I thought it would be helpful to provide an overview on AI and tech in recruitment, with the areas that directly affect you as an applicant.

I'm a bit short on time this week, due to an urgent work requirement, so I've used Claude to build on previous posts, research current stories and link it together.

In a sense this is a useful way to do it, because many such articles are written by AI, and many are flawed by hallucinations, flattening nuance, and overstating claims.

So you should challenge this in the same way you apply critical thinking to any 'thought leadership'.

That said, it covers the points I intended, and you can read previous articles in this series for my own researched takes - Some Truths about the ATS and AI; De Facto Automation and You are a good starting point, including explanations for what you may be experiencing, and how assumptions may hold you back.

I've also cross checked against ChatGPT, Gemini and Perplexity, which flagged minor overstatements around the Pymetrics study and technical omissions around Eightfold. But then those same inaccuracies are why professional talent acquisition functions are building safeguards in how AI is used.

Don't forget the Summer update to A Career Breakdown Kit is out now. Buying a copy supports my work, complements the frees tools and resources available on my website, and will improve your odds.

AI in Recruitment: What It Actually Does, What It Doesn't, and What Jobseekers Need to Know

1. How AI Is Currently Used in Recruitment

The phrase "AI in recruitment" has become so broad as to be almost meaningless. It covers everything from a spellchecker in an email template to a neural network scoring your personality through browser games. To make sense of it, it helps to separate AI from the technology it sits alongside, and both from the marketing language wrapped around them.

Workflow automation. The most widespread and least controversial use of AI in recruitment is the automation of administrative tasks. Scheduling interviews, sending confirmation emails, parsing CVs into structured data fields, generating interview summaries from transcripts — these are genuine productivity tools. They reduce the time a recruiter spends on process and increase the time available for judgement. Automated transcription of interviews, powered by speech-to-text models, falls into this category. It is mundane, useful, and rarely the thing anyone means when they say "AI is screening you out."

AI-assisted sourcing. Some recruitment platforms use machine learning to suggest candidates from databases or professional networks. LinkedIn Recruiter, for example, uses algorithms to surface profiles that match a search brief based on historical engagement patterns and role data. The sourcing layer is probabilistic — it suggests who a recruiter might want to look at, but a human still decides whether to make contact. This is AI as a recommendation engine, closer to how a streaming service suggests a film than how a judge passes sentence.

Generative AI in candidate communications. A growing number of recruiters and in-house teams use large language models (ChatGPT, Claude, Gemini) to draft job advertisements, outreach messages, and rejection emails. This is widespread, lightly regulated, and — depending on the competence of the person prompting the model — ranges from genuinely helpful to transparently generic. Candidates are increasingly doing the same thing in reverse: using AI to tailor CVs and write cover letters. The result is a kind of AI-on-AI performance, where neither side is entirely sure what was written by a person.

2. The Technology AI Sits Alongside

It is important to distinguish AI from the broader recruitment technology stack, because the two are routinely conflated — often deliberately.

The Applicant Tracking System (ATS). An ATS is a database. It stores applications, organises them by vacancy, tracks candidate status, and allows recruiters to search, filter, and communicate. The core function of an ATS — receiving and storing applications — is not AI. It is data management. Workable, Greenhouse, Bullhorn, Lever, and dozens of others all perform this function. The ATS does not, by default, make decisions. It stores decisions made by people and, in some configurations, applies filters set by people.

Automated communications. Most ATS platforms allow automated email sequences: acknowledgement of application, invitation to interview, rejection notification. These are triggered by a recruiter changing a candidate's status, not by an algorithm deciding the candidate's fate. The automation is in the sending, not the judging.

Ranking, filtering, and knockout questions. Here the picture becomes more nuanced. Many ATS platforms allow recruiters to set mandatory criteria — often called knockout questions. "Do you have the right to work in the UK?" "Do you hold a valid HGV licence?" If a candidate answers "no" to a mandatory question, the system may automatically reject or deprioritise them. This is filtering by rule, not filtering by algorithm. The rule was written by a human. Whether that rule is sensible, lawful, or proportionate is a question about the person who configured it, not the software that executed it.

Some ATS platforms also offer keyword matching or scoring, where applications are ranked by how closely the CV text matches the job description. This is closer to AI in the technical sense, but it is crude pattern matching, not deep learning. It rewards candidates who mirror the language of the advert, which creates a perverse incentive to keyword-stuff — a behaviour that helps no one and which much of the career advice industry actively encourages.

