Hiring by Vibes: The Weak Science Behind AI Personality Screening
You sit before your laptop, alone in your bedroom, staring at a webcam. The screen displays a question about how you would handle a difficult customer. You have two minutes to respond. As you speak, an algorithm analyses every micro-expression on your face, the cadence of your voice, the words you choose, and even the pauses between your sentences. Somewhere in a data centre, machine learning models are calculating your “employability score.” They are determining whether your vibes are right.
Welcome to the new frontier of recruitment, where “good vibes” have been quantified, algorithmatised, and sold back to employers as objective science.
An estimated 99 per cent of Fortune 500 companies now use some form of automation in their hiring process. The World Economic Forum reported in March 2025 that roughly 88 per cent of companies use AI for initial candidate screening. And an October 2024 survey of business leaders found that approximately seven in ten companies allow AI tools to reject candidates without any human oversight. The promise is efficiency, objectivity, and the elimination of human bias. The reality is proving far more complicated.
The Machinery Behind the Vibe Check
The technical architecture of AI-powered hiring tools reads like science fiction adapted for the human resources department. These systems deploy a constellation of technologies: facial expression analysis, voice tone detection, linguistic pattern recognition, and neuroscience-based games, all working in concert to produce a single output, a score that purports to capture something ineffable about a candidate's personality and potential.
HireVue, one of the most prominent vendors in this space, built its reputation on video interview technology that went far beyond simply recording candidates' responses. According to the Washington Post, the company's system used candidates' computer or cellphone cameras to analyse their facial movements, word choice, and speaking voice before ranking them against other applicants based on an automatically generated “employability” score. The company had been used by over 100 employers to evaluate more than a million job candidates.
The granularity of this analysis was remarkable. HireVue examined what they called “Facial Action Units,” which could make up 29 per cent of a person's interview score. According to HireVue's own documentation, 10 to 30 per cent of a candidate's score was based on facial expressions, with the remainder calculated from language use. The AI claimed to determine everything from how excited someone seemed about work tasks to how they might behave around angry customers, all derived from facial movements and voice patterns.
Pymetrics, another major player in the space (now acquired by Harver), took a different approach, using neuroscience-based games to assess candidates. Founded by Harvard and MIT-trained neuroscientist Frida Polli, the company developed 12 neuroscience mini-games that take less than half an hour to measure 90 cognitive, social, and emotional traits. “The whole idea behind Pymetrics is that instead of using a resume, you are looking at people's cognitive, social, and emotional aptitudes,” Polli has stated. “It's also much more future-facing and potential-oriented, rather than backwards-facing and only talking about your past experiences.”
The appeal to employers is obvious. Traditional hiring is expensive, time-consuming, and riddled with inconsistency. A single corporate job posting can attract hundreds or thousands of applications. Human recruiters are limited by time, attention, and their own unconscious biases. AI systems promise to process vast numbers of candidates quickly while applying consistent criteria. When Unilever adopted an AI-powered recruitment funnel partnering with Pymetrics for its Future Leaders programme, the company processed 250,000 applicants in months. Time-to-hire was reportedly reduced by 90 per cent, from four months to four weeks, saving the company over one million pounds in recruitment costs.
The Scientific Foundation Begins to Crack
Yet the scientific claims underlying these systems have come under withering scrutiny. The most damaging revelation came from HireVue's own research, which showed that facial analysis contributed only 0.25 per cent to actual job performance prediction. Candidates were being scored heavily on factors that had virtually no correlation with their ability to do the job.
AI researchers have been considerably less diplomatic. Some have described the technology as “digital snake oil,” an unfounded blend of superficial measurements and arbitrary number-crunching unrooted in scientific fact. They argued that analysing a human being this way could penalise non-native speakers, visibly nervous interviewees, or anyone else who did not fit the model for look and speech.
Sandra Wachter, Professor of Data Ethics at the Oxford Internet Institute, has been particularly scathing. She has stated that emotion AI has “at its best no proven basis in science and at its worst is absolute pseudoscience.” Her assessment reflects a growing consensus among researchers that the technology reproduces historical forms of pseudoscience based on the concept of quantifiable and unequally distributed emotional capacity.
Lisa Feldman Barrett, Professor of Psychology at Northeastern University and a leading expert on emotion, has highlighted the fundamental problem with these systems. “The topic of facial expressions of emotion, whether they're universal, whether you can look at someone's face and read emotion in their face, is a topic of great contention that scientists have been debating for at least 100 years,” she has observed. The science is nowhere near settled enough to stake someone's livelihood on it.
