A hiring tool that does not publish its fairness audit is asking you to trust it on faith. We do not think that is a defensible position for an AI-native vendor in 2026. This post is the audit, the methodology, and the cadence, in full.
The 4/5ths rule (originally from the US Uniform Guidelines on Employee Selection Procedures, 1978) says: the selection rate for any protected group should be at least 80% of the selection rate for the highest-selected group. If white candidates are selected at 30% and Black candidates are selected at 18%, the ratio is 0.60. That is a 4/5ths failure and a disparate impact flag.
We apply the 4/5ths rule per protected group, per role family, per stage of the pipeline. Not as a single sitewide number. A sitewide ratio can be 0.95 and hide a 0.50 failure on backend engineering roles in DACH. The granularity matters.
We do not infer protected characteristics from voice, video, or written content. The only data we use is the optional, self-reported demographic data captured at the start of the candidate experience, on a separate consent screen, with a "prefer not to say" option for every field.
We run the audit weekly, automatically, across all candidates who completed at least the role-fit assessment. The results land in two places:
When a per-role-family ratio drops below 0.80, the role is flagged. The flag triggers two actions: a human-review pause on automated advancement decisions, and an engineering-team investigation of the rubric calibration. We have hit this flag twice in beta, both times on small-sample roles (under 30 candidates). The pause held until the role accumulated enough sample to retest.
On /trust/fairness, every quarter:
We do not publish per-customer fairness reports without that customer's consent. We do publish the methodology, the model card for each role family (on /trust/model-cards), and the human-oversight surface (on /trust/responsible-ai).
We will not run a "blind hiring" mode that hides protected characteristics from the model. That is theatre. The model does not see protected characteristics in the first place: the input is the candidate's answers, scored against a rubric. The fairness audit measures whether the rubric is biased, not whether the model can guess somebody's gender from their voice.
We will not score on tone of voice, accent, video appearance, typing speed, or any signal that does not directly bear on the competency being assessed. The model card for each role family lists exactly which signals are weighted.
We will not auto-reject based on the model score alone. Every automated decision has a human-review path, surfaced to the candidate at the point of decision. EU AI Act Article 22 compliance is built in, not bolted on.
SOC 2 Type II is in progress with Prescient Assurance and Drata; the report ships with V1 in Q4 2026.
The people building Picked. Method posts, model cards, fairness audits, product opinions. Edited and signed off by the engineering and research leads.