Every AI Screen Has a False-Negative Rate. Most Vendors Won’t Tell You Theirs.

Across our pilot campaigns, spanning 1,099 completed voice interviews, the AI auto-rejected 60%+ of candidates before a recruiter opened a file. That number gets asked about constantly, usually as “is 60% too harsh?” It’s the wrong question to lead with. The harder number, the one almost nobody in AI screening publishes, is how many of that 60% were good candidates the system got wrong.

Every screening system, human or AI, has a false-negative rate: qualified people it turns away. A tired human recruiter skimming a resume pile late in the evening has one. So does a voice AI scoring communication in a 15-minute conversation. The difference is that human error is assumed and mostly invisible, while an AI system’s errors are logged, timestamped, and technically knowable — which makes it strange that so few vendors, including ourselves before we started pushing on this, actually report the number.

What a false negative looks like here

A false negative isn’t a candidate who got a fair shot and lost. It’s a candidate who could do the job and got screened out anyway, because of a bad connection, a nervous first two minutes, an accent the model handled less gracefully than it should have, or a question that didn’t map cleanly onto how that candidate would actually explain their experience. It’s a different failure mode from bias in the legal sense, and a different one again from the auto-reject rate simply reflecting a lopsided applicant pool. It’s the system being wrong about someone it should have advanced.

That distinction matters because the three get conflated constantly. A high auto-reject rate is not evidence of a high false-negative rate — most rejected candidates in a 3,000-applicant campus drive genuinely aren’t a fit for the role. But some slice of that 60%+ is the system’s mistake, not the candidate’s shortfall, and that slice is the number that should keep a vendor honest.

Why the number stays invisible

Part of it is that measuring a false-negative rate is genuinely harder than measuring throughput. You need to go back and re-evaluate rejected candidates by a method you trust more than the one that rejected them, which is expensive and slow, and there’s no commercial incentive to volunteer a number that makes your product look worse. Vendor decks lead with time saved and cost per interview because those numbers only ever look good. A false-negative rate is the one metric where publishing anything above zero is an admission, which is exactly why it’s worth asking for.

Siddarth, whose pilot feedback shaped a lot of how we think about this, put it in one line early on: don’t reject good profiles — prioritize communication first, not keyword or credential density. That’s an operating principle, not a guarantee. We built our classification around it specifically because a system tuned to reject on communication ability, not resume proxies, has fewer ways to wrongly filter out someone who could actually do the job. It doesn’t get the number to zero. Nothing does.

The honest version

We don’t have a clean published false-negative rate to hand you, and we’d trust a vendor less, not more, if they claimed a precise one this early in the category’s life. What we do have is a practice: every pilot customer is told to pull a sample of the rejected batch each drive and spot-check it by hand, specifically looking for candidates the system got wrong, not just candidates who were correctly screened out. What a 60%+ auto-reject rate actually means covers the shape of that number; this is about the part of it that deserves scrutiny rather than acceptance.

SHRM’s 2025 AI-in-HR survey found that 88% of HR leaders already see AI screening as a compliance risk. Most of that concern is framed around bias and disclosure. Less of it is framed around plain accuracy — whether the tool is simply wrong about people sometimes, in ways that have nothing to do with protected categories and everything to do with a 15-minute conversation being an imperfect instrument. Both concerns deserve the same scrutiny, and right now only one of them gets asked about in vendor evaluations.

The take

A vendor who can’t tell you their false-negative rate probably hasn’t measured it, and a vendor who claims a suspiciously precise one probably hasn’t measured it rigorously either. The honest middle ground is a vendor who tells you what they check for, how often, and what happens when they find a mistake. Ask that question in your next vendor evaluation before you ask about pricing. It tells you more about whether the tool respects the candidates it rejects than any accuracy claim on a sales deck.

Want to see the spot-check process we run on our own rejected batches? Get in touch and we’ll walk you through it against a sample from your own funnel.