The criticism extends beyond facial analysis to emotion recognition more broadly. Research published in 2024 found that emotion recognition technologies discriminate on the basis of race, gender, and disability. In one study by Lauren Rhue, emotion AI consistently interpreted Black subjects as having more negative emotions than white subjects, even when facial expressions were identical. Another study found that an emotion recognition system read Black faces as angrier than white faces, even when both were smiling to the same degree.
These are not theoretical concerns. They have real consequences for real people. In a US study published in 2024, workers expressed concern that emotion recognition systems would harm their wellbeing and impact work performance. They were fearful that inaccuracies could create false impressions about them.
Training Data and the Reproduction of Bias
The deeper problem with AI hiring tools lies not in their algorithms per se, but in what those algorithms learn from. Machine learning systems are only as good as the data they are trained on. And when that data reflects decades of hiring decisions made by humans with their own biases, the algorithm does not eliminate bias; it codifies it.
The most notorious example remains Amazon's experimental recruiting engine, first reported by Reuters in 2018. The company had been building computer programmes since 2014 to review job applicants' resumes with the aim of mechanising the search for top talent. The tool used artificial intelligence to give job candidates scores ranging from one to five stars.
The problem was the training data. The AI tool was trained on ten years' worth of resumes the company had received. Because the technology sector is male-dominated, the majority of those resumes came from men. The result was that the system was unintentionally trained to prefer male candidates over female candidates.
The discrimination was not subtle. The system reportedly penalised resumes containing the word “women's” or the names of certain all-women's colleges. Meanwhile, it favoured words such as “executed” and “captured,” which are apparently deployed more often in the resumes of male engineers. Amazon edited the programmes to make them neutral to these particular terms, but that was no guarantee the machines would not devise other ways of sorting candidates that could prove discriminatory. The company ultimately disbanded the team because executives lost hope for the project.
The Amazon case illustrates a fundamental tension in AI hiring. These systems are typically trained on data from “top performers” at a company. If a company's existing workforce lacks diversity, particularly at senior levels, then the algorithm will learn to select candidates who resemble that homogeneous group. As Meredith Whittaker, co-founder of the AI Now Institute, has observed, “Firms that are using such software may not have diverse workforces to begin with, and often have decreasing diversity at the top.”
Whittaker has documented how this pattern repeats across the AI industry. “There are an increasing pile of examples where we see that these systems embed biased and discriminatory logics,” she has stated. “In almost every case, these biases are effectively replicating histories of discrimination, so against women, against Black people, against trans people.”
The AI Now Institute's 2019 report “Discriminating Systems” documented how workforce discrimination in AI labs, dominated by a narrow demographic, causally propagates biases into deployed systems. The problem is not that AI is inherently biased. The problem is that AI amplifies and automates the biases embedded in historical data and in the teams that build these systems.
Cultural Fit Becomes Algorithmic Conformity
Perhaps nowhere is the potential for algorithmic discrimination more acute than in the assessment of “cultural fit.” This concept, long a staple of hiring discourse, has always carried discriminatory risks. AI systems that claim to measure cultural fit take those risks and amplify them.
Katherine Klein, Professor of Management at Wharton, has characterised cultural fit as “an incredibly vague term, and it's a vague term often based on gut instinct.” According to diversity researchers, the vague use of “fit” is one of the top contributors to homogenous hiring.
The problem is that hiring managers often conflate cultural fit with personal similarity. Instead of looking for people who share the company's values, they look for people who share their own background and interests. As one expert noted, “What you're going to get is a copy of your existing employees,” and “in many instances, it is a form of discrimination.”
In a 2012 paper for the American Sociological Review, Lauren A. Rivera investigated this dynamic through 120 interviews with professionals involved in hiring at US investment banks, law firms, and management consulting firms. She found that the most common thing employers looked for at the job interview stage was “similarity” in hobbies, experiences, and self-presentation styles. When interviewers said they “clicked” or “had chemistry” with a candidate, they often meant that they shared a similar background.
When AI systems are trained to identify “cultural fit,” they risk encoding these same preferences in algorithmic form. The system learns from historical data about who was hired and promoted. If past hiring reflected homogeneity, the algorithm will perpetuate that homogeneity, now with the veneer of scientific objectivity. “Good vibes” becomes a proxy for candidates who mirror the demographics and mannerisms of those already in power.
Professor Wachter's research at Oxford has explored how AI creates what she calls “artificial immutability,” using features like opacity, vagueness, instability, involuntariness, and invisibility to make discriminatory groupings seem natural and inevitable. Candidates in online interviews may be assessed by facial recognition software that tracks facial expressions, eye movement, respiration, or sweat. The criteria for evaluation are hidden from candidates, who have no way to know why they were rejected or how to improve.
