Somewhere in your rejected pile is the candidate you were looking for. If your ATS filters CVs by keywords, that isn't a risk — it's a statistical inevitability, because keyword filtering doesn't measure whether someone can do the job. It measures whether they used your vocabulary to describe it.
This piece takes the two approaches apart mechanically: what a keyword filter actually computes, what a rubric-based AI evaluation computes instead, and why the difference decides who you meet.
What a keyword filter actually does
A keyword filter reduces a CV to a term-match test: does the document contain "Kubernetes," "stakeholder management," the exact title you searched? It has three structural failures, and none of them are fixable with better keyword lists.
It's blind to synonyms and phrasing. The engineer who wrote "orchestrated containerized deployments" doesn't match "Kubernetes." The career-changer whose transferable evidence lives under unexpected job titles doesn't match anything. The strongest people often describe their work in their own words — precisely the behavior filters punish.
It's trivially gamed. Everyone knows the game: paste the job ad's vocabulary into the CV (some in white text). Keyword filters actively select for candidates who optimize documents, against candidates who optimized for the work.
It's binary and unexplained. Match or no match — no notion of how well, no reasons, no audit trail. When a filter rejects four hundred people, nobody can say what any individual rejection was based on, which is exactly the black-box problem in its oldest form.
What rubric scoring does instead
Rubric-based evaluation inverts the computation. Instead of asking "does this document contain my words?", it asks "how well does this candidate's evidence satisfy each of my criteria?" — and modern AI can read a CV's meaning well enough to answer with reasons.
| Keyword filter | Rubric-based AI scoring | |
|---|---|---|
| Reads | Term frequency | Meaning, in context |
| Handles synonyms & phrasing | No | Yes — evidence counts however it's worded |
| Career-changers | Filtered out | Evaluated on transferable evidence |
| Gamed by buzzword stuffing | Yes | No — claims without evidence score poorly |
| Output | Pass/fail | 0–10 per criterion + written justification |
| Explains itself | No | Every score cites what it found |
| Missing information | Silent rejection | Explicit "Cannot evaluate" |
In Rubrily, every CV gets a per-criterion score with an evidence-citing justification, document-quality signals, and red flags for a human to weigh — and when a criterion has no evidence either way, the answer is "Cannot evaluate," not a quiet zero. The result is a ranking you can interrogate: click any score, read why, check it against the CV.
The asymmetry that matters: who each system rejects
Both systems reject most of a big pile — the question is on what basis. A keyword filter's false rejections cluster on your best nonstandard candidates: career-changers, people from adjacent stacks, people who write plainly instead of in job-ad dialect. Its false accepts cluster on the well-optimized mediocre. That's a double selection error pointed exactly the wrong way.
Rubric scoring's errors, when they happen, are visible ones — a justification that doesn't survive contact with the CV — which makes them catchable in review and fixable at the criterion level. An error you can see is a process improvement; an error you can't see is a lost hire, forever.
Is AI CV screening just a smarter keyword filter?
No — the computation is different in kind, not degree. A keyword filter matches terms; rubric-based AI evaluation reads the document against defined criteria and writes a justification for each score. The practical tests: synonyms and unusual phrasing still score, buzzword stuffing doesn't, and every result explains itself.
What to do with the pile you already have
One underrated feature of rubric scoring is that it's retroactive. The four hundred CVs your keyword filter processed last quarter are still sitting in your system — re-evaluating them against a real rubric routinely surfaces strong candidates the filter discarded. (Rubrily bulk-evaluates up to 30 CVs at a time and ranks the results; your "talent pool" may already contain your next hire.)
FAQ
Don't recruiters still need keywords for sourcing? Searching for candidates by keyword is fine — search is a recall tool and a human reviews the results. The failure mode is keyword filtering: unattended, binary rejection with no evaluation behind it. Use keywords to find; use rubrics to judge.
What about knockout questions — aren't those filters too? Legitimate hard requirements (work authorization, licensure, location) are fine as explicit knockout questions, because they're transparent and job-related by construction. The problem is smuggling skill assessment into term matching — skills deserve evaluation, not string comparison.
Can rubric scoring be biased? Any evaluation encodes its criteria's assumptions, which is exactly why criteria should be explicit, job-related, and human-reviewed — and why every score should carry a justification you can audit. Rubric scoring doesn't make bias impossible; it makes it inspectable, which keyword filtering never was.
Stop filtering by vocabulary. Score every CV against your rubric — with reasons — and see who was hiding in the pile. Start free →