Dataset Bias Auditor
Upload a CSV dataset or use a sample to analyze bias and representation imbalances.
Same score, unequal mistakes
The recidivism sample's risk score never sees race. Yet among people who did not go on to reoffend, it flags Black defendants as high risk far more often than White defendants. Selection rates hide this; you need the confusion matrix, which needs ground truth.
Work one row out by hand first: count the people right of the line, then the shaded ones among them. Then load the Recidivism sample below and check the auditor gets the same number.
Data Source
Audit Configuration
Which outcome benefits the person. This is a value judgement, not a property of the data — you have to make it.
Data Preview
Class Distribution
Distribution of the target variable across protected attribute groups.
Group Representation
Proportion of each group in the dataset.
Correlation Heatmap
Pearson correlation between numeric features.
Fairness Metrics
Selection-rate metrics need only the decision column. Error-rate metrics also need ground truth, and are disabled without it.
Intersectional Analysis
Bias can be amplified at intersections of multiple attributes (e.g., race × gender). This analysis checks for compound effects.