The Disability and Neurodiversity Gap
The limitations of AI hiring tools become especially stark when applied to candidates with disabilities or neurodiverse traits. These systems are typically trained on data from neurotypical, able-bodied candidates. Anyone who communicates differently, whether due to deafness, autism, or other conditions, is at immediate disadvantage.
This issue came into sharp focus in March 2025, when the American Civil Liberties Union, Public Justice, and ACLU of Colorado filed a complaint with the Colorado Civil Rights Division and the Equal Employment Opportunity Commission against Intuit and HireVue. The complaint was filed on behalf of an Indigenous and Deaf woman who was denied a promotion allegedly due to discrimination based on her disability and race.
The complainant, identified as D.K., communicates using American Sign Language and English with a deaf accent. Despite receiving positive supervisor feedback and bonuses every year since 2019, she was rejected for a seasonal manager position after completing a HireVue video interview. According to the complaint, the HireVue platform did not provide consistent subtitles for all audio content.
When D.K. requested human-generated captioning, Intuit allegedly denied this accommodation, telling her that HireVue's software included subtitling capabilities. When she began the interview, no subtitling option was available, forcing her to rely on Google Chrome's automated captioning, which is sometimes incomplete and inaccurate. She was rejected for the promotion and allegedly received AI-generated feedback recommending that she “practice active listening,” a suggestion that the ACLU argues demonstrates how her hearing disability disadvantaged her in the process.
Research has consistently shown that such AI systems perform worse when evaluating non-white and deaf or hard of hearing speakers. The complaint alleges that HireVue's hiring assessment platform discriminated against deaf and non-white individuals, violating the Colorado Anti-Discrimination Act, the Americans with Disabilities Act, and Title VII of the Civil Rights Act.
HireVue CEO Jeremy Friedman has stated that the complaint “is entirely without merit” and that Intuit did not use a HireVue AI-based assessment in this instance. Intuit has similarly stated that “the allegations in the complaint are entirely without merit.” The case remains ongoing.
The legal implications extend beyond this single case. As one legal expert noted, “employers can still be held responsible for AI-related discrimination, even if the tool was developed and implemented by a third-party vendor.”
The Regulatory Landscape Takes Shape
Regulators around the world are beginning to grapple with the implications of AI in hiring. The most comprehensive framework to date is the European Union's AI Act, which entered into force on 1 August 2024. Under this regulation, AI systems used for recruitment and selection are explicitly categorised as “high-risk” due to their potential impact on individuals' access to employment and their future careers.
The EU AI Act covers AI systems intended to be used for the recruitment or selection of individuals, including to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates. Usage of AI for deciding on promotions, termination of work-related contractual relationships, allocation of tasks, and monitoring or evaluation of the performance and behaviour of workers is also classified as high-risk.
That classification is settled law. Its consequences are not yet in force. The obligations that attach to high-risk systems, covering risk management, data quality, transparency, human oversight, record-keeping, registration and incident reporting, were due to bind providers and deployers from 2 August 2026. They will not. Under the EU's Digital AI Omnibus, provisionally agreed between the Council and Parliament on 7 May 2026 and subsequently finalised, the compliance deadline for stand-alone Annex III systems, the category containing recruitment, selection, promotion and termination, was deferred to 2 December 2027. For AI embedded in regulated products under Annex I, the deadline moved to 2 August 2028. The stated rationale was that industry needed additional time to prepare for compliance and that the harmonised technical standards on which compliance depends were not yet ready. Whatever the merits of that argument, the practical effect is that the world's most comprehensive AI hiring regime will not constrain a single hiring decision until the closing weeks of 2027.
One part of the Act did not slip. The prohibition in Article 5(1)(f) on inferring emotions from biometric data in the workplace took effect on 2 February 2025 and remains in force. The European Commission's review of the Act in November 2025 specifically declined to soften the list of prohibited practices, and the omnibus delay did not touch it. The asymmetry is worth sitting with. The broad, technical, engineering-heavy obligations on high-risk systems, the risk assessments and documentation and audit trails, slipped by sixteen months. The outright ban on the single most pseudoscientific practice in the sector held. Reading emotions off a candidate's face is unlawful in the European workplace today; building a high-risk hiring algorithm without a risk management system is not, at least not yet. The most indefensible practice is the one that stayed banned, and it arguably stayed banned because it was indefensible. There was no serious industry case to be made for a technology whose scientific foundations Wachter has described as absent.
