The algorithm said no. 3.4 million times.
What Stanford’s landmark study tells us about the AI hiring system nobody is auditing
A new paper from Stanford’s Digital Economy Lab just dropped, and if you work in talent, HR, or workforce policy, you need to read it.
The researchers analysed 3.4 million real job applicants submitting 4 million applications to 156 employers across 11 sectors. Every single application was assessed by algorithms built by the same vendor.
What they found should change how we think about hiring AI.
The same people are being rejected. Everywhere.
When multiple employers use the same algorithmic vendor to screen candidates, they don’t just get similar results. They get identical ones. A candidate rejected by one employer’s algorithm is highly likely to be rejected by another’s, because the underlying system is the same.
The researchers call this “algorithmic monoculture.” And the data shows it has a direct, measurable impact on racial equity.
25.87% of all applications submitted by Black candidates go to positions that adversely impact them under Title VII, the relevant US employment discrimination standard. For Asian applicants, the figure is 14.74%. If Asian applicants were hired at the same rate as the most-selected group for each position, 29,000 additional applications would have been recommended.
These are not marginal numbers.
The bias was being hidden by how we measure it
Here’s what makes this study particularly important. Previous research found minimal adverse impact in algorithmic hiring data. Not because the bias wasn’t there, but because prior studies looked at aggregate vendor data rather than individual positions.
When you blend all positions together, disparities average out. When you look at each position separately, as Title VII actually requires, a very different picture emerges.
The researchers found that adverse impact was being systematically concealed by aggregate-level analysis. Position-by-position measurement reveals what blended data hides.
You need to apply 25 times to have a fair shot
The study also ran simulations to understand how widely a candidate would need to apply in order to get at least one recommendation. Under normal, independent hiring decisions, 10 applications would give you a 99.9% chance of at least one positive outcome.
Under algorithmic monoculture? You need 25.
For 10% of applicants who submitted 4 applications, the algorithm rejected them from every single position. Not because they lacked capability. Because the same system, applied at scale, reached the same conclusion repeatedly.
| Independent employers | 10 |
|---|---|
| Shared algorithm | 25 |
| Independent employers (expected) | about 4% |
|---|---|
| Shared algorithm (observed) | 10% |
Source: Bommasani et al., Algorithmic Monocultures in Hiring (FAccT 2026), Figures 2a and 4. Redrawn by Gradstack, CC BY 4.0.
This is what “AI screening” looks like in practice
Over 90% of US employers now use hiring algorithms to screen candidates. More than 60% of the Fortune 100 use one vendor’s tools.
When an applicant doesn’t make it through, they rarely know why. There is no explanation. No feedback. No pathway to reconsider. The algorithm said no, and that was the end of it.
The researchers are clear that this isn’t just a fairness problem. It’s a structural one. Concentrated reliance on a small number of systems creates correlated failures across the entire market. A flaw in one vendor’s model doesn’t affect one employer. It affects all of them, simultaneously.
What this means for hiring, and for Gradstack
This research reinforces something we built Gradstack around from day one.
AI should not be used to source, screen, or make hiring decisions. That is not what it is for.
At Gradstack, AI is used for one purpose only: matching verified capability data to verified workforce seats ® and role requirements inside those seats. The VCI® (Verified Capability Index) and DHP® (Digital Human Profile) give employers a multi-layered, evidence-based picture of what a candidate can actually do. Not an algorithmic score. Not a black-box recommendation. Verified capability, matched transparently.
The candidate is not filtered out before a human sees them. The decision remains with the employer, informed by evidence rather than replaced by it.
The Stanford study documents what happens when AI replaces human judgement at the screening stage. Gradstack is built on the position that AI should never do that. It should give humans better information to make better decisions.
The difference matters. Especially for the 29,000 Asian applicants who should have been recommended. And the 25.87% of Black applications directed to positions already stacked against them.
Verified capability is not just a better hiring model. Right now, it is a more ethical one.
Read the full Stanford study: algorithmichiring.github.io