3. Niche Applications: Pymetrics, Eightfold, and the Assessment Layer

Beyond the ATS and its associated tooling sits a smaller but more consequential category: third-party AI assessment vendors. These are the tools that genuinely attempt to make or influence hiring decisions using algorithmic models, and they are the ones generating the most significant legal and ethical scrutiny.

Pymetrics (now part of Harver). Pymetrics asks candidates to play a series of 12–16 online games designed to measure cognitive and behavioural traits such as risk tolerance, processing speed, and altruism. A machine learning model, trained on the gameplay data of a company's existing employees, then scores each candidate as "recommend" or "do not recommend." The model does not assess CVs. It does not read cover letters. It plays games with you and decides whether you resemble the people already doing the job.

The training methodology deserves scrutiny. The "good" benchmark is whoever currently holds the role — not top performers, not people independently assessed as effective, just incumbents. The "bad" benchmark is not people who failed in the role or were dismissed; it is random profiles. The model therefore learns one thing: do you look like the existing workforce, or do you look like a stranger? This is similarity matching dressed as science, and it carries an obvious risk of reproducing whatever demographic composition already exists within the client organisation. Your gameplay score is also cached for 330 days, meaning a single session can follow you across multiple applications to different employers.

Eightfold AI. Eightfold operates a "Talent Intelligence Platform" that claims to maintain data on over one billion workers globally. It scores job applicants on a scale of zero to five based on their perceived likelihood of success, drawing on scraped and aggregated personal data. Unlike pymetrics, Eightfold's legal exposure centres not primarily on bias but on transparency — or rather, the absence of it. A January 2026 class action (Kistler et al. v. Eightfold AI, Inc.) alleges that Eightfold operated as an unregistered consumer reporting agency, compiling and using personal data to score candidates without the disclosures, consent mechanisms, or dispute rights required by the Fair Credit Reporting Act. The lawsuit was brought by former EEOC Chair Jenny R. Yang and the nonprofit Towards Justice.

HireVue. HireVue provides automated video interview technology used by over 60% of the Fortune 100. Its platform has used automated speech recognition to generate transcripts and, historically, facial expression analysis to score candidates. In March 2025, the ACLU filed charges with the EEOC and the Colorado Civil Rights Division on behalf of a deaf, Indigenous woman who alleged that HireVue's platform discriminated against her when she applied for a promotion at Intuit. She requested real-time captioning accommodation, was allegedly denied, and subsequently received feedback to work on "effective communication." HireVue dropped its facial analysis feature in 2021 following sustained criticism, but the speech recognition and transcript-based assessment remain. A separate 2024 lawsuit alleged that CVS required applicants to take HireVue video assessments that amounted to an unlawful lie detector test under Massachusetts law; CVS settled privately.

Aon Consulting. In May 2024, the ACLU filed a complaint with the FTC challenging three of Aon's AI hiring tools — ADEPT-15, vidAssess-AI, and gridChallenge — as discriminatory against people with disabilities and certain racial groups. The ACLU also challenged Aon's marketing claim that its tools were "bias-free."

4. Marketing Versus Usage: The AI Bandwagon

There is a considerable gap between what AI vendors claim their products do, what those products actually do, and what employers understand they have bought. This gap is not accidental. It is commercially useful.

The recruitment technology market is crowded, and "AI-powered" has become the default differentiator. A platform that sends automated emails now markets itself as using AI. A Boolean search with a drag-and-drop interface becomes "intelligent talent matching." A chatbot that asks three screening questions becomes "conversational AI." The inflation is so routine that the label has lost most of its informational content. When everything is AI, nothing is.

For jobseekers, this creates a specific problem: it is nearly impossible to know, from the outside, what technology is actually being applied to your application. You might be told a company "uses AI in hiring" and reasonably conclude that a sophisticated algorithm is scoring your CV. In practice, the "AI" might be an automated acknowledgement email. Or it might be a pymetrics game that decides your fate before a human sees your name. The marketing makes no distinction, and the companies buying these tools often do not fully understand the distinction either.

The World Economic Forum's widely cited claim that "over 90% of employers use AI in hiring" illustrates the problem. The underlying data, drawn from an Accenture survey, found that 90% of employers use recruitment software — an ATS or recruitment marketing system — to filter or rank candidates. That is not the same as saying 90% use artificial intelligence in any meaningful sense. An ATS is a database with search functionality. Calling it AI is like calling a filing cabinet a robot because it has a label on the front.

This matters because the gap between marketing and reality feeds directly into jobseeker anxiety. If you believe sophisticated AI is screening you out, you will spend time and money on "ATS-beating" strategies, keyword-stuffing services, and resume optimisation tools — many of which are sold by the same ecosystem that created the anxiety in the first place. The commercial incentive to overstate what AI does in recruitment is enormous, and it runs in both directions: vendors overstate capability to sell to employers, and career coaches overstate threat to sell to jobseekers.

