DEI in Hiring

01Where bias typically enters a pipeline

Resume screening is the most-studied entry point: research going back decades has repeatedly found that identical resumes receive different callback rates when only the applicant's apparent name is varied — a pattern documented across multiple countries and industries, not a single study's artefact. Keyword-based ATS filtering compounds this by rewarding resume vocabulary that correlates with particular educational or professional backgrounds, filtering out equally qualified candidates who describe the same experience differently.

Referral-heavy sourcing tends to reproduce the demographic makeup of an existing team, since people's networks are usually more homogeneous than the broader qualified labour pool. And unstructured interviews — the classic "culture fit" conversation with no fixed rubric — carry some of the weakest evidence of validity of any hiring method commonly used, precisely because they leave the most room for an interviewer's unexamined preferences to operate.

02What anonymisation actually does and doesn't do

Before any CV reaches Expertini's AI extraction step, it passes through an eleven-pattern stripping process removing names, contact details, physical addresses and postal codes, social/professional profile URLs, and language suggestive of age, gender, nationality, or marital status. The AI performing semantic reading never has access to who the candidate is — only what they've documented having done.

This is a meaningful, mechanical reduction in exposure — not a guarantee of zero bias. Pattern-based stripping can miss an unusual name format, and writing style itself can sometimes carry weak, indirect signals a determined statistical analysis might detect even after explicit identifiers are removed. We describe the mechanism honestly rather than claim a stronger guarantee than pattern-matching can deliver.

03Evidence-based matching versus keyword matching

Separately from anonymisation, matching itself moved from keyword overlap to semantic evidence — detailed on the Talent Matching page — specifically because keyword filtering systematically disadvantages candidates who describe real experience in non-standard vocabulary: career changers, candidates from under-represented educational backgrounds, and people writing in a second language. Reading for evidence rather than string matches removes this specific, well-documented filtering bias.

04What structure adds on top of anonymisation

Anonymisation addresses what the AI sees; structured evaluation (detailed on the Structured Hiring page) addresses what happens afterward — every candidate for a role is measured against the same dimensions with the same weights, and every interview stage uses the same scoring rubric across candidates. Structure alone has meaningfully reduced bias in decades of hiring research even without any AI involved at all; combining it with anonymised, evidence-based screening compounds the effect rather than relying on either mechanism in isolation.

05Stated limitations

None of this eliminates bias entirely, and any vendor claiming their product does should be treated with real scepticism. A biased job description — one written with unnecessarily narrow requirements that exclude qualified candidates for no job-relevant reason — will still produce biased CMS scores, because the scoring faithfully measures against whatever requirements were stated. Human interview judgement, even when structured, still carries residual bias structure alone can reduce but not fully remove. And final hiring decisions remain human judgement calls the system informs but does not make.