AI Bias Testing For Candidate Scoring

Nova tests AI-assisted candidate scoring for adverse-impact signals using a synthetic cohort with demographic-linked signal variation. The public summary below shows one product-level test for a representative role.

Read Methodology
500
Profiles Tested
5
Test Slices Clear
5
Demographics Analyzed
0
Groups Below Concern Threshold

Public Bias Test Summary

This synthetic run checked whether generated profile groups received materially different scoring outcomes across demographic-linked signals.

Overall Bias Evaluation Summary

Clear
Sex
Clear
Race/Ethnicity
Clear
Race x Sex
Clear
Age
Clear
Disability

Sex

Clear

No adverse-impact signal in this synthetic run

Male
91.9% selected0.99 ratio
Female
92.6% selected1.00 ratio

Race/Ethnicity

Clear

No adverse-impact signal in this synthetic run

Asian
92.3% selected0.97 ratio
Hispanic
94.8% selected1.00 ratio
White
91.3% selected0.96 ratio
Black
90.2% selected0.95 ratio

Age

Clear

No adverse-impact signal in this synthetic run

Over 40
100% selected1.00 ratio
Under 40
84.8% selected0.85 ratio

Disability

Clear

No adverse-impact signal in this synthetic run

No disability
93.8% selected1.00 ratio
Has disability
90.5% selected0.96 ratio

Evaluation Methodology

Disparate Impact Analysis

Compares selection rates across demographic groups to identify adverse-impact signals using the Four-Fifths Rule benchmark.

Synthetic Resume Testing

We score synthetic resumes with broadly comparable role-relevant qualifications and varied demographic indicators, such as names, locations, schools, and age-related experience patterns.

Clear

Impact ratio >= 80%
No material signal

Review

60% <= Impact ratio < 80%
Human review needed

Concern

Impact ratio < 60%
Concern signal

Intersectional Analysis

Clear

Analysis of overlapping demographic factors did not show an adverse-impact signal in this synthetic run

Race/EthnicitySexSamplesSelection RateImpact Ratio
AsianMale5887.9%0.91
HispanicMale6196.7%1.00
HispanicFemale7393.2%0.96
WhiteMale6792.5%0.96
WhiteFemale6090%0.93
BlackMale7290.3%0.93
BlackFemale5090%0.93
AsianFemale5996.6%1.00

Testing Details

Evaluation Standards

  • Disparate Impact Analysis (Four-Fifths Rule)
  • Intersectional signal review
  • Demographic signal analysis
  • Threshold impact-ratio review

What This Run Shows

  • 500 synthetic profiles for one representative role
  • Demographic-linked signal variation
  • Group and intersectional impact-ratio review
  • Public methodology documentation

Report Details

Run Date
May 26, 2025
Scoring Range
1-10
Scoring Path
Nova evaluation harness
Sample Generation
Synthetic Profiles

Review Nova’s Scoring Approach

See how Nova supports responsible AI-assisted hiring with criteria-based scoring, explainable outputs, human review, and bias testing.