ProPublica publishes the COMPAS investigation

ProPublica published its COMPAS investigation on May 23, 2016.^[1] The chronology dates reporting, rather than a software launch, an individual arrest or a judicial finding.

What exactly is being compared?

Chouldechova’s analysis distinguishes predictive parity from error-rate balance. Its hypothetical penalty model is not a causal audit of court decisions.^[2]

Kleinberg and colleagues provide a separate formal framework and identify future arrest as the risk tool’s outcome proxy. That proxy is not a direct measurement of all offending.^[3]

These questions make the investigation a useful entry into Fairness Tradeoffs. Ask which population enters each denominator, how the outcome is recorded, what the score means and how it is used. A dispute over a metric and a dispute over a decision’s justification require different evidence.

A reading sequence

Start with the original investigation and a reporter’s published interview excerpts, then compare the two research papers. Alexandra Chouldechova connects the empirical reanalysis with metric definitions. The incidents guide contrasts this investigation with chatbot failures and an identification-led arrest.

Review here establishes publication identity and selected formal/reanalysis arguments. Original vendor responses, underlying criminal records, code and downstream sentencing effects were not independently audited. The entry does not adjudicate every contested allegation.