False positives and false negatives in AI candidate screening, explained

Ask a vendor how accurate their AI screening is and you’ll get a percentage. Ask what happens to the candidates it’s wrong about, and most go quiet. That second question matters more, because being wrong about a candidate isn’t one problem. It’s two, and they aren’t equally dangerous.

A false positive is a candidate who scores well and doesn’t work out. A false negative is a candidate who would’ve been great and never made your list. Both are AI candidate screening errors. Only one of them announces itself.

That difference, not a better accuracy score, is the actual argument for a screening tool that ranks everyone and rejects no one automatically. Here’s why.

Accuracy is the wrong thing to worry about first

Every hiring tool that ranks candidates will be wrong sometimes. That’s not a flaw you can shop your way around. Resumes are incomplete, interviews are a snapshot, and any system, human or software, is working from partial evidence about a stranger.

So the real question was never “how accurate is this.” Perfect accuracy isn’t on the table, for a recruiter or an algorithm. The question that actually matters is what happens after the tool gets it wrong. Does that mistake get caught, or does it disappear.

That’s a workflow question, not a model question.

The two mistakes AI screening can make

Statisticians have names for the two ways any test can be wrong, and they map cleanly onto hiring.

A false positive is a candidate who ranks well, gets interviewed, maybe gets hired, and turns out to be a mismatch. The resume looked strong, the one-way interview answers sounded right, but the job itself reveals what the screening missed.

A false negative is the opposite: a candidate who would’ve done the job well, but the ranking put them at the bottom, or a resume-parsing gap knocked their score down, and nobody ever gave them a second look.

Picture a role that pulls 240 applications, the kind of pile a front-desk or coordinator posting draws in a week. AI ranks all 240 against the criteria you set. You realistically have time to seriously look at the top 15 to 20.

Somewhere past rank 100, both kinds of mistakes are sitting in that list. A false positive at rank 8 is about to cost you an interview. A false negative at rank 140 is about to cost you a hire you’ll never know you missed.

Why this isn’t the same as the bias question

It’s worth separating this from a related but different problem. Documented bias in AI resume screening is about a ranking that’s systematically unfair to a group of candidates. False positives and false negatives happen even in a system with zero bias, because ranking hundreds of strangers from incomplete information is just a hard prediction problem.

Bias makes the errors uneven across groups. This is about the errors themselves, and what happens to them once they exist.

The candidate you interview and the one you never see

Here’s the asymmetry that actually matters.

The mistake you can point to

A false positive is self-correcting, on a delay. You interview the person, or you hire them and find out in the first month.

It costs you time, and sometimes real money. Industry survey data from CareerBuilder puts the average cost of a bad hire at $14,900, and nearly three in four employers said they’d made one. That’s a real cost. But it’s a cost you notice, because you lived through it.

The mistake you never see

A false negative doesn’t work that way. If a candidate never gets surfaced, you never meet them, never find out what you missed, and never get a signal that anything went wrong. The search just looks normal. You filled the role, the process felt fine, and the person who would’ve been your best hire this year is working somewhere else, and you have no idea they ever applied.

On a small team, that math cuts the wrong way. A bad hire is a hole you feel every day for months. A missed great hire is a hole you never notice, because there’s nothing there to point at.

It’s not a smaller loss. It’s a silent one, and silent losses don’t get fixed, because nobody’s asking anyone to fix them.

There’s a second version of this same problem: speed. A candidate ranked low today because your criteria weren’t quite dialed in yet might have taken another offer by the time anyone corrects it. Slow screening loses good candidates even when the ranking eventually gets fixed, because the person doesn’t wait around for round two.

A more accurate model still can’t find someone it never showed you

The obvious response is “so make the AI more accurate.” It’s a reasonable instinct, and it’s incomplete in a way that matters.

A better model doesn’t rescue what’s already gone

Better accuracy shrinks how often either mistake happens. It doesn’t change what happens after one does. If your screening tool auto-rejects anyone below a cutoff score, a small improvement in the model’s accuracy doesn’t rescue the false negatives that already got cut.

They’re gone before the improvement ships. You’d need the model to be right about a specific stranger, every time, which no honest vendor can promise you.

A human reviewer can’t catch what they never see

This is also where “just add a human reviewer” runs into its own limit, not because people are careless, but because of what they’re shown. A University of Washington study had 528 people make hiring calls alongside AI recommendations. When the AI’s judgment was wrong but not obviously so, people followed it about 90% of the time anyway.

