AI in recruitment: what it does, where it fails, and the bias question

AI in recruitment is software that does parts of a recruiter's work: drafting job descriptions, finding candidates, screening CVs, answering applicants, and running or scoring first-round interviews. It saves time on the high-volume steps. The main risk is bias. A model trained on past hiring decisions, or one that reads a stand-in for a protected characteristic such as a name, an accent or a gap in a CV, can turn away a whole group of applicants faster than a recruiter could. Using it with care means knowing what the tool scores, comparing its results across groups of applicants, and keeping a person responsible for the hiring decision.

What AI does at each stage of hiring

The UK Information Commissioner's Office (ICO) audited developers and providers of AI recruitment tools in 2024 and published the results in November 2024. It sorted the tools into sourcing, screening and selection. Its audit left out generative AI that drafts job adverts or chats with applicants, which SHRM's 2024 survey shows employers also use.

Sources: the ICO's audit outcomes report (November 2024) for sourcing, screening, assessment and interviews; SHRM's January 2024 survey for job descriptions and applicant messages.

How many employers use AI in hiring

In the 2024 survey, nearly 2 in 3 of the organisations that use AI in recruiting use it to help write job descriptions. About 1 in 3 use it to review or screen CVs, to communicate with applicants during the interview process, or to automate candidate searches. Nearly 9 in 10 said they use AI to save time or work more efficiently. The 2025 press release says 43% of organisations "now integrate AI into HR functions, with 51% utilizing it for recruitment tasks such as resume screening and candidate communication."

Candidates are warier. Pew Research Center surveyed 11,004 US adults from 12 to 18 December 2022 and published the results on 20 April 2023. In that survey, 66% said they would not want to apply for a job with an employer that uses AI to help make hiring decisions, and 71% opposed AI making the final hiring decision, against 7% in favour. More of them, 47%, thought AI would do better than humans at "evaluating all job applicants in the same way" than thought it would do worse (15%).

What is an AI recruiter?

People use "AI recruiter" for two different products. The first is an AI agent that works through recruiting tasks on a recruiter's behalf, such as searching for candidates who fit a role. On 3 September 2025, LinkedIn announced that Hiring Assistant, which it calls "LinkedIn's first AI agent for recruiters", would be globally available in English by the end of that month. The second is an AI interviewer that runs the first-round interview and scores the answers; the formats are covered in what is an AI interview.

Either kind leaves the employer subject to anti-discrimination law. New York City's guidance on automated hiring tools, for example, reminds employers that federal, state and city laws prohibit discrimination, and that they must comply with those laws whatever a tool's bias audit shows.

Where AI in recruitment has failed

A CV model that learned to prefer men

Reuters reported on 10 October 2018 that Amazon had built a model to score job applicants' CVs from one to five stars. It was trained on CVs submitted to the company over 10 years, most of them from men. According to people familiar with the project, the model penalised CVs that included the word "women's", as in "women's chess club captain", and downgraded graduates of two all-women's colleges. Amazon edited the programs to be neutral to those terms, but that "was no guarantee that the machines would not devise other ways of sorting candidates that could prove discriminatory." The team was disbanded by the start of 2017. Amazon's recruiters looked at the tool's recommendations but never relied solely on its rankings, the sources told Reuters.

A filter that rejected older applicants

In September 2023, iTutorGroup agreed to pay $365,000 to settle a suit brought by the US Equal Employment Opportunity Commission (EEOC). According to the EEOC, the company programmed its tutor application software to automatically reject women aged 55 or older and men aged 60 or older, and so rejected more than 200 qualified US applicants because of their age. Here the bias came from a rule the company programmed into its software.

Language models that rank CVs by name

University of Washington researchers took more than 550 real CVs and varied the names on them, using names associated with white and Black men and women. Three large language models ranked the CVs against more than 500 real job listings, in over three million comparisons. The models favoured white-associated names 85% of the time and Black-associated names 9% of the time, and male-associated names 52% of the time against 11% for female-associated names. They never preferred names perceived as Black men's over names perceived as white men's. The university published the study on 31 October 2024.

Speech recognition that mishears some speakers

Interview tools that score a transcript depend on speech recognition, which does not work equally well for every speaker. A 2020 study in PNAS tested speech recognition systems from Amazon, Apple, Google, IBM and Microsoft on 19.8 hours of recorded conversational interviews with 42 white and 73 Black speakers. The average word error rate was 0.35 for Black speakers and 0.19 for white speakers.

Natalie Sheard's 2025 study in the Journal of Law and Society reports what HireVue's own documentation says about its transcription: a word error rate below 10% on average for US English speakers, rising to between 12% and 22% for non-native English speakers depending on their country of origin, and 22% for non-native speakers from China.

A vendor taken to court

In Mobley v. Workday, applicants allege that Workday's AI applicant recommendation system discriminated by race, age and disability. On 16 May 2025, a federal court in California granted preliminary certification of a collective action on the age claim. Mobley and four other plaintiffs over 40 allege they applied for hundreds of jobs through Workday and were rejected almost every time without an interview. The court described the critical issue as "whether that system has a disparate impact on applicants over forty." The order did not decide that question; at that stage the claims were allegations.

What a regulator found in its audits

The ICO's 2024 audits found that many providers monitored their tools for accuracy and bias, but in some cases accuracy was not tested. Some tools let recruiters search for candidates in a way that filtered out people with certain protected characteristics. Others estimated candidates' gender, ethnicity and other characteristics from their application, or only from their name, which the ICO found was "not accurate enough to monitor bias effectively", and often done without a lawful basis or the candidate's knowledge. Some collected far more personal information than they needed. The ICO made 296 recommendations and 42 advisory notes, and says all of its recommendations were accepted.

AI bias in hiring: how it gets in and how it is measured

How bias gets in

The cases above show four routes:

1. Training data that records past decisions. Amazon's model learned from a decade of CVs, most of them from men.
2. Proxies. A feature can stand in for a protected characteristic: a name for race or sex in the University of Washington study, or, Sheard notes, gaps in employment history for gender, since women are more likely to have taken time out to care for children.
3. Measurement. A transcript with more errors for some accents gives those candidates worse material to be scored on. Sheard also points out that the time limits an employer sets for answers may disadvantage applicants from non-English-speaking backgrounds.
4. Configuration. iTutorGroup's age rule and the search filters the ICO found were settings that people chose.

How it is measured

The usual measure compares selection rates. A group's selection rate is the share of its applicants that the tool passes. Divide each group's rate by the rate of the group selected most often, and you get the impact ratio.

The US Uniform Guidelines on Employee Selection Procedures set the benchmark known as the four-fifths rule. A selection rate for any race, sex or ethnic group that is less than four-fifths (80%) of the highest group's rate "will generally be regarded by the Federal enforcement agencies as evidence of adverse impact." Smaller differences can still count when they are significant in both statistical and practical terms. As an illustration, if a screen passes 50 of 100 male applicants and 30 of 100 female applicants, the women's impact ratio is 30% divided by 50%, or 0.6, below the 0.8 line.

New York City's Local Law 144 requires this calculation in an audit. An employer may not use an automated employment decision tool unless an independent auditor has done a bias audit within one year of its use, a summary of the results is public, and candidates have been notified. The audit must at least calculate selection or scoring rates and impact ratios across sex, race and ethnicity, and their intersections. The law does not require any particular action when the ratios are low; anti-discrimination law still applies.

Deleting the protected field from the data does not remove its proxies: Amazon made its model neutral to particular terms and still could not rule out other discriminatory sorting. Inferring candidates' demographics in order to monitor the tool is not reliable either, according to the ICO, which pointed to a provider that collected demographic information directly from candidates, optionally and only for monitoring bias, as good practice.

What the rules ask

Illinois has its own law on AI-analysed video interviews, set out with the other rules for AI interviews in what is an AI interview. This is general information, not legal advice; check the official source for your case.

Questions to ask before you buy an AI recruiting tool

1. What exactly does it score, and what does it leave out? Ask for the list of inputs.
2. Does it analyse video, voice or facial expressions? In the EU, emotion recognition at work is banned.
3. How accurate is its transcription for the accents and languages of your applicants? Ask for word error rates by group, the kind of figure HireVue's documentation gives.
4. Has an independent auditor checked it for bias? On whose applicants, and what were the impact ratios by group?
5. Can you run your own selection-rate check on your applicants, with demographic data the candidates give you?
6. Who makes the decision, and can a person see the evidence behind every score?
7. What personal data does it collect, what else is it used for, and how long is it kept?
8. How does a candidate ask for an adjustment, or challenge a result?

Assessment methods differ in how well they predict job performance in the first place; structured interviews do best in the research summarised in skills-based hiring.

What ROUND.1 scores, and what it does not

ROUND.1 is a hiring marketplace for Hong Kong and Singapore, built by h.work. Each candidate takes one AI first-round video interview, in a web browser with a camera and a microphone. It lasts about 20 minutes by default (roles can be shorter), 30 at most, and each role is interviewed in one language, English or Mandarin Chinese. Before it starts, the candidate consents to the processing of their video and voice under Hong Kong's PDPO and Singapore's PDPA.

The same interview is reused for the other roles the candidate chooses, so a candidate is not interviewed again for each company. A company first sees the scored evaluation without the candidate's name or contact details, which open to the company when it shortlists the candidate after the interview. A person at the company decides whom to hire; ROUND.1 does not make the hiring decision.

The studies and cases above are about AI hiring tools in general. None of them measures ROUND.1's interview. The full list of facts and limits is on ROUND.1, explained, and the candidate's side of the format is in one-way video interviews.