Stability Audit

This repository contains code used to audit the stability of personality predictions made by two algorithmic hiring systems, Humantic AI and Crystal. This codebase supports the 2021 manuscript entitled "External Stability Auditing to Test the Validity of Personality Prediction in AI Hiring," authored by Alene K. Rhea, Kelsey Markey, Lauren D'Arinzo, Hilke Schellmann, Mona Sloane, Paul Squires, and Julia Stoyanovich.

Code

The Jupyter notebook analysis.ipynb reads in the survey and system output data, and performs all stability analysis. The notebook begins with a demographic summarization, and then estimates stability metrics for each facet experiment as described in the manuscript.

Spearman's rank correlation is used to measure rank-order stability, two-tailed Wilcoxon signed rank testing is used to measure locational stability, and normalized L1 distance is used to measure total change across each facet. Medians of each facet treatment are estimated as well. Results are saved to the results directory, organized by metric and by system (Humantic AI and Crystal). Subgroup analysis is performed for rank-order stability and total change. Highlighting is employed to indicate correlations below 0.95 and 0.90, and Wilcoxon p-values below the Bonferroni and Benjamini-Hochberg corrected thresholds. Scatterplots are produced to compare the outputs from each pair of facet treatments. Boxplots illustrate total change. Boxplots comparing relevant subgroup analysis for each facet are produced as well.

Data

Survey

Anonymized survey results are saved in data/survey.csv. Columns described in the table below.

Humantic AI and Crystal Output

Output from Humantic AI and Crystal is saved in the data directory. Each run is saved as a CSV and is named with its Run ID. Tables 3 and 4 in the manuscript (reproduced below) provide details of each run. Each file contains one row for each submitted input. Participant_ID provides a unique key, and output_success is a Boolean flag indicating that the system successfully produced output from the given input. Wherever output_success is true, there will be numeric predictions for each trait. Crystal results contain predictions for DiSC traits, and Humantic AI results contain predictions for DiSC traits and Big Five traits.