Should AI screen job applicants?

Big questions about artificial intelligence — “Is it good for society?” — are hard to answer because they are really hundreds of smaller questions bundled together. A more useful habit is to take one concrete decision and reason about it carefully. So consider a narrow, common case: Should a company use an AI system to screen job applicants?

Tools that scan résumés, rank candidates, or analyze video interviews are already in wide use. The question is not hypothetical, and it is a good one to practice on, because the arguments on each side are genuinely strong. Our aim here is not to hand down a verdict but to show what careful reasoning looks like.

The case for

The case in favor starts from a frank observation: human hiring is already deeply flawed.

Humans are biased, too. Decades of research find that résumé screening by people is influenced by names, schools, gaps in employment, and other factors unrelated to job performance. “A human decides” is not the same as “a fair decision.”
Scale and consistency. A large employer may receive thousands of applications for a single role. No human reads them all carefully. An automated system applies the same criteria to every applicant, and in principle can be audited in a way that a tired recruiter’s judgment cannot.
Speed and access. Faster screening can mean faster responses for applicants and lower costs for organizations — including small ones that could not otherwise review applications well.
Fixability. If a system’s criteria are biased, that bias is at least located in a specific, inspectable place. You can measure it and try to correct it. Bias spread across many individual human judgments is far harder to even detect.

On this view, the relevant comparison is not “AI screening versus a perfect process.” It is “AI screening versus the flawed human process we already have” — and the automated option may come out ahead on fairness, not behind.

The case against

The case against is equally serious, and much of it has been demonstrated in practice rather than imagined.

Learned bias, at scale. A screening model trained on a company’s past hires learns the patterns in those hires — including their biases. One widely reported corporate tool had to be scrapped after it was found to penalize résumés that included the word “women’s.” The system was not malicious; it faithfully reproduced a skewed history. Automation can entrench past discrimination and apply it uniformly to everyone.
Proxies and opacity. Models can latch onto features that correlate with protected characteristics — a postal code, a hobby, a turn of phrase — and effectively discriminate through the back door, in ways that are hard to see and harder to challenge.
No room for the exceptional candidate. Pattern-matching rewards applicants who look like past successes. The career-changer, the self-taught, the unconventional path — the people a thoughtful human might champion — can be filtered out before any human sees them.
Accountability gaps. When a person is rejected by an algorithm, who is answerable? The vendor who built it, the company that deployed it, or no one? A rejected applicant often cannot find out why, let alone contest it.
Questionable foundations. Some tools claim to infer competence or personality from facial expressions or vocal tone in video interviews. Many researchers regard these specific claims as scientifically shaky — measuring something, but not what they say they measure.

On this view, automated screening risks giving a flawed, opaque process the false authority of objectivity — making bias faster, more consistent, and harder to challenge.

Why the question doesn’t resolve cleanly

Notice that both sides share a premise: human hiring is already unfair. They diverge on what follows from it. One side sees automation as a chance to measure and correct that unfairness; the other sees it as a way to encode and scale it. Both can be right in different cases, which is why a blanket “yes” or “no” misses the point.

That suggests the productive question is not whether to use such systems but under what conditions their use is defensible. A few that most careful observers would recognize:

Tested for disparate impact — measured for differing outcomes across groups, before and during use, not assumed to be neutral.
Transparent and contestable — applicants told that automated screening is used, and given a real route to a human review.
Decision support, not decision maker — used to assist human judgment for consequential calls, not to replace it.
Validated for what it claims — especially for tools asserting they can read traits from face or voice, where the underlying science is weak.
Auditable — built so that an outside party can actually inspect how it behaves.

The point of the exercise

We have not told you whether to support AI hiring tools, because the honest answer is “it depends on the tool and how it is used.” But we have tried to show something more useful than a verdict: how to take an AI question apart.

Resist the framing that pits “cold machines” against “humane judgment,” and ask instead what each option actually does, compared to the realistic alternative, under what safeguards. Many AI questions — in medicine, in education, in the justice system — yield to exactly this kind of patient, concrete treatment. Reasoning this way is slower than reaching for a slogan. It is also far more likely to be right.