21 June 2024 she was invited into the assessment step. As described in the complaint it runs about three hours: roughly a dozen timed recorded video questions, mostly management scenarios with a few tax-law questions, plus essay and multiple-choice sections. The instructions and the questions are delivered audibly, by a recording of a person speaking, and portions of it are alleged to have carried no subtitles at all.
D.K. is Deaf. She communicates in English with a deaf accent and in ASL. The invitation email offered only a technical-support address and carried no accommodations information, so she initiated a request herself, from familiarity with the process. She asked for Communication Access Real-time Translation (CART) — human-generated real-time captioning — and deliberately not for ASL interpretation, because interpreters she had worked with before could not handle tax concepts. She did not ask to skip the video interview: it was never offered, and she says she feared that asking would count against her. The response, as alleged, was to deny CART and tell her the software's built-in subtitles could be turned on. When she started the assessment, she says, no subtitle option existed. She completed the three hours on her browser's automatic captions, which she describes as less accurate, particularly on complicated words.
That sequence is the employer-side allegation in full, and it stands independent of everything that follows. A process existed. She used it. It produced an answer. The answer named a control, and the control was not there when it mattered.
What happened to her recorded answers is the part the parties dispute at its foundation. The complaint pleads a three-stage mechanism: automatic speech recognition of the spoken answers, a second system interpreting the transcript, and machine-learning scoring against job competencies. It pleads this on information and belief, inferring it from one artifact. On 13 August 2024 an automated rejection told her Intuit had decided to move forward with other candidates. On 25 September a feedback message advised her to give more concise and direct answers, to adapt her communication style to different audiences, and to practise active listening. From its return address and its generic register she believed a machine had written it, and the complaint reads that advice as a screen restating a disability as a deficiency.
HireVue's chief executive answers flatly that the complaint is entirely without merit, that it rests on an inaccurate assumption about the technology used in the interview, and that Intuit did not use a HireVue AI-based assessment. Intuit separately says the allegations are entirely without merit and that it provides reasonable accommodations to all candidates. Nothing has been determined. If the vendor's account is right, the same facts describe human reviewers who watched a recorded interview and let its result stand over five seasons of their own evidence — which changes what the case is about without touching the accommodation allegation at all.
On 31 October 2024 D.K. exercised her right under Colorado law to inspect her personnel file, to learn how the decision had been made. The complaint, filed the following March, alleges that only minimal documents were produced.
Two earlier events make the charge a notice case as well as a screening case. After the 2023 season, as a member of Intuit's own Accessibility Team, she told the Team's Chair that the HireVue step could exclude deaf applicants. She was told it would be looked into. Nothing was ever reported back, and the following year the same step was required of her. And in 2020, per the complaint, Intuit's own AI call-monitoring system — which scored how closely agents followed scripts, from speech-recognition transcripts of customer calls — read her deaf accent as deviation from script and produced one artificially low indicator against otherwise strong customer-satisfaction, resolution-count and response-speed measures. Intuit's response, as alleged, was to reassign her from phone calls to chat. The metric was not corrected, the low indicator stayed in her record, and the complaint pleads it may also have been considered when the promotion was denied.
The technical spine of the charge is published research, and it is worth stating at the scope the research has. A 2025 study in The Laryngoscope measured four widely used commercial speech-recognition services against a twenty-four-speaker corpus and found a mean word error rate of 52.6 percent for d/Deaf and hard-of-hearing speakers, against 5.0 percent for normal-hearing speakers — the authors' own "ten times higher" — rising to 85.9 percent for speakers with the lowest speech-intelligibility classification. The complaint cites it, along with peer-reviewed findings of higher error rates for non-white speakers, which it extends on information and belief to speakers of Indigenous dialects of English. That study measured four general-purpose services. It did not measure HireVue's system, or this interview, and the step from that literature to this platform is precisely what is contested.
On 19 March 2025 the ACLU, the ACLU of Colorado, Public Justice and Eisenberg & Baum LLP filed the Complaint of Discrimination with the Colorado Civil Rights Division and the EEOC. As of 28 August 2026 no probable-cause determination, dismissal, right-to-sue notice, court filing or settlement had become public. Administrative investigations are confidential, so seventeen months of silence means no public development, and not that nothing happened.
The sociotechnical reading
Read this deployment as a question about what an employer does with what it already knows. Every other automated screening case in this domain is about a stranger: an applicant the employer has never met, about whom the screen is the only evidence there is. That is the standard argument for a standardised gate, and it is a serious one. This case is the other situation entirely. The employer had five seasons of direct measurement of this person — her customer-satisfaction scores, her resolution counts, her response speed, her bonuses, her supervisors' written praise, a promotion already granted, and a manager on the hiring team who had worked alongside her. The information the assessment was built to approximate was already in the building, and, as alleged, a three-hour standardised gate is what the decision ran on.
That framing changes what the accommodation failure means. A denied accommodation on a hiring screen is a barrier to being assessed. A denied accommodation on an internal promotion screen is a barrier to being assessed by an instrument that was, in this instance, redundant. And the shape of the denial is unusual enough to be worth naming precisely: this is not an employer without an accommodation process. There was a process, the employee knew it well enough to start it herself, and it returned a disposition. The disposition pointed at a product feature. Nobody, as alleged, checked that the feature was there. The failure lives on the link between deciding an accommodation and delivering one, which is a link most governance frameworks do not draw at all, because they assume that a decision to provide a control is the same event as the control existing.
Then notice how the warning travelled. The person best placed in the entire organization to see this problem coming was a Deaf employee who sat on the employer's own accessibility advisory body. She raised it, specifically, about this step, with the body's chair, a season in advance. She was told it would be looked into. The organization then required the unchanged step of her. That is a complete governance loop in which every component functioned as designed — a body existed, a member raised a concern, the chair received it — and the loop still produced nothing. A structure that hears and does not act is harder to fix than one that never hears, because from the inside it looks like working.
The observability problem is not a detail of this case; it is a party to it. What separates "an AI scored her badly" from "people watched her video and chose someone else" is, on the public record, the return address and register of one email. HireVue's denial occupies exactly that gap and cannot be answered from outside it. D.K. did the one thing the law let her do about it — she requested her personnel file — and the charge alleges she got very little. So the two pleaded layers stand on different ground: the accommodation-denial claim survives whatever the truth of the mechanism is, while the scoring claim depends on it. Any honest reading has to carry both, and to notice that the party with the evidence is the party whose denial the evidence would test.
Finally, the same substrate appears twice, four years apart, in two systems owned by two different companies. Automatic speech recognition transcribed her customer calls in 2020 for a script-adherence metric, and — on the charge's account — transcribed her interview answers in 2024. In 2020 the employer's response to a metric that misread her speech was to move her away from the metric. Nothing carried that lesson forward. That is not a technology failure and not a policy failure; it is a missing connection between two things the same organization already knew, which is the most ordinary way an institution fails to learn.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.