The regulation has extraterritorial reach, meaning US employers can be covered even without a physical EU presence if AI outputs are intended to be used in the EU. Penalties are tiered, and the tiering matters. Engaging in a prohibited practice, which includes workplace emotion recognition, can attract fines of up to 35 million euros or 7 per cent of global annual turnover, whichever is higher. Breaches of the obligations governing high-risk systems carry a lower ceiling of 15 million euros or 3 per cent. That the harsher of the two tiers is the one already operative is the clearest signal Brussels has sent about which practice it considers beyond redemption.
In the United States, regulation has proceeded in a more patchwork fashion. New York City's Local Law 144, which took effect in January 2023 with enforcement beginning in July 2023, requires organisations using automated employment decision tools to undergo annual bias audits from independent third-party auditors. The audit must assess the tools' disparate impact on employment decisions for candidates based on protected categories such as sex, ethnicity, and race.
The law also imposes transparency requirements. Organisations must make the date of the most recent bias audit, a summary of results, and distribution date of the tool publicly available on their websites. Employers must notify candidates at least ten business days before using such a tool.
However, enforcement has proven challenging. A recent state comptroller audit found that while the New York City Department of Consumer and Worker Protection surveyed websites and bias audits of 32 companies and identified only a single issue of non-compliance, the comptroller's office identified at least 17 instances of potential non-compliance. DCWP officials acknowledged that identifying non-compliance is difficult because if an employer does not take the required steps, it is hard to identify that they are violating the law.
Illinois has been particularly active in regulating AI in employment. The state's Artificial Intelligence Video Interview Act governs employers' use of AI analysis in video interviews, requiring various notices, consents, and data management practices. In August 2024, Governor J.B. Pritzker signed HB 3773, amending the Illinois Human Rights Act to expressly regulate AI for employment decisions. That law took effect on 1 January 2026 and now binds employers in the state. It prohibits the use of AI in ways that discriminate against employees on the basis of protected characteristics, even where the discrimination is unintentional; it bars the use of zip codes as proxies for protected classes; and it requires employers to give notice when AI is used in employment decisions.
The obligations arrived without a map, however. The Illinois Department of Human Rights temporarily withdrew its proposed rules on the use of artificial intelligence in employment, leaving the statutory duties fully in force while the precise contours of compliant notice remain undefined pending further rulemaking. Employers are bound by requirements whose detailed shape has not been settled. This is a familiar pattern in the field, where the duty to be accountable consistently outruns any specification of what accountability actually looks like.
The Illinois Biometric Information Privacy Act adds another layer of exposure. Employers using AI facial recognition in conjunction with video interviews, such as to analyse facial expressions, speech patterns, and other non-verbal cues, could face liability under both the video interview act and the biometric privacy law.
The Legal Reckoning Begins
The courtroom is becoming a new battleground over AI hiring discrimination. The EPIC complaint against HireVue filed with the Federal Trade Commission in November 2019 charged that the company falsely denied using facial recognition and failed to comply with baseline standards for AI decision-making. While HireVue subsequently removed facial analysis from its assessments, the complaint highlighted the opacity surrounding these systems.
In July 2024, CVS privately settled a proposed class action lawsuit filed by a job applicant who claimed the company broke Massachusetts law by requiring prospective employees to undergo what legally amounted to a lie detector test. The lawsuit alleged that applicants were required to take HireVue video interviews using Affectiva's AI technology to track facial expressions and assign an “employability score.”
The most consequential of these cases is Mobley v. Workday, which alleges discrimination against Black, older and disabled applicants arising from Workday's algorithmic screening practices. It has proceeded in California as a nationwide collective action under the Age Discrimination in Employment Act, covering applicants who applied for positions through Workday's platform on or after 24 September 2020 and who were aged forty or over at the time. The court authorised notice to that collective, and the window for applicants to opt in closed on 7 March 2026. A ruling handed down on 6 March 2026 rejected Workday's contention that the ADEA does not reach job applicants at all, an argument that, had it succeeded, would have removed algorithmic screening at the application stage from the statute's reach entirely.
Then came a decision that ought to trouble anyone who believes bias audits are a sufficient answer. On 29 May 2026, the court denied the plaintiffs' motion to compel production of Workday's own bias-testing data, the internal evidence of whether its tools produced disparate outcomes. The material was held to be protected by attorney-client privilege. Workday's lawyers had curated the data used in the testing, and the results had been used in the provision of legal advice rather than for a business purpose. The court also rejected the argument that Workday had waived that privilege by publicly acknowledging outside litigation that it conducted bias testing at all. The case remains in discovery, with no trial date set.