A reviewer who only sees a score, or only sees the candidates who cleared a cutoff, has nothing to catch a false negative with. The mistake isn’t in front of them. It’s already been removed.

So neither fix, on its own, solves the actual problem. A better model reduces how often you’re wrong. It does nothing for the wrongness you can’t see. The only thing that reaches a false negative after the fact is a workflow where it’s still there to be found.

Every candidate stays in the list, even the one ranked 140th

This is the part of the design that matters more than the score itself. Truffle, the candidate screening platform we build, ranks every candidate against the criteria you set for the role, whether that’s resumes, a one-way interview, an assessment, or all three. Nothing gets removed from your pipeline based on that ranking. Nobody’s status changes until you take an action on them.

That design choice is aimed directly at the asymmetry.

What catches a false positive

A false positive, you’ll likely catch anyway, because the review workflow shows you the reasoning behind a high score, not just the number, so you can spot one that doesn’t hold up before you ever pick up the phone.

What catches a false negative

A false negative needs something different: the candidate has to still exist somewhere you can find them. In practice, that’s the “For Review” queue holding all 240 candidates, not the top 20. It’s being able to open candidate 140’s profile and read their actual answers, not just trust the number next to their name.

Maybe a referral mentions someone you don’t remember seeing. Maybe you’re staring at your top 10 and something feels thin, so you scroll further than you planned to. Every one of those moments only works because the candidate is still there.

None of this promises you’ll always find the false negative. Nobody can promise that; it’s a real limitation, not a marketing line. What it promises is that the option to look is never taken away from you by the software.

AI surfaces the ranking and the reasoning behind it. You decide who’s worth a second look, including the people who ranked lower than you expected.

Ask what happens to the candidates it’s wrong about

“How accurate is your AI” is the question a vendor wants you to ask, because it’s the one they can answer with a number that sounds reassuring. It’s not the question that protects you.

The question that actually protects you

That question is narrower: when this tool is wrong about someone, what happens to them? If the honest answer involves an auto-reject, a hard cutoff, or a pipeline that quietly drops anyone below a threshold, you’ve built a system where your false negatives are gone before you’d ever know to look for them. If the answer is that every candidate stays reviewable and nothing advances or disappears without you saying so, the tool’s mistakes stay recoverable, which is the most any screening process, human or AI, can honestly promise.

Pew Research found Americans oppose letting AI make the final call on who gets hired by a 71 to 7 percent margin. Most people asked that question probably aren’t picturing a false negative specifically. But that’s the instinct underneath it: nobody wants the final say quietly handed to a system that can be wrong about them and never tell anyone.

That’s the real case for keeping a human on every call, and it’s a sturdier one than “compliance requires it.” Underneath the legal question is a simpler one: which kind of wrong can you afford, and which one can you not afford to lose sight of.

Ready to see what a ranked list that never auto-rejects anyone looks like on your own candidates? Truffle’s plans start at $49 a month, with a 7-day free trial and no credit card required. Set your criteria, upload a role’s worth of resumes, and see who’s still in your list, all the way to the bottom.

Frequently asked questions about false positives and false negatives in AI screening

What’s the difference between a false positive and a false negative in candidate screening?

A false positive is a candidate the screening ranked well who turns out not to fit the role once you actually interview or hire them. A false negative is a candidate who would’ve been a strong fit but got ranked low or missed entirely, so you never found out. Both are normal outcomes of ranking people from incomplete information, not signs that a tool is broken.

Can AI screening eliminate false negatives?

No. No honest AI system can promise it will never rank a strong candidate too low, the same way no recruiter can promise they’ll never overlook someone. What a well-built tool can do is keep every candidate visible and reviewable instead of removing anyone automatically, so a false negative is still findable if you go looking.

Does Truffle auto-reject candidates who score low?

No. Every candidate sits in your queue with their match score and the reasoning behind it until you personally advance, hold, or reject them. Nothing is removed from your pipeline by the ranking itself.

Is a false positive or a false negative worse for a small business?

They cost you differently. A false positive shows up as a bad hire you can measure: the broad industry estimate for a bad hire runs around $14,900. A false negative rarely gets measured at all, because you never meet the person you missed. That makes it easy to underweight and expensive to ignore.