Four ways an AI screener misleads you, and the one question that catches all of them

It ranks the wording rather than the work, hides what it could not match, has no sense of time, and cannot be asked why. A checklist you can use on any vendor, including us.

2026-09-294 min

In 2026 the hiring loop closed on itself. Candidates generate applications with AI, auto-apply tools send hundreds a day, and on the other side an AI screener ranks them. Noise ranking noise. The screening layer was supposed to be the fix for the volume; instead it became the reason the volume stopped carrying information.

Here are four specific ways these tools mislead the person reading their output. None of them require bad faith from the vendor. All of them are invisible unless you go looking.

1. It ranks the wording, not the work

A screener trained on documents learns what a good document looks like. The candidate who writes "spearheaded the migration of a high-throughput billing platform" outranks the one who writes "moved billing to Postgres, 2023" — although the second sentence contains a verifiable fact and the first contains an adjective.

This was always true of human screening too. What changed is that writing the first sentence now costs nothing: the model that reads it also writes it. When the cost of a signal goes to zero, the signal goes with it.

What to do instead: rank on claims that carry a date and a source, not on phrasing. An adjective is not a claim. "I ran the weekly client call for two years" is.

2. It hides what it could not match

Ask a vendor what happens to a requirement no candidate meets. Many systems quietly drop it, because dropping it makes the demo look better — every candidate scores high, the tool looks confident.

The result is a number that cannot be compared with any other number. Six of ten requirements met becomes "60%" in an honest system and "100%" in one that counted only the six. You cannot tell which you are looking at from the number.

What to do instead: demand the requirement-by-requirement list behind any percentage. If the tool cannot produce one, the percentage is a guess with a decimal point. We wrote up how ours is computed, including the version of the query that had this exact bug.

3. It has no sense of time

Python, 2014 and Python, last week are the same string. To a screener working from a document, they are usually the same signal — and a document rarely says which it is, because a CV lists employers in order and leaves the skills in a block at the bottom with no dates at all.

Time matters in both directions. A skill from eight years ago may be stale. A skill from last month may be the reason this person is exactly right for a role nobody else has done yet. A system that cannot tell them apart is throwing away the most decision-relevant thing it has.

What to do instead: require a time on every claim, and accept unknown as an answer. Our agent asks rather than guessing, because a confidently wrong date is worse than an honest gap — and an age is computed from a birth date when it is read, never stored as a number that goes stale on its own.

4. It cannot be asked why

This is the one that will stop being optional. From August 2026 the EU AI Act treats recruitment AI as high-risk, which brings transparency, human oversight and record-keeping with it. A score with no reasons attached meets none of the three.

It is also simply bad practice. A rejection you cannot explain is a rejection you cannot defend — to the candidate, to your own hiring manager, or to anyone who asks six months later why the shortlist looked the way it did.

What to do instead: five questions to ask any vendor, with our answers to all five.

The one question

If you remember nothing else from this article:

For this candidate, show me each requirement, whether it was met, by what evidence, and in the candidate's own words.

A tool that can answer it is a tool whose output you can defend. A tool that cannot is asking you to trust it, which is a different product from the one it is selling you.

Ask us the same question. The answer is a screen in the product, not a slide.