Bias risk assessment

Last updated: 14 July 2026

Monitoring program adopted 14 July 2026 — first cycle scheduled

Most bias statements in this industry are one sentence long and unfalsifiable. This one is a register: the specific ways an instrument like Stori's could produce biased outcomes, named plainly, with what the design does about each — and an honest account of what is not yet testable and why.

The risk register

Historical bias absorbed as ground truth

The classic failure: a model trained on past hiring outcomes learns the biases in those outcomes. Structurally absent here — Stori does not train on hired / not-hired outcomes, so there is no rejection pile whose patterns could be learned.

Speech delivery leaking into scores

The sharpest risk for any interview-based tool: fluency, accent, hesitation, or atypical prosody depressing scores — a disability and language-background artifact, not capability. Design response: evaluation reads transcript substance only; delivery carries no signal, positive or negative, as a written rule in every evaluation prompt. Delivery-invariance testing (same facts, varied speech style) checks the rule holds in practice.

Proxies for protected characteristics

Signals like ZIP code, school prestige, or graduation year standing in for race, class, or age. Design response: protected characteristics are not evaluation criteria; ZIP is not used as a proxy; location is used only as job-relevant proximity to a stated work location; scoring inputs are enumerated and audited against a proxy list.

Personality as a gate

Trait instruments have a long history of screening out atypical profiles. Design response: the self-report never contributes to ranking, prioritization, filtering, or exclusion — it is candidate self-insight and employer context only.

Unequal abstention

A subtler failure: the system saying “insufficient signal” more often for some groups, quietly shrinking their evidence. Design response: abstention-rate parity is part of the monitoring program, alongside score parity.

Modality disadvantage

A voice-first interview could disadvantage candidates who cannot complete one comfortably. Design response: adjustments (extra time, restarts, alternative format) on request before the interview; because scoring reads the transcript, a text-based completion preserves the instrument — and text-vs-voice parity is a standing test.

What we can test today — and what we cannot yet

Runnable now, without demographic data: delivery-invariance, text-vs-voice parity, abstention parity across speech styles, length-robustness, and the scoring-input proxy audit. These target the mechanisms above directly.
Not yet computable: group-level statistics by race, sex, age, disability, or veteran status — because Stori does not collect demographic data from candidates.
What we refuse to do about that: infer demographics from names, photos, voices, or geography. Inference is itself a protected-characteristic estimation with error concentrated in minority groups, and data-protection regulators treat inferred characteristics as the sensitive data itself.
The path: a voluntary, clearly optional self-identification step, stored separately from all scoring inputs, invisible to employers and to the scoring pipeline, used solely for aggregate bias monitoring. When group-level results exist, their dated summaries will be reflected here.

On audits

Where an employer's use of Stori makes it subject to an independent bias-audit requirement, Stori provides the structured, per-candidate data an auditor needs to compute selection rates and impact ratios. And a standard worth stating: a selection-rate parity audit is a floor, not fairness — random scoring passes it perfectly. The tests above are aimed at the level where an instrument like this one actually fails or holds: whether the system ever mistakes how someone speaks, or who they are, for what they have done.

This document describes how Stori is designed. It is not legal advice and does not determine whether any specific employer deployment is lawful. Questions from legal or procurement teams: support@onestori.com.