The ACLU has also filed a complaint with the FTC against Aon over AI personality tests.
Ifeoma Ajunwa, Professor of Law at the University of North Carolina and founding director of the AI Decision-Making Research Program, has documented the legal challenges these cases present. “The law requires that you prove either intent to discriminate or show a pattern of discrimination,” she has explained. “Automated hiring platforms actually make it much harder to do either of those. And a lot of times, the algorithms that are part of the hiring system are considered proprietary, meaning that they're a trade secret. So you may not actually be able to be privy to exactly how the algorithms were programmed and also to exactly what attributes were considered.”
The Workday privilege ruling is that warning made concrete, and then some. Ajunwa's concern was that trade secrecy would place the algorithm itself beyond a plaintiff's reach. What the May 2026 decision established is that evidence of the algorithm's effects can be placed beyond reach too, by the straightforward expedient of routing the testing through counsel. A company can test its systems for discrimination, learn the answer, tell the world that it tests, and still keep the findings from the very people whose claims those findings might substantiate. The audit becomes a compliance artefact rather than an accountability mechanism. This is the accountability gap in its purest form: not an absence of scrutiny, but scrutiny sealed by privilege, conducted for the benefit of the party being scrutinised.
In testimony before Congress, Ajunwa argued that government measures are urgently needed to regulate automated hiring systems that often discriminate against women, military veterans, formerly incarcerated people, people with disabilities, and others. She has argued for independent audits and compulsory data retention, contending that an employer's failure to audit its automated hiring platforms for disparate impact should serve as prima facie evidence of discriminatory intent under Title VII.
Ajunwa's paper “Automated Video Interviewing as the New Phrenology” draws an explicit parallel between these technologies and discredited pseudoscientific practices of the past. The comparison is not rhetorical flourish. It reflects a genuine concern that we are watching history repeat itself, now with the authority of silicon rather than callipers.
The Promise of Debiased AI
Despite the mounting criticism, proponents of AI hiring tools argue that the technology, properly designed, could actually reduce discrimination rather than perpetuate it. Frida Polli, who co-founded Pymetrics and went on to serve as Chief Data Science Officer at Harver following the 2022 acquisition, has been a vocal advocate for this position. She has since moved on to other work, holding a position as a Visiting Innovation Scholar at MIT's Schwarzman College of Computing and founding Alethia AI in 2023.
“We are fundamentally biased, and we can't help it,” Polli has stated. “I personally think that if we really want to change diversity in the workforce, we are never going to get there with humans.” She argues that algorithms are more trainable than humans: “It's hard to remove bias from algorithms, but it is possible. It is not possible to remove bias from humans.”
Pymetrics designed its assessments to be “completely free of gender and ethnic bias,” arguing that measuring cognitive skills like memory through actual tests is gender-blind. The company audits its algorithms with a formula made publicly available on GitHub. Polli was instrumental in passing New York City's bias audit law, the first such regulation in the nation.
The Unilever case study is often cited as evidence that AI hiring can promote diversity when properly implemented. The company reported that the use of AI in sourcing led to an increase in talent diversity by 16 per cent. It employed what it called “the most diverse ethnic and gender employee class so far.” The company attributed this success to using bias-free data sets in training AI systems and maintaining human supervision in the use of the systems.
Yet even success stories come with caveats. With smaller companies, top performers may be very similar in how they think and work. Training models to identify candidates based on how they match with top performers could potentially exclude innovative thinkers. The very diversity that makes teams effective might be screened out by systems optimised for conformity.
Research by University of Washington found significant racial, gender, and intersectional bias in how state-of-the-art large language models ranked resumes. A large-scale randomised experiment found that leading AI models systematically favour female candidates while disadvantaging Black male applicants, even when qualifications are identical. The bias is complex and intersectional, making simple fixes inadequate.
Professor Wachter's work has shown that the majority of popular bias tests and tools, 13 out of 20 examined, do not live up to the standards of EU non-discrimination law. In response, she developed a bias test called Conditional Demographic Disparity that meets EU and UK standards. Amazon and IBM have implemented it in their cloud services. In 2024, this test was used to uncover systemic bias in education in the Netherlands, leading to a formal apology from the Dutch Minister for Education.
What Candidates Face
For job seekers, the rise of AI hiring tools creates an opaque and often bewildering landscape. You may never know that an algorithm rejected you, let alone why. The criteria for success are hidden. The appeals process is non-existent.
