Algorithmic monocultures in hiring: what the Stanford HAI write-up suggests, and how to interpret it in a UK context

Summary (for skimmers)

At a glance (US study)

Stanford HAI reports on a large-scale US study of algorithmic screening in hiring, analysing 4 million applications from 3.4 million applicants across 150 employers in 11 sectors, all assessed via a single third-party vendor.

Using the EEOC’s four-fifths rule at the position level, the researchers report evidence of adverse impact affecting Black and Asian applicants in a substantial share of postings.

They also document “systemic rejection” patterns consistent with algorithmic monoculture effects, where shared reliance on one vendor can correlate outcomes across employers.

Scope and limitations (read this first)

Key terms (plain definitions)

This study looks at what happens at the screening stage, before a human interview. Applicants submit applications, the vendor’s system produces a label like “recommend” or “do not recommend”, and that label is sent to employers to inform next steps.

The researchers then check for “adverse impact” using the EEOC’s four-fifths rule, and they emphasise that you need to measure this per job, not just as one big average across all roles.

They also test whether outcomes become more similar across employers when many employers rely on the same vendor, which is where the “algorithmic monoculture” and “systemic rejection” risk comes in.

Method in plain English

• Algorithmic hiring / screening: Software that helps sort, rank, or filter applicants.
• Adverse impact (US framing): A pattern where a protected group is selected or recommended at a meaningfully lower rate than another group.
• Four-fifths rule: A common threshold used to flag potential adverse impact when one group’s selection rate is less than 80% of the most-selected group.
• Algorithmic monoculture: Many employers relying on the same vendor or model, creating shared dependencies.
• Systemic rejection: A pattern where applicants are rejected everywhere they apply, at a rate higher than expected if decisions were independent.

What the Stanford HAI write-up reports

Dataset

• 3.4 million people
• 4 million applications
• 150 employers
• 11 industry sectors
• One third-party vendor’s screening outputs

Findings: adverse impact appears at the job level

Stanford HAI reports that, when evaluated per position, 26% of Black applicants and 15% of Asian applicants applied to roles where the AI system disadvantaged their group under the four-fifths rule.

A key methodological warning is that pooling results across all jobs can hide discrimination. If one group is recommended more often in one type of role and less often in another, the average can look “fine” while role-level outcomes remain unequal.

Findings: correlated outcomes and systemic rejection

The write-up reports that applicants screened through the same vendor across multiple applications experience more correlated outcomes than expected under independence.

It states that 10% of applicants who submit four applications are rejected everywhere they apply.

The authors argue this is a market concentration risk. When one vendor dominates screening in a sector, correlated failures can scale.

Ten percent of applicants who submit four applications are rejected from all the places to which they apply.

AI screening tools bring together three properties that should not co-exist in high-stakes decision-making: They are pervasively adopted, highly consequential, and opaque to the public.

In the same study, 26% of Black applicants and 15% of Asian applicants applied to roles where the AI system showed adverse impact against their racial group.

The researchers estimate that if the AI had recommended Black and Asian candidates at the same rate as the most-favoured group, 40,000 more applications would have progressed to the next stage.

So this is not just tough luck in a tough market. It is bias and systemic rejection, at scale, inside tools most candidates cannot see or challenge.

When one vendor’s model sits upstream of hundreds of employers, a single black box can quietly shape who gets a shot, and who gets shut out everywhere.

• Human route in: employers can use structure and tools, but humans remain accountable for decisions, especially in high-stakes screening.
• Bias-aware by design: we assume bias is likely, so safeguards and monitoring are part of the design, not a retrofit after harm.
• No monoculture scoring: we avoid relying on a single opaque pass or fail score that can be reused across employers as a universal filter.
• Explainable matching: we prioritise explainability so candidates can understand why they matched, and what to do next, rather than receiving an unchallengeable no.
• Strengths-first, not pedigree-first: we aim to surface potential and fit, not reward polish, school brand, or confidence signals.

The Stanford HAI write-up is a warning about the risks of opaque, high-stakes screening at scale. It also gives a clear build brief for anyone designing hiring pathways for young people.

What good looks like (and what we’re building at inyourroots®)

Pull quotes (from the Stanford HAI write-up)

UK guidance to pair with this research

For UK context, we recommend reading:

1. 1.ICO: Automated decisions can streamline the hiring process, with the right safeguards in place
2. 2.ICO: Recruitment rewired
3. 3.ICO: Here’s what jobseekers need to know about automated recruitment decisions
4. 4.UK Government: Responsible AI in Recruitment

Practical implications (inyourroots® lens)

• It focuses on one vendor at scale, so readers should be careful about generalising to every tool, especially newer systems built on language models.
• It measures screening recommendations (recommend / do not recommend), not necessarily final hiring decisions.
• It does not test compliance with the UK’s Equality Act 2010, it uses US legal framing (EEOC, Title VII).
• It does not prove the same rates of adverse impact exist in every country or every UK hiring process, because the dataset is US-based.

What this does not prove (and why that matters)

• Measure fairness per role, not just overall.
• Reduce monoculture risk by avoiding over-reliance on a single vendor, and by maintaining human oversight.
• Offer alternative routes for entry-level applicants, especially those without traditional CV signals.
• Design for explainability and support, so applicants understand what’s expected and how to improve.

References (links)

• Stanford HAI article: https://hai.stanford.edu/news/ai-hiring-tools-can-yield-racial-bias-and-systemic-rejection
• Project page: https://algorithmichiring.github.io/
• arXiv paper: https://arxiv.org/abs/2605.27371
• ICO (automated decisions and hiring safeguards): https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/03/automated-decisions-can-streamline-the-hiring-process-with-the-right-safeguards-in-place/
• ICO (Recruitment rewired): https://ico.org.uk/about-the-ico/what-we-do/recruitment-rewired/
• ICO (jobseekers and automated recruitment decisions): https://ico.org.uk/about-the-ico/media-centre/news-and-blogs/2026/03/here-s-what-jobseekers-need-to-know-about-automated-recruitment-decisions/
• UK Government (Responsible AI in Recruitment): https://www.gov.uk/government/publications/responsible-ai-in-recruitment-guide/responsible-ai-in-recruitment