ProPublica tested COMPAS recidivism predictions against actual reoffense outcomes for bias

Evidence

If a facial-recognition system is wildly inaccurate for people of one race or gender but not another, or if someone is denied bail because of a statistical model that has never been audited and that no one in the courtroom—including the judge, attorneys, and defendant—understands, this is a problem. — Brian Christian, The Alignment Problem, 61 (#exemplifies, via linked claim)

Categorization

• corroborated: 2 independent passages support this claim
• XREF: connects to: Machine-learning bias often stems from unrepresentative training data; Science only recently began predicting parolee behavior after release; ProPublica used 2014 recidivism outcomes to test the risk model’s accuracy and bias (+6 more)

Related

• ProPublica used 2014 recidivism outcomes to test the risk model’s accuracy and bias (similarity 0.85)
• Science only recently began predicting parolee behavior after release (similarity 0.81)
• Untested risk models can detain people on factors having no real predictive value (similarity 0.81)
• Model opacity makes algorithmic decisions unaccountable in courtrooms (similarity 0.80)
• COMPAS errors fall unequally on Black and White defendants (similarity 0.80)
• Machine-learning bias often stems from unrepresentative training data (similarity 0.80)
• Prior biased policing leaks into race-neutral variables like conviction history (similarity 0.79)

Broader Connections

• Model opacity makes algorithmic decisions unaccountable in courtrooms #exemplifies (0.60) — Brian Christian, The Alignment Problem, 61