Prior biased policing leaks into race-neutral variables like conviction history
Evidence
In a criminal justice context, a history of treating one demographic group differently creates redundant encodings all over the place. Policing a minority neighborhood more aggressively than others, for instance, means that all of a sudden something as seemingly neutral as length-of-rap-sheet, i.e., number of previous convictions, can become a redundant encoding of race. — Brian Christian, The Alignment Problem, 296
Categorization
- XREF: Directly anticipates the feedback-loop discussion later in the chapter (Lum and Isaac), and echoes the book’s theme of systems sculpting the reality they predict. — connects to: Partial debiasing can make bias problem worse by hiding it — Brian Christian, The Alignment Problem, 210; ProPublica tested COMPAS recidivism predictions against actual reoffense outcomes for bias; Parole outcomes vary predictably by measurable background factors — Brian Christian, The Alignment Problem, 244 (+8 more)
- corroborated: 2 independent passages support this claim — Brian Christian, The Alignment Problem, 296
Broader Connections
- ProPublica tested COMPAS recidivism predictions against actual reoffense outcomes for bias
#expands(0.60) - Machine-learning bias often stems from unrepresentative training data
#expands(0.55) - COMPAS errors fall unequally on Black and White defendants
#expands(0.55) — Brian Christian, The Alignment Problem, 279 - Religious identity can be racialized when signifiers trigger profiling
#expands(0.55) - US law requires proof of racist intent, not just disparate impact, to win
#complicates(0.50) — Harriet A. Washington, A Terrible Thing to Waste, 470 - Parole outcomes vary predictably by measurable background factors
#complicates(0.40) — Brian Christian, The Alignment Problem, 244