Part 1: AI in Hiring: What the Tools Do and What the Rules Say

AI in hiring, from CV screening and AI interviews to candidates who apply with AI themselves, then what the evidence says about quality and bias, and the EU AI Act and GDPR rules that apply. Part 1 of 3 on AI and HR.

Tue, 27 Oct 2026 · 15 min

Made for you, by AI, from the sources below.

Listen to it in AI and HR

Transcript

The script the two voices were given, with the sound cues taken out. A voice may say a word a little differently, so the audio is the episode.

Mia: Welcome to Walking Points. Workday says its software was used to reject 1.1 billion job applications between September 2020 and this year. That is the company's own figure, and it sits at the centre of a lawsuit that is still running.

Adam: Wait, a billion rejections by software? Who decided that was acceptable?

Mia: That is the question this whole episode circles. The topic is AI in hiring: the software that now sits at almost every step between a job advert and a job offer. This part was made in October 2026. We'll take three things in order. First, what these tools actually do at each step. Second, what the evidence says about whether they pick better people and whether they are biased, with a twist about candidates who now apply with AI too. And third, the rules that matter most in Europe: the EU AI Act, and a GDPR rule about decisions made solely by a machine. It earns a whole episode because hiring is where a software decision changes someone's life, and where Europe has written its most specific law.

Adam: Okay, so start with the basics. I picture a CV filter, and that's about it.

Mia: That's the step most people mean, but the funnel is longer. At the top, algorithms on job boards and social platforms decide who even sees an advert. Then sourcing tools search databases and profiles for people who might fit. Then comes CV screening: software parses applications, matches them against the job description, and ranks or filters them. After that come assessments, meaning games and tests that score skills, one-way video interviews where you answer recorded questions, and newer AI voice agents that hold a live phone conversation. Chatbots answer questions and book interviews, and at the end some tools summarise interviews or suggest who gets the offer.

Adam: So there's hardly a step left where a person is the first one to look at you.

Mia: Roughly, though it differs a lot between employers. And two older stories explain why this is sensitive. In 2018 Reuters reported that Amazon had scrapped an experimental recruiting tool that rated candidates from one to five stars. It was trained on ten years of the company's own CVs, which came mostly from men, and it learned to downgrade CVs containing the word women's, as in women's chess club captain. Then in January 2021 the video-interview company HireVue dropped facial-expression analysis from its product, after criticism that scoring faces could disadvantage people with disabilities, among others.

Adam: So were the Amazon engineers just careless?

Mia: No, and that's the lesson. A model trained on past hiring decisions learns the past, habits included. Nobody had to program the bias in, because it was already in the data.

Adam: Fine, but that's an old experiment that failed. Is there any evidence the newer tools actually pick better people?

Mia: The honest summary is that the evidence is thin, mixed, and mostly produced by the vendors themselves. There are a few strong independent exceptions. The best is a field experiment by Brian Jabarian of the University of Chicago Booth School and Luca Henkel, called Voice AI in Firms. They randomly assigned about 70,000 applications, for 48 customer-service style job postings at a recruitment firm in the Philippines, to be interviewed either by a human recruiter or by an AI voice agent. Applicants interviewed by the AI were 12 percent more likely to receive job offers. More of them actually started, retention was better, and the productivity of those hired did not decline. And 78 percent of applicants chose the AI interviewer when they were given the option.

Adam: That sounds suspiciously like an advertisement for the machine. What's the catch?

Mia: There is a real limit. It's one setting: high-volume entry-level work in the Philippines, with the AI acting as an interviewer inside a process that humans still ran. It doesn't show that AI picks better managers, or that it works the same in a Danish or German workplace.

Adam: So why did it work at all?

Mia: The authors' reading is that the AI interviews were more structured and consistent, while still responsive to the candidate, so they collected more hiring-relevant information. Decades of hiring research favour structured interviews. A consistent interviewer is exactly what a machine finds easy to be.

