Anti-bias in recruiting - what actually helps, what just looks good

One - structured interviews. Double the predictive validity vs unstructured (Schmidt+Hunter 1998). Any structured scoring process reduces the weight of implicit preferences. The single largest lever.

Two - write requirements profiles before publishing the role. If you 'adjust' requirements after reading ten applications, that's real-time bias amplification. Write the profile, lock it, publish.

Three - multiple independent voices per application. Each interviewer scores before discussion. Prevents mood pile-on where one loud sceptic drags everyone else's score down.

Four - blind first-round screening. Hide photo, name, date of birth from the first screening view. For structured scoring these aren't relevant factors anyway - this just removes temptation. Effect measured in several studies (Goldin+Rouse 2000 for orchestras).

Five - veto rights with reason, not consensus mandate. Three clear yeses and one no without basis, consensus wins. With basis, veto wins. Protects both against 'we just like them' and 'one person can block'.

Six - audit your own funnel. After three months check score distribution by gender and origin. If your top-20% systematically under-represents a group, you have a problem in the requirements profile or KI config. KI BMS surfaces this in Reports.

One - diversity statement on the careers page without process change. Studies show: the statement alone leads to fewer applications from under-represented groups, because they read it as a 'quota hint'. Use the statement only when the process behind it has actually changed.

Two - 'diversity training' once a year for hiring managers. Scientific evidence: short trainings have no measurable effect on behaviour even two weeks later, let alone a year.

Three - gender-neutral language in the job posting without other change. Pure language correction without structural adjustment shifts applicant demographics only marginally. Necessary but far from sufficient.

Structured scorecards are first-class. KI pre-sort exposes scoring rationale - if the KI learns a systematic pattern unrelated to the requirements profile, it shows up in the reasoning. Retention + anonymisation are built-in. Per-role score-distribution audits live in Reports.