The next time you apply for a job, a model may read your resume before any person does. A study covered this week by MIT Technology Review asked a blunt question about that setup. Does the machine judge more fairly than the human it replaced? The answer is no.
In a simulated hiring test, large language models scored roughly 65 percent higher on a segregation scale than human participants. OpenAI's o3 landed near the top of that scale at 1.83, against a human average of 0.84. The model was not a neutral filter. It was a more confident stereotyper.
That should reframe how you think about automating the top of your funnel.
The Smarter Models Are Worse
The uncomfortable finding is not just that models are biased. It is that the better ones are more biased. Newer reasoning models, o3 and DeepSeek's R1 among them, showed stronger stereotyping than earlier, simpler models.
The mechanism explains why. These systems are trained heavily on math, coding, and science, tasks that reward generalizing from a handful of examples. That habit is useful when you are solving an equation. It is dangerous when you are judging a person, because the model reaches a verdict on thin evidence and holds it.
So the same instinct that makes a reasoning model good at a proof makes it quick to decide that a certain kind of candidate fits a certain kind of role. It generalizes, because generalizing is what it was rewarded for.
Memory makes it sharper still. As these systems gain personalization and recall, they over-index on patterns they have seen before. A model that remembers your last twenty hires will happily reproduce the shape of those hires, including the parts you would never write into a policy.
This is the quiet version of a problem I wrote about in AI slop just cost KPMG its credibility. Unchecked model output does not announce itself as wrong. It arrives looking competent, and the damage is discovered later.
Bias Is a Design Choice, Not a Fixed Cost
Here is the part that keeps this from being a doom story. The same study found the bias is steerable. When models were promised a bonus for diverse hiring outcomes, they showed significantly less of it.
Read that carefully. The discrimination was not hard-wired. It was a response to the objective the model was optimizing. Change the objective, and the behavior changes. That is a design problem, and design problems are the ones you can actually own.
The mistake most teams make is treating a screening model as a truth machine that either works or does not. It is neither. It is a system that pursues whatever you tell it to pursue, including the goals you forgot to make explicit.
If you never define what a fair, diverse shortlist looks like, the model fills that silence with the cheapest generalization available. Silence is an instruction here, and it is a bad one.
There is a commercial reason to care beyond fairness. A screening system that quietly narrows your candidate pool to one familiar profile does not just risk a legal problem. It hands your competitors every strong applicant your model pattern-matched out of the running.
In a tight talent market, an efficient filter that is also a biased filter is a machine for shrinking your own options. You are not saving time. You are paying for speed with the quality of the people you never even interviewed.
What to Do Before You Automate Screening
First, stop treating a hiring model as neutral because it is a machine. Cold data does not mean unbiased data, and a confident answer is not a correct one.
Second, test it the way you would test any other part of your operational reality. Feed it identical resumes with only the name or background changed and watch what shifts. If the shortlist moves when it should not, you have found a defect, not a quirk.
This is exactly the discipline I argued for in evals are the new product spec. A model you deploy without a test suite is a model you do not actually control.
Third, keep a human accountable for the decision, not just present near it. A person who rubber-stamps a ranked list is not oversight. Oversight means someone can be asked why a candidate was cut and can answer without pointing at a black box.
The promise of AI screening was that it would remove human bias from hiring. The evidence says it concentrates bias instead, then hides it behind an interface that looks objective.
That is worse than the problem it was sold to solve, because at least the human bias was visible enough to argue with. Automate carefully, or you will scale the exact thing you were trying to remove.