Adam: So the machine wins by being repeatable rather than by being clever, which is a less dramatic story than the one the vendors tell.

Mia: Right, and the wider pattern is the counterweight. Most vendor claims of predictive accuracy come without a method anyone can check, and the tools are usually validated against the employer's past decisions, or against managers' ratings. That's the Amazon problem again: a tool that is good at predicting what the old process would have done is not necessarily good at predicting who would be good at the job. So AI is clearly good at speed and consistency, plausible as a structured interviewer, and unproven as a judge of who will excel.

Adam: Which brings us to bias. Is there a hard number?

Mia: There is. Kyra Wilson and Aylin Caliskan at the University of Washington tested three widely used text-embedding models, the kind of model that sits under many CV-matching tools. They used 554 real CVs and 571 job descriptions, producing more than three million comparisons, and they changed only the first name on otherwise identical CVs. They used 120 names strongly associated with Black women, Black men, white women or white men. The models favoured white-associated names in 85.1 percent of cases, and female-associated names in only 11.1 percent. In some tests, Black male names lost to white male names in up to 100 percent of the comparisons. It was presented at the AAAI/ACM Conference on AI, Ethics, and Society.

Adam: Mia, hold on. Are those the tools employers actually buy, or something from a lab?

Mia: Something from a lab, and that matters. These were off-the-shelf research models in a controlled setup, not a vendor's product with its own safeguards. The study shows what the raw material is like, which is exactly why audits of the actual product matter.

Adam: And when a product does discriminate, who's on the hook, the employer or the company that built it?

Mia: That is what the US case Mobley v. Workday is about. A quick word first: this is a briefing, not legal advice, and a lawyer is who you'd ask about a real case. An applicant over 40 says Workday's screening tools rejected him across many employers. Judge Rita Lin in Northern California reasoned that when a vendor's AI performs a traditional hiring function such as screening, ranking or rejecting, the vendor can be treated as the employer's agent, so anti-discrimination law can reach the vendor, not only the employer. In February the court authorised notice to a nationwide group of applicants aged 40 and over who were rejected through Workday since September 2020. That is where the billion-plus figure comes from. And nothing is decided: no court has found that Workday discriminated, these are rulings on whether the case can go forward.

Adam: So for anyone buying one of these tools, the vendor's tool did it is not much of a defence.

Mia: Not one anyone should count on. The other US example is New York City's Local Law 144, in force since 2023. An employer using an automated employment decision tool for hiring or promotion must have an independent bias audit within the last year, publish the results, and give candidates notice. In December 2025 the New York State Comptroller audited how the city enforces it, and called enforcement ineffective. The city agency had received only two complaints in nearly two years, and 75 percent of test calls to the 311 line about such tools were misrouted. The city's own review of 32 companies found one possible violation, and the auditors found at least 17.

Adam: So the law exists, but nobody's really checking.

Mia: The takeaway isn't that the law is dead. A disclosure-and-audit regime only works if someone checks, and the city has promised to adopt most of the recommendations.

Adam: You mentioned a twist earlier, that candidates are using AI as well.

Mia: They are, and it changes the picture. Surveys reported this year say that 78 percent of European candidates use AI in their job search, by one summary, and that 77 percent of hiring teams regularly meet AI-generated or AI-assisted applications, up from 53 percent a couple of years earlier. LinkedIn receives about 11,000 applications every minute. Those come from secondary roundups of industry surveys, so treat them as indicative, not exact. A Robert Half survey in March found that 67 percent of US HR leaders say reviewing AI-generated applications has slowed hiring.

Adam: So it's an arms race. Candidates write with AI, employers screen with AI, and then candidates tune their applications to beat the filter.

Mia: That's the loop, yes.

Adam: But an arms race usually leaves a winner standing, and here it sounds like both sides just burn more effort. What actually gets destroyed?

