概述
It became the central case study in algorithmic fairness after a 2016 ProPublica investigation found that its errors fell unevenly by race. The debate that followed showed that common definitions of fairness can mathematically conflict.
深入探讨
In May 2016, ProPublica published "Machine Bias." It analyzed COMPAS scores for roughly 7,000 people arrested in Broward County, Florida, in 2013 and 2014, and checked who was charged with a new crime within two years. Race is not an input to COMPAS, which uses criminal history and a questionnaire. Even so, ProPublica found that among people who did not reoffend, Black defendants were nearly twice as likely as white defendants to have been labeled higher risk. Among people who did reoffend, white defendants were more often labeled lower risk. Put simply, false positive rates were higher for Black defendants and false negative rates were higher for white defendants. ProPublica also reported that overall accuracy was modest, around 60 percent. Northpointe replied that the tool was fair by a different standard. At each score level, Black and white defendants reoffended at similar rates, so a given score meant the same thing regardless of race. This property is called calibration, or predictive parity. Each side's numbers were largely correct. Researchers then showed why both could be right. Jon Kleinberg, Sendhil Mullainathan and Manish Raghavan, and separately Alexandra Chouldechova, proved that when two groups have different base rates of the outcome, a score cannot be calibrated and also have equal false positive and false negative rates across groups. The only exception is a perfect predictor. So choosing a fairness definition is a value judgment, not a technical detail. In State v. Loomis (2016), the Wisconsin Supreme Court allowed judges to consider COMPAS at sentencing. It ruled that the score could not be the deciding factor and required written cautions to accompany it. A 2018 Dartmouth study by Julia Dressel and Hany Farid found that untrained online volunteers predicted rearrest about as accurately as COMPAS. A common misconception is that leaving race out removes bias. Correlated inputs such as prior arrests can carry historical policing patterns into the score.
战略影响
风险与安全
灾难性和日常的人工智能危害都取决于谁了解风险以及谁能够采取行动。
更清晰的判决
公众和专业素养决定强有力的安全政策在政治上是否可行。
打破炒作
清晰的解释可以减少炒作、实验室公关和模糊道德剧场的影响。
The Future of COMPAS and Bias in Recidivism Algorithms
Risk assessment tools are still used in many US pretrial and sentencing settings, and the debate continues. Some reform groups back them as more consistent than unaided judgment, while others argue they entrench historical bias. Likely developments include requirements for local validation, public reporting of error rates by group, and access for defendants to how scores are produced, though the pace varies by jurisdiction. The impossibility result is now a standard part of fairness teaching and shapes audits in hiring, lending and healthcare. It does not settle which definition to use. That remains a policy choice about which errors a society is more willing to accept.
现实世界的实施
In a county that uses COMPAS, a pretrial officer combines a defendant's questionnaire answers with their criminal history. The result is a set of decile scores from 1 to 10 that appears in the report sent to the judge.
After State v. Loomis, a Wisconsin judge who receives a COMPAS report also receives written cautions. They warn that the method is proprietary, that the scores describe groups rather than individuals, and that questions