Recruitment AI review urges broader evaluation rules

A systematized narrative review on arXiv argues that AI in recruitment has shifted from matching candidates and vacancies toward multi-stage workflows that retrieve evidence, compare applicants and can support or carry out actions. The authors say they organised 40 representative works, alongside industrial and legal sources, using a search and coding process updated through late July 2026 with targeted updates into early September.

Their central claim is that evaluation methods have lagged behind the systems themselves. The review traces a move from similarity scoring to reciprocal suitability, from a single model to a compound workflow, and from offline prediction toward evaluation tied to evidence and productivity.

What the review found

The authors say persistent gaps arise because behavioural labels can mix up exposure, preference and qualification; private or synthetic datasets can limit external validity; and final output scores can hide failures inside a pipeline. Within their coded set, they report that privacy was not directly evaluated and that no row jointly evaluated utility, fairness, privacy and security.

Those findings come with explicit limits. The paper says it is not a prevalence estimate of the whole field and that its observations describe the coded set rather than every hiring product or study. It is also a preprint, not peer-reviewed evidence. Even so, for employers and vendors using AI in recruitment, the paper offers a more disciplined way to ask what a system actually proves, as opposed to what a dashboard implies.