Mia: The signal. A polished CV used to carry information, such as effort and writing ability. When anyone can produce one in seconds, it no longer separates people. So your arms-race picture needs a correction: the thing being fought over is the thing being used up. Tools that screen on CV text alone are screening on a signal that has stopped working. Structured, skills-based steps later in the process become more valuable, and live voice interviews, like the ones in that Philippines study, are one way back to a signal.

Adam: Okay, so that leaves the question I can't shake. If software ranks people and a recruiter just works down the list, who is actually deciding?

Mia: That is exactly what the rules try to answer, and it's the second half of the episode. Start with the EU AI Act, Regulation 2024/1689. It sorts AI by risk, and Annex III, point 4 is the employment list. It covers AI used to place targeted job adverts, to analyse and filter applications, and to evaluate candidates in interviews or tests. All of that counts as high-risk. It also covers promotion, termination, task allocation and monitoring, which is the next part of this series.

Adam: And the dates? I thought the high-risk rules were due this summer.

Mia: They were, for 2 August. The Digital Omnibus package moved stand-alone Annex III systems to 2 December 2027. The Council gave final approval on 29 June, and the change is in force. What did not move: the banned practices, including the ban on emotion recognition in the workplace, have applied since 2 February 2025, and so has the AI literacy duty. The transparency duties in Article 50 apply from that same 2 August date as planned.

Adam: Meaning a recruiting chatbot already has to say it's a chatbot.

Mia: Yes, even though the high-risk rules for the screening behind it come later. And for an employer, the deployer in the Act's language, high-risk means a list of duties. Use the system according to the provider's instructions. Give human oversight to people with the competence and authority to act on it. Keep the logs, watch the system, report problems, and make sure the input data is relevant. Before putting it into use at work, inform the workers' representatives and the affected workers. People subject to a decision made with help of such a system are entitled to a clear explanation of its role. The vendor carries the heavier burden, and fines for breaching high-risk obligations go up to 15 million euros or 3 percent of worldwide turnover.

Adam: What if the employer says they only use the tool to help, and a human decides?

Mia: Assisting a decision still counts, so that does not take a system off the list. And an employer that substantially modifies a tool, or puts its own name on it, can take on the provider's duties.

Adam: Then the GDPR rule you mentioned. How does that differ?

Mia: Article 22 gives a person the right not to be subject to a decision based solely on automated processing that has legal or similarly significant effects. A job rejection is the standard example: recital 71 itself names e-recruiting practices without any human intervention. There are exceptions, such as contract, law or explicit consent, but then the employer must give safeguards, at least the right to human intervention, to express a view and to contest the decision.

Adam: Right, but in most companies a recruiter does look at the list. So surely that's not solely automated.

Mia: Adam, that's where the Court of Justice of the EU's SCHUFA judgment of 7 December 2023 comes in. SCHUFA is a credit agency whose score was sent to lenders who drew strongly on it. The Court held that producing the score was itself an automated decision, because in practice the lender nearly always followed it. Apply that to hiring: if a tool ranks candidates and the recruiter in practice only works down the top of its list, the ranking can be the decision, and a human rubber stamp does not save it. The test is whether a person has real authority and actually looks.

Adam: That sounds like a big deal, and I'd guess it's also not that simple.

Mia: Fair. That reading was made about credit, and legal commentators note that how far it extends to hiring is still being worked out by national courts and regulators. As for how the two sets of rules fit: the AI Act regulates the tool and the way it is run, while the GDPR protects the person in any decision made about them. A hiring tool in the EU can fall under both at once, and meeting one does not satisfy the other.

Adam: So back to the billion rejections. The real question was never whether software was involved, it was whether anyone with authority looked.

Mia: Right, and that gives you one step to try this week. If you have any say in hiring, ask whoever runs your hiring software one thing: when it ranks candidates, what does a person actually do with that ranking, and can they overrule it? The reason is the SCHUFA test, because if the answer is that people just work down the list, then the ranking is the decision.

Mia: That was part one of three. Part two is called AI and the people you already employ, and it is about why the AI Act bans emotion recognition at work.

Heard something wrong? Tell us