HireVue does not give candidates access to their assessment scores or the training data, factors, logic, or techniques used to generate each algorithmic assessment. You receive a rejection, or you do not. The system has spoken.
A survey found that only 12.9 per cent of Australian adults support face-based emotion recognition technologies in the workplace. Respondents viewed facial analysis as invasive, unethical, and highly prone to error and bias.
The power asymmetry is profound. Job seekers, particularly those early in their careers or from marginalised backgrounds, have little leverage to question or challenge these systems. They must perform for the algorithm, adapting their self-presentation to what they guess the machine wants to see, without knowing what that actually is.
Some candidates are fighting back. Research published in 2024 documented how smart candidates are learning to detect AI interview bias and protect themselves from algorithmic discrimination. But this places an additional burden on job seekers, requiring them to become experts in a technology that should be serving them fairly.
Rethinking What We Measure
The fundamental question underlying this entire debate is what we are actually trying to measure when we hire someone. Traditional interviews, for all their flaws, at least have face validity. You talk to a person. You assess whether you can work with them. The criteria may be subjective, but they are human.
AI hiring tools promise to replace this subjectivity with objectivity. But objectivity requires valid measures. And the measures these systems use (facial expressions, vocal tone, word choice, neuroscience game performance) have not been validated as predictors of job performance. HireVue's own research showed facial analysis contributed only 0.25 per cent to job performance prediction. What, exactly, are we measuring?
The answer may be nothing more than conformity. Systems trained on historical data learn to identify candidates who look, sound, and behave like people who were hired in the past. In a homogeneous workplace, that means selecting for homogeneity. In a discriminatory system, that means perpetuating discrimination.
Joy Buolamwini, founder of the Algorithmic Justice League and co-author of the landmark Gender Shades study, has documented how commercial AI systems fail to recognise people equally. Her research with Timnit Gebru found that the error rate for light-skinned men was 0.8 per cent, compared to 34.7 per cent for darker-skinned women. IBM ended its facial recognition programme partly in response to this research.
Buolamwini's experience was personal before it was academic. While working on a facial-recognition-based art project at the MIT Media Lab, she discovered that commercial AI systems could not consistently detect her face due to her darker skin. “A white mask was a closer fit to what the system had learned was a face than my actual human face,” she has observed. Fortune magazine named her “the conscience of the AI revolution.”
Gebru, who co-founded Black in AI and later the Distributed Artificial Intelligence Research Institute, has spent years documenting how AI systems replicate discrimination. Named one of Time's most influential people of 2022, she has argued that facial recognition is too dangerous to be used for law enforcement and security purposes at present. The same concerns apply to employment.
The Regulatory Winds Are Shifting
It is tempting to read the past two years as a one-way ratchet, with the era of unaccountable AI hiring drawing steadily to a close. The evidence no longer supports that reading. The trajectory is contested rather than inevitable, and during 2026 it moved in both directions at once.
Consider Colorado. SB 24-205, signed in 2024, was the most ambitious state AI statute in the United States: a duty of care imposed on developers and deployers of high-risk systems to protect consumers from algorithmic discrimination, with employment squarely within scope. It never took effect. In roughly six weeks in the spring of 2026, the entire edifice came down. A federal court stayed its enforcement. The US Department of Justice joined a constitutional challenge to it. The state attorney general announced he would not enforce it pending rulemaking. Then the legislature repealed it outright, replacing it on 14 May 2026 with SB 26-189, a markedly narrower framework governing automated decision-making technology, due to take effect on 1 January 2027. The replacement abandons the original's duty of care around algorithmic discrimination altogether. In its place sit notice requirements, a right to an explanation of an adverse outcome within thirty days, and human review where commercially reasonable. The attorney general has indicated he will not enforce that statute either until rulemaking concludes.
Set that alongside the European Union's sixteen-month deferral of its high-risk obligations, and alongside the New York City enforcement record, where a state comptroller's audit identified seventeen instances of potential non-compliance that the enforcing agency had not caught, and a rather different picture emerges. Binding constraints on AI hiring are arriving more slowly and more thinly than the legislative wave of 2024 appeared to promise, even as the underlying technology proliferates and the vendors multiply.
This is not a counsel of despair. Real constraints did land, and they have held. The EU's prohibition on workplace emotion recognition is in force, carries the Act's heaviest penalties, and survived both a Commission review and an omnibus delay that touched almost everything else around it. Illinois HB 3773 is now law and binds employers in that state today. New York City's disclosure regime, however weakly policed, has at least made bias audits a visible expectation rather than a private courtesy. The direction of travel still points towards accountability. What has changed is that nobody can now say with confidence how fast it is moving, or how far it will go.