5. The Origins of ATS Mythology

Few statistics in recruitment have done more damage than the claim that "75% of CVs are rejected by ATS before a human ever sees them." It is repeated on LinkedIn, in career coaching courses, on TikTok, in national media, and in the marketing copy of resume optimisation services. It has shaped how millions of jobseekers approach applications. And it is, as far as anyone can determine, entirely made up.

The Preptel origin. The 75% figure has been traced, through multiple independent investigations, to a company called Preptel — a resume optimisation service that used the statistic in its sales pitch around 2012. Preptel went out of business in August 2013. No research methodology, no sample size, no survey, and no peer-reviewed source were ever published to support the claim. Career consultant Christine Assaf documented this origin in her investigation and concluded that the statistic was created without any study, survey, or context. The figure drifts between 70%, 75%, and 88% across different sources precisely because there was never an original value to anchor it. A real statistic has one number and one method. This one has three numbers and no method.

The citation chain runs from Preptel's marketing through an uncited Forbes mention, then through CNBC, Yahoo News, CIO, and an expanding network of career advice blogs — each citing the one before, none citing a study. It is a textbook example of circular citation: a marketing claim repeated often enough that it acquired the appearance of fact.

The Hidden Workers report. The claim is frequently bolstered by reference to the Harvard Business School and Accenture report Hidden Workers: Untapped Talent (2021). This is a serious piece of research, surveying over 8,000 workers and 2,250 executives across the US, UK, and Germany. It does find that 88% of employers agreed that qualified, higher-skilled candidates were being screened out. But it attributes this to overly rigid hiring criteria set by people — not to autonomous algorithmic rejection. The report found that employers configured their systems with excessively narrow requirements (specific job titles, precise years of experience, zero tolerance for employment gaps) and that these human-defined rules excluded capable candidates. The software executed the filter. A person wrote it.

The critical distinction, which the career advice industry almost universally ignores, is between the ATS as a tool that stores and organises applications and the ATS as an autonomous gatekeeper that rejects candidates without human involvement. The Hidden Workers report describes the former being badly configured, not the latter making independent decisions. Knockout questions — "Do you have the right to work in the UK?", "Do you hold a CSCS card?" — are the most common form of automated filtering, and they are binary eligibility checks set by recruiters, not algorithmic judgements made by machines.

An Enhancv survey of 25 US recruiters found that 92% confirmed their ATS does not auto-reject on formatting or content. Only 8% configured any form of content-based auto-rejection at all. Jobscan, the largest ATS optimisation vendor, states plainly that an ATS does not reject resumes — it stores them and allows recruiters to search using keywords.

How the myth sustains itself. The 75% figure persists because it serves multiple commercial interests simultaneously. Resume optimisation services need jobseekers to believe their CVs are being machine-rejected so they will pay for "ATS-friendly" formatting. Career coaches need a dramatic hook. LinkedIn influencers need engagement. And AI vendors, paradoxically, benefit from an inflated perception of their own products' power, because it makes their offering sound more consequential than it may be. The myth is self-reinforcing because everyone in the supply chain profits from it.

6. The Stanford Study: What It Found, What It Didn't, and How It Was Misrepresented

In May 2026, a Stanford-led research team published Algorithmic Monocultures in Hiring at the ACM Conference on Fairness, Accountability, and Transparency (FAccT). The paper examined 4.2 million job applications submitted through pymetrics between 2018 and 2022, across 156 employers and 1,746 positions. It generated immediate and extensive media coverage, much of which treated it as a verdict on AI hiring as a whole. It is not.

What the study found. Pymetrics had previously audited its own fairness by pooling all applicants together across all employers and positions. Pooled like that, it passed the standard adverse-impact screen: Black applicants were recommended at 52.5%, white applicants at 58.3%. The Stanford researchers correctly observed that this is the wrong way to audit a selection tool. US employment discrimination law evaluates adverse impact position by position, not as a company-wide average. When they split the data into individual roles, bias that the average had concealed became visible: approximately 10.6% of positions showed adverse impact against Black applicants, and roughly a quarter of Black applications landed in those affected roles. The researchers estimated that if the tool had recommended Black and Asian candidates at the same rate as the most-favoured group, approximately 40,000 additional applications would have advanced.

This is a genuine and important finding: company-level audits can mask position-level discrimination. It applies to any selection tool, not just pymetrics, and any organisation auditing hiring fairness only at the aggregate level should take note.

