AI-Supported Candidate Screening: What Helps, What Doesn’t

What “automated screening” actually covers

The label hides very different systems: keyword rules, CV parsing, semantic matching against a competency profile, gamified tests, video analysis. Only the first three matter for most SMEs. Hunkenschroer and Luetge (2022) map AI recruiting along the process chain: the closer the system sits to a rejection, the higher the ethical and legal bar [1].

• Pre-sort: rank against defined competencies — recruiters decide who advances
• Knock-out rules: hard filters (e.g. a mandatory licence) — only if the rule is job-related and documented
• Not the same thing: auto-reject with no human review — that is the Article 22 case

Vendor claims vs documented practice

Raghavan, Barocas, Kleinberg, and Levy (2020) reviewed public materials from vendors of algorithmic pre-employment assessments [2]. The pattern: many promise bias reduction; few disclose what the model was trained on, what it predicts (historical performance? tenure? “culture fit”?), or which fairness metric they even test. Without that, “our AI is fair” is marketing, not evidence.

• Inspect the prediction target: learning from past hires also learns their bias
• Do not “remove” protected class while leaving proxies (postcode, school name, CV gaps) unchecked
• Ask for validation against job performance — not accuracy on a vendor dataset
• Technical de-biasing can conflict with anti-discrimination law if it treats groups differently [2]

Algorithmic discrimination is not an edge case

Köchling and Wehner (2020) systematically reviewed 36 studies on algorithmic decision-making in HR recruitment and development [3]. Two layers show up separately: objective fairness (differential validity and prediction across groups) and perceived fairness (how applicants experience the process). Either can damage candidate experience and legal posture — even if the model “performs” internally. Barocas and Selbst (2016) show disparate impact often arrives through seemingly neutral features, not explicit discrimination [4].

• Historical labels are not ground truth for “a good hire”
• A model that weights alma mater or employer prestige can proxy class and origin
• Perceived opacity lowers acceptance even when rankings are internally consistent [3]

Validate it like any other selection procedure

AI screening is a personnel selection procedure. The SIOP Principles (2018) apply whether a human or a model produces the score: construct, criterion, local validation, documentation, ongoing monitoring [5]. A tool that “only ranks” still needs a job-related reason for that score.

The EU hard stop: support, don’t decide

GDPR Article 22 prohibits decisions based solely on automated processing that significantly affect a person — a rejection in hiring counts. The EU AI Act treats AI used in employment and worker management as high-risk. In practice: scores may prioritise; humans must own and be able to explain the decision. That is not red tape for its own sake. It matches the research: unattended models scale error [2][3].

• No fully automated rejection from a ranking
• Privacy notice: that AI is used, for what, on which legal basis
• Explainability: which features moved the score, in language a candidate can use
• Human-in-the-loop must be real — rubber-stamping a ranking is not review

When AI screening is still the right call

The alternative to a bad model is rarely “no tool”. It is an overloaded team reading CVs in inbox order — also biased. Screening helps when the role profile leads and the model follows.

• Write competencies and knock-outs first (see the structured-interview guide)
• Match against that profile, not against “people we hired before”
• Show recruiters the rationale (which skills hit), not only a number
• Sample: periodically compare rankings to a human shortlist and explain gaps
• Log overrides — that is your audit trail, not a defect

Conclusion

AI screening is an accelerator, not a judge. The literature on vendor claims, algorithmic discrimination, and validation says the same thing as GDPR and the AI Act: disclose criteria, keep humans deciding, measure fairness instead of asserting it. Done that way, you keep the time gain without inheriting a black-box liability.