HireVue itself has evolved. The company removed facial analysis from its assessments in 2021, acknowledging that the technology “wasn't worth the concern.” The company now states that its video assessments only evaluate language, specifically how candidates talk about their experiences and past actions pertaining to competencies critical to the particular role. But voice analysis and linguistic pattern detection carry their own bias risks.
The market is adapting. Some companies are shifting from “culture fit” to “culture add,” seeking candidates who bring different perspectives rather than candidates who conform to existing norms. Research shows structured interviews are twice as predictive of job performance compared to unstructured ones, and they reduce the scope for bias. There are better ways to hire than asking an algorithm to judge someone's vibes.
But the technology continues to proliferate. New AI tools emerge constantly, promising ever more sophisticated assessments of personality, potential, and fit. The vendors may change. The fundamental problems remain.
A Mirror That Distorts
In the end, AI hiring tools hold up a mirror to the organisations that use them. They reflect the biases embedded in historical data. They reproduce the homogeneity of existing workforces. They encode the preferences of the people who build and train them. They are not neutral arbiters of talent. They are amplifiers of existing power structures.
The promise of eliminating human bias through technology is seductive. It offers absolution from the uncomfortable work of confronting discrimination directly. But there is no algorithmic shortcut to equity. Machines learn what humans teach them. If we teach them our biases, they will apply those biases at scale.
The question is not whether AI should play a role in hiring. Automation has benefits: efficiency, consistency, and the capacity to process far more candidates than any human recruiter. The question is whether we will demand transparency, accountability, and scientific validity from these systems, or whether we will allow “good vibes” to become another way of saying “people like us.”
Meredith Whittaker has called for algorithm audits that include experts, advocacy groups, and academics reviewing them and studying their effects on different populations. “We think that's not happening today,” she has stated, “and it could lead to serious problems as AI takes off.”
Ifeoma Ajunwa has argued that audits should be mandated and that failure to audit should constitute evidence of discriminatory intent. Sandra Wachter has developed bias tests that meet legal standards. Joy Buolamwini and Timnit Gebru have shown that these systems fail the people they are supposed to serve fairly.
The tools exist. The research is clear. The regulations are uneven: deferred in Brussels, repealed in Denver, thinly enforced in New York, and binding in Springfield. What remains to be seen is whether employers will choose accountability over convenience in the absence of anyone reliably compelling them to, and whether job seekers will be protected from algorithms that claim to see into their souls but see only reflections of the status quo.
Your next job interview may be judged by a machine. The question is what that machine has learned about who deserves a chance.
References and Sources
Dastin, J., “Amazon scraps secret AI recruiting tool that showed bias against women”, Reuters, October 2018. https://www.reuters.com/article/us-amazon-com-jobs-automation-insight-idUSKCN1MK08G
Electronic Privacy Information Center, “In re HireVue: Complaint and Request for Investigation, Injunction, and Other Relief”, submitted to the Federal Trade Commission, November 2019. https://epic.org/documents/in-re-hirevue/
Harwell, D., “A face-scanning algorithm increasingly decides whether you deserve the job”, The Washington Post, 22 October 2019. https://www.washingtonpost.com/technology/2019/10/22/ai-hiring-face-scanning-algorithm-increasingly-decides-whether-you-deserve-job/
Buolamwini, J. and Gebru, T., “Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification”, Proceedings of Machine Learning Research, volume 81, 2018. https://proceedings.mlr.press/v81/buolamwini18a.html
University of Washington News, “AI tools show biases in ranking job applicants' names according to perceived race and gender”, University of Washington, 31 October 2024. https://www.washington.edu/news/2024/10/31/ai-bias-resume-screening-race-gender/
American Civil Liberties Union, “Complaint of Discrimination: D.K. v. Intuit Inc. and HireVue Inc.” (redacted), filed with the Colorado Civil Rights Division and the Equal Employment Opportunity Commission, March 2025. https://assets.aclu.org/live/uploads/2025/03/Redacted-HireVue_Intuit-Complaint-of-Discrimination_Redacted.pdf
Ajunwa, I., “Automated Video Interviewing as the New Phrenology”, Berkeley Technology Law Journal, volume 36, number 3, page 1173, 2022. https://btlj.org/wp-content/uploads/2023/01/0008-36-3-Ajunwa_Web.pdf
Ajunwa, I., “An Auditing Imperative for Automated Hiring Systems”, Harvard Journal of Law & Technology, volume 34, number 2, 2021. https://jolt.law.harvard.edu/assets/articlePDFs/v34/5.-Ajunwa-An-Auditing-Imperative-for-Automated-Hiring-Systems.pdf
US Equal Employment Opportunity Commission, “Testimony of Dr. Ifeoma Ajunwa”, public meeting on navigating employment discrimination in AI and automated systems, 31 January 2023. https://www.eeoc.gov/meetings/meeting-january-31-2023-navigating-employment-discrimination-ai-and-automated-systems-new/ajunwa
Oxford Internet Institute, “Current discrimination laws failing to protect people from AI-generated unfair outcomes”, University of Oxford, 2022. https://www.oii.ox.ac.uk/news-events/ai-creates-unintuitive-and-unconventional-groups-to-make-life-changing-decisions-yet-current-laws-do-not-protect-group-members-from-ai-generated-unfair-outcomes-says-new-paper/
West, S.M., Whittaker, M. and Crawford, K., “Discriminating Systems: Gender, Race and Power in AI”, AI Now Institute, April 2019. https://ainowinstitute.org/publications/discriminating-systems-gender-race-and-power-in-ai-2
Harvard Business Review, “AI, Accountability, and Power with Meredith Whittaker”, Exponential View podcast, November 2023. https://hbr.org/podcast/2023/11/azeems-picks-ai-accountability-and-power-with-meredith-whittaker
New York City Department of Consumer and Worker Protection, “Automated Employment Decision Tools (AEDT)”, City of New York. https://www.nyc.gov/site/dca/about/automated-employment-decision-tools.page
Office of the New York State Comptroller, “Enforcement of Local Law 144, Automated Employment Decision Tools”, 2 December 2025. https://www.osc.ny.gov/state-agencies/audits/2025/12/02/enforcement-local-law-144-automated-employment-decision-tools
Fowler, M.A., “Expanding Use of Artificial Intelligence Into Employment and Labor Practice: Legislative Response and Legal Implications”, Chuhak & Tecson, published by the National Law Review, 21 November 2025. https://natlawreview.com/article/expanding-use-artificial-intelligence-employment-and-labor-practice-legislative
Seyfarth Shaw, “Illinois Department of Human Rights Temporarily Withdraws Proposed Rules on Use of Artificial Intelligence in Employment”. https://www.seyfarth.com/news-insights/illinois-department-of-human-rights-temporarily-withdraws-proposed-rules-on-use-of-artificial-intelligence-in-employment.html
European Commission, “Regulatory framework for AI”, Shaping Europe's Digital Future. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
Pinsent Masons, “Law delaying EU's 'high-risk' AI rules finalised”, Out-Law, 6 July 2026. https://www.pinsentmasons.com/out-law/news/law-delaying-eu-high-risk-ai-rules-finalised
Finnegan, “Colorado Replaces Landmark AI Act: An Overview of the New SB 26-189 Framework”, 2026. https://www.finnegan.com/en/insights/articles/colorado-replaces-landmark-ai-act-an-overview-of-the-new-sb-26-189-framework.html
Duane Morris, “California Federal Court Clarifies Limits On AI Bias Testing And Applicant Data Disclosure In Mobley v. Workday”, Class Action Defense Blog, 2 June 2026. https://blogs.duanemorris.com/classactiondefense/2026/06/02/california-federal-court-clarifies-limits-on-ai-bias-testing-and-applicant-data-disclosure-in-mobley-v-workday/
Quartz, “People are terrible judges of talent. Can algorithms do better?”, 2019. https://qz.com/work/1742847/pymetrics-ceo-frida-polli-on-the-ai-solution-to-hiring-bias
Rivera, L.A., “Hiring as Cultural Matching: The Case of Elite Professional Service Firms”, American Sociological Review, volume 77, number 6, pages 999-1022, December 2012.
Knowledge at Wharton, “Is Cultural Fit a Qualification for Hiring or a Disguise for Bias?”, University of Pennsylvania. https://knowledge.wharton.upenn.edu/article/cultural-fit-a-qualification-for-hiring-or-a-disguise-for-bias/
HR Dive, “AI hiring software was biased against deaf employees, ACLU alleges in ADA case”, 2025. https://www.hrdive.com/news/ai-intuit-hirevue-deaf-indigenous-employee-discrimination-aclu/743273/
SHRM, “HireVue Discontinues Facial Analysis Screening”. https://www.shrm.org/topics-tools/news/talent-acquisition/hirevue-discontinues-facial-analysis-screening
Tim Green UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk
Listen to the free weekly SmarterArticles Podcast