Semantic AI Recruiting

01Where AI genuinely helps

Three categories of AI recruiting tooling have reasonably solid track records. Semantic search and matching — understanding that "led a cross-functional engineering team" and "managed a multidisciplinary technical group" describe the same experience despite sharing almost no vocabulary — is a genuine advance over keyword Boolean search, which has frustrated both recruiters and candidates for two decades. Scheduling automation removes a specific, well-bounded coordination problem. And generative drafting, used as a starting point a human edits rather than a final output, speeds up the unglamorous work of writing job descriptions and outreach messages.

What these three have in common: the AI's output is either easily verified (a scheduled meeting either happened or it didn't) or explicitly treated as a draft a human will review. The failure mode below shows up when that verification step disappears.

02Where the category has a documented failure mode

The riskiest use of AI in hiring is asking a model to directly produce a hiring-relevant score or ranking — "rate this candidate 1 to 100" — and then acting on that number with limited human review. Large language models are next-token predictors; asked for a score, they will confidently produce one, and that score will not necessarily be the same number if you ask again with the same inputs. This isn't a bug particular to any one vendor's implementation — it's intrinsic to how generative models work, and it means a naively-built "AI resume screener" is closer to a random number generator with a persuasive explanation attached than to a repeatable measurement.

The consequence in hiring specifically is serious: a score that can't be reproduced can't be defended if a rejected candidate — or a regulator — asks why. Several jurisdictions now require employers using automated hiring tools to be able to explain and audit the criteria (New York City's Local Law 144 and the EU AI Act's provisions on high-risk AI systems are the two most-discussed examples as of this writing; we're not lawyers and this isn't legal advice, but the regulatory direction of travel is unambiguous). An unreproducible score is close to indefensible under that kind of scrutiny by construction, regardless of how accurate it happens to be on average.

03The architectural fix: separate extraction from scoring

Expertini's approach, detailed in full on the dedicated Candidate Match Score page, is to give the AI exactly one job — reading comprehension — and never let it touch the arithmetic. The model reads an anonymised CV against a job description and returns a structured list of which competencies it found evidence for, each tagged with a short citation of that evidence. It does not return a score, a ranking, or a recommendation. A deterministic formula, running as ordinary code outside the model, converts that structured extraction into the final number.

This split matters because it makes each half independently checkable. You can inspect what the model actually found (the extraction) and separately verify the arithmetic (the formula) — and because the formula is fixed, identical inputs always produce an identical score, on any date, run by any recruiter.

The published paper tests exactly this on a deliberately hard case — a strong engineer whose CV shares almost no vocabulary with the job description, applying to a role with one mandatory requirement he lacks:

The paper's comparative test: the same candidate scores roughly 32 percent under keyword matching, 92 from an unconstrained LLM, and 75 from the hybrid framework — the only architecture that both surfaces him for review and flags the genuine blocker (Syed, 2026, Fig. 2)

You can download the full paper as a PDF and check the methodology — and its acknowledged limitations — yourself.

04What this means in practice for a hiring team

In day-to-day use, the difference is mostly invisible until something goes wrong — which is the point. A recruiter ranking fifty applicants sees a Kanban board and a score, the same experience a keyword ATS or a black-box AI tool would offer. The difference surfaces when a candidate disputes a rejection, when a works council or legal team asks how the tool works, or eighteen months later when someone tries to reproduce a decision from an audit log. In the deterministic architecture, all three of those moments have a real answer. In a stochastic one, they don't.