What the study did not find. The paper's central concept is "algorithmic monoculture" — the idea that when multiple employers use the same AI vendor, a single biased model can lock a candidate out of the entire market. The headline version of this claim — that candidates are being "rejected everywhere" by the same algorithm — barely survives contact with the study's own data. 84% of applicants in the dataset applied to exactly one position. Over 95% applied to one or two. Only 0.02% (522 people out of 3.4 million) applied to ten. When the researchers directly simulated the worst case — running 1,000 applicants against all 495 pymetrics models — not a single person was rejected by every model. The worst-off individual was still recommended for 52 roles. The monoculture mechanism exists in theory, but the data shows it affecting almost nobody in practice.

The specificity problem. The study examined one vendor: pymetrics. It did not examine any ATS. It did not examine CV screening. It did not examine keyword matching, ranking algorithms, or any of the other technologies that the phrase "AI in hiring" is routinely taken to mean. Pymetrics is a niche behavioural game platform, acquired by Harver in 2022, that trains its models on current employees as the benchmark for "good" and random profiles as the benchmark for "bad." There is no evidence in the study — and the authors acknowledge this in their limitations section — that the tool predicts job performance at all. It is a similarity matcher. It measures whether you resemble the existing workforce, not whether you can do the job.

The paper's opening cites HireVue's market dominance — used by over 60% of the Fortune 100 — to establish the scale of algorithmic hiring. But HireVue is a different product, a different company, and a different technology category, and it does no work whatsoever in the study's analysis. The citation creates an impression of ubiquity that the data does not support.

The media amplification. The researchers were careful. Their limitations section acknowledges that their findings may not generalise beyond one game-based tool, that they never measured whether pymetrics predicts performance, that they cannot determine whether rejected candidates would have been good hires, and that nothing in the paper proves illegality. The problem is not the research. It is what happened after publication: abstract became press release, press release became headline, headline became influencer thread, and each stage dropped a qualifier the authors had deliberately included. "May be a distinctive feature of centralised algorithmic assessment" became "AI is rejecting you everywhere." A study of one niche vendor became a referendum on recruitment technology as a whole.

For jobseekers, the practical takeaway is narrower than the coverage suggests. If you encounter a pymetrics-style game assessment, know that the tool has documented fairness issues and no published evidence of predictive validity. If you are applying through a standard ATS, this study tells you nothing about what is happening to your application, because it did not examine that process.

7. AI Hiring in the Courts: The Cases That Matter

The legal landscape around AI in recruitment is developing rapidly, with several significant cases establishing precedents that will shape how employers and vendors use algorithmic tools.

Mobley v. Workday, Inc. (Case No. 3:23-cv-00770, N.D. Cal.) — This is widely regarded as the bellwether case for AI hiring discrimination. Derek Mobley filed a class action in February 2023 alleging that Workday's AI-powered applicant screening system discriminated on the basis of race, age, and disability. Mobley had applied to over 100 jobs through Workday over seven years and was rejected within minutes each time. The case has survived multiple attempts by Workday to secure dismissal and has escalated significantly through 2026.

Key rulings include the court's acceptance that Workday can be considered an "agent" of employers — meaning it is not merely a neutral tool provider but is acting on employers' behalf in making hiring decisions, which triggers direct liability under anti-discrimination law. In January 2026, Workday argued that the Age Discrimination in Employment Act (ADEA) does not cover job applicants, only current employees. In March 2026, Judge Rita Lin rejected this argument, pointing to legal precedent and EEOC guidance interpreting the law as covering candidates. Preliminary collective action certification was granted in May 2025, and applicants had until March 2026 to opt in.

In May 2026, a discovery ruling denied the plaintiffs' attempt to compel production of Workday's bias-testing data (protected by attorney-client privilege) and Workday's customers' applicant data, but ordered production of Workday's EEO-1 and OFCCP documents as relevant to the company's knowledge of potential demographic disparities. Plaintiffs filed an amended complaint in March 2026 reasserting dismissed California state and disability claims. The case now presents exposure not only for Workday but for the more than 10,000 employers using its AI-powered hiring tools.

Kistler et al. v. Eightfold AI, Inc. (Contra Costa County Superior Court, filed January 2026) — This case takes a different angle from Mobley. Rather than alleging bias in outcomes, it alleges secrecy in process. The plaintiffs argue that Eightfold AI operated as an unregistered consumer reporting agency under the Fair Credit Reporting Act by scraping personal data on over one billion workers, scoring applicants from zero to five on their "likelihood of success," and discarding low-ranked candidates before any human review — all without the statutory disclosures, consent, or dispute mechanisms the FCRA requires. The suit was brought by former EEOC Chair Jenny R. Yang and the nonprofit Towards Justice.

The distinction matters: Mobley attacks discriminatory outcomes; Eightfold attacks opaque process. One says the algorithm was biased. The other says the algorithm existed in secret. Both point in the same direction — that AI hiring vendors cannot avoid accountability by characterising themselves as passive tool providers.

iTutorGroup (EEOC Settlement, August 2023) — The first EEOC settlement involving AI hiring discrimination. iTutorGroup, a virtual tutoring company, had programmed its recruitment software to automatically reject female applicants over 55 and male applicants over 60. This was not subtle algorithmic bias; it was explicit age-based exclusion coded directly into the system. The company settled for $365,000. EEOC Chair Charlotte Burrows stated that employers cannot rely on AI to make employment decisions that discriminate on the basis of protected characteristics, regardless of whether the discrimination is carried out by a person or a machine.

ACLU v. Aon Consulting (FTC complaint, May 2024) — The ACLU challenged three of Aon's AI hiring assessment tools as discriminatory against people with disabilities and certain racial groups, and disputed Aon's marketing claim that its tools were "bias-free."

ACLU / EEOC charges against HireVue and Intuit (March 2025) — Filed on behalf of a deaf, Indigenous woman who alleged HireVue's automated video interview platform discriminated against her when she applied for an internal promotion at Intuit. The platform's automated speech recognition allegedly performs worse for speakers with different speech patterns. Both HireVue and Intuit denied the allegations.

Harper v. Sirius XM Radio (ongoing) — Arshon Harper, an African American applicant, alleges that 149 of his job applications to Sirius XM Radio were rejected by the company's AI hiring tool, despite his qualifications meeting or exceeding the listed requirements. The case is ongoing.

8. What This Means for Jobseekers

The honest summary is less dramatic than the headlines but more useful.

Most recruitment technology is administrative. Your CV is stored in a database, searched by a recruiter using keywords, and filtered by eligibility criteria that a person configured. The ATS is not reading your CV and passing judgement; it is filing your CV and making it searchable. The quality of the search depends on the quality of the person searching.

Where genuine AI assessment exists — pymetrics games, Eightfold scoring, HireVue video analysis — it is concentrated in large-volume hiring at major employers, not across the market as a whole. If you are applying to an SME with 200 employees, the chances of encountering algorithmic screening of any sophistication are low. You are far more likely to encounter a recruiter who has not read your CV properly than a machine that has read it too precisely.

The 75% rejection statistic is a commercial fiction. The Stanford study is about one niche tool, not about recruitment technology in general.

The court cases are real and important, but they concern specific vendors and specific practices, not a universal condition.

What you can usefully do: write a clear, well-structured CV that describes what you have done and what you are good at, in language that relates to the roles you are applying for. Do not keyword-stuff. Do not pay for "ATS-beating" services built on a debunked statistic. Do not assume that silence after an application means a robot rejected you — it is far more likely that a person has not yet looked, or has looked and chosen someone else, or has been pulled onto a different priority and will never reply. These are human failures, not algorithmic ones, and they predate AI by decades.

If you are asked to complete a game-based assessment, a video interview scored by software, or any process where a third-party tool appears to be making decisions about your candidacy without human review, you are entitled to ask what technology is being used and how it informs the decision. In the UK, the Equality Act 2010 and GDPR's provisions on automated decision-making provide some framework for challenge, though enforcement remains limited. In the US, the legal landscape is evolving rapidly through the cases described above.

The best defence against bad technology in recruitment is not better technology on the candidate side. It is transparency, regulation, and the kind of scrutiny that the Stanford researchers, the courts, and — occasionally — the industry itself are beginning to apply.

Sources consulted: Bommasani et al., "Algorithmic Monocultures in Hiring," FAccT 2026; Fuller & Raman, "Hidden Workers: Untapped Talent," Harvard Business School / Accenture, 2021; Placementist, "Fear-farming the Stanford AI hiring study," June 2026; Christine Assaf, "Your Job Application Was Rejected by a Human, Not a Computer"; Mobley v. Workday, Inc., No. 3:23-cv-00770 (N.D. Cal.); Kistler et al. v. Eightfold AI, Inc. (Contra Costa County, January 2026); EEOC v. iTutorGroup, August 2023; ACLU complaints against Aon Consulting (May 2024) and HireVue/Intuit (March 2025); Enhancv recruiter survey; Jobscan ATS documentation.