The class of 2026 is entering the toughest job market in years. Entry-level hiring has slowed, while AI has driven the barrier to submitting applications to zero. The result: employers are receiving nearly three times the volume of entry-level applications they saw in 2022.
Ninety percent of US employers use AI to screen and rank candidates, and most rely on the same handful of third-party vendors. Stanford HAI researchers tracked 4 million applications submitted by 3.4 million individuals — spanning 150 employers, 1,700 positions, and 11 industries — all evaluated by the same AI hiring vendor.
The conclusion is stark: AI hiring tools not only exhibit racial bias, but because multiple companies share the same algorithm, a candidate rejected by one is also rejected by the others.
40,000 Missing Recommendations
The study used the EEOC’s “four-fifths rule” to measure adverse impact: when a group’s recommendation rate falls below 80% of the highest-recommended group’s rate, the position is flagged as discriminatory. Title VII employment law treats this as prima facie evidence of discrimination.
The results: 26% of Black applicants and 15% of Asian applicants applied for positions where the AI discriminated against their racial group. If the AI recommended Black and Asian candidates at rates comparable to the highest-recommended group (typically white applicants), an additional 40,000 applications would have advanced to the next round.
There’s a statistical trap worth noting. If you aggregate recommendation results across all positions — treating the vendor as a “single giant hiring process” — the data shows no adverse impact. That’s because the AI may frequently recommend Black applicants for some roles (e.g., warehouse positions) while rarely recommending them for others (e.g., finance). The two patterns cancel each other out in the aggregate pool, making everything look fair. But disaggregate by position, and the discrimination is right there.
Algorithmic Monoculture
“Algorithmic monoculture” is a theoretical concept the research team previously proposed: when multiple decision-makers rely on the same algorithm, the algorithm’s biases get systematically amplified. This study is the first to validate that hypothesis with real-world data.
The key finding: when job seekers submit applications to multiple positions screened by the same AI vendor, the probability of being rejected by all positions is significantly higher than the statistical-independence baseline. Among applicants who submitted four applications, 10% were rejected across the board.
The research team compared this against the largest prior hiring-decision dataset — 83,000 applications sent simultaneously to 108 Fortune 500 companies, without restricting to AI users — and found that the control group’s all-rejection rate matched the statistical-independence expectation.
This means market concentration is the critical variable: when a single AI hiring vendor dominates screening in a given industry, the probability of systematic candidate exclusion rises.
The Vendor’s Statistical Shell Game
The study also revealed a methodological loophole vendors use to evade discrimination accusations.
If you aggregate all positions processed by a vendor into an overall assessment, discriminatory patterns across different positions cancel each other out, and the aggregate numbers look clean. But this ignores a basic fact: job seekers don’t apply to “vendors” — they apply to specific positions. Getting recommended for a warehouse role and rejected for a finance role — those two outcomes don’t “cancel out” anything in statistical terms, because they’re different life trajectories.
This loophole also exists at the legal level. The EEOC’s adverse impact assessments are typically conducted by position, but AI vendors can argue for “system-level” evaluation — mixing all positions together to “average out” the discrimination signal.
Three Traits That Should Never Coexist
The research team captured the structure of the problem in a single sentence: “AI hiring tools simultaneously possess three traits that should never appear together: widespread adoption, high stakes, and external opacity.”
When an automated decision system:
- Covers 90% of employers
- Determines whether someone gets an interview
- Operates with logic invisible to the outside world
All three conditions met at once, you’re looking at a black-box power node with no checks and balances.
The study’s most valuable contribution isn’t confirming that “AI has bias” — that’s already known. It’s quantifying how market concentration amplifies individual bias into systematic exclusion. “AI is biased” is old news. “How the same algorithm gets one person rejected by every company simultaneously” — that’s the new problem.
The New Variable: LLMs and Agents
The research team’s conclusion flags a trend worth watching: next-generation hiring tools are starting to use language models and AI agents. These models are more capable, less predictable, and harder to audit for bias.
Given the current progress of LLMs in code generation and writing, hiring screening is shifting from “keyword matching + structured scoring” to “conversational assessment + holistic judgment.” The latter is harder to audit — because the basis for judgment is no longer a set of discrete scoring dimensions, but an end-to-end black-box reasoning process.
This article draws on publicly available information and community discussions. If you have deeper first-hand experience with this topic, corrections and additions are welcome.