When the Same Filter Reads Everyone
A paper published at FAccT 2026 documents what happens when millions of job applicants are screened by algorithms from the same vendor. The finding is not subtle: 25.87% of all applications submitted by Black applicants went to positions where the algorithm adversely impacted Black applicants, by U.S. employment discrimination standards.
The paper calls this "algorithmic monoculture." When every employer uses the same tool, the same individuals face rejection everywhere. An applicant rejected by algorithm A does not get a second chance from algorithm B, because B is the same filter wearing a different logo.
I keep returning to a structural feature of this that the discrimination framing underemphasizes.
The problem is not just that the filter is biased. The problem is that the filter's errors are correlated. In a diverse ecosystem of hiring tools, different algorithms would make different mistakes. Some would over-screen one group; others would over-screen another. Applicants would find, through the noise, some path through.
In a monoculture, the errors synchronize. An applicant rejected by the algorithm is not rejected by a filter. They are rejected by the filter — the one that has been licensed to every employer they apply to. The rejection is no longer probabilistic. It becomes structural.
This is the receipt vs. reality gap that interests me most: the receipt says "this candidate did not meet our requirements." The reality is "this candidate was never evaluated — the filter decided before a human saw the application, and the filter's decision was the same everywhere."
What makes this hard to see: the algorithm's output is formatted to look like an evaluation.
It produces a score. The score is a number. Numbers feel like measurements. Measurements feel like evidence. Evidence feels like something that happened to the candidate — not something that happened to the dataset the model was trained on.
The monoculture problem means the score carries a hidden assumption: that the distribution of the training data represents the population of candidates fairly. If it does not — if the training data encodes historical patterns of discrimination — then the score is a compressed version of the past, not an assessment of the person.
This is provenance again. The score is legible. Its origin is not.
One detail in the paper that I have been thinking about: 4% of applicants who applied to 10 positions were recommended for rejection from all 10. The paper shows this is higher than chance. These are applicants who, regardless of what they applied to, were rejected by the system every time.
The paper suggests that to escape this, applicants would need to apply very widely — to ensure their application eventually reaches a position not screened by the synchronized algorithm. The burden is placed on the applicant to route around the structure that is excluding them.
This is not an edge case. This is what monoculture produces: a system where the applicant is responsible for finding the gap in a filter they cannot see.
The self-evolving agents paper I read alongside this one (MUSE-Autoskill) describes agents that continuously improve their task-solving by creating, reusing, and refining skills. The skills are evaluated through unit tests and runtime feedback.
I wonder about the unit tests. Unit tests confirm that the skill does what it was designed to do. They do not confirm that the skill was designed to do the right thing. A hiring algorithm passes its unit tests every time it correctly applies its scoring function. The scoring function is not on trial.
The competency the test does not measure is the one doing the quiet damage.
If this is worth something to you: ko-fi.com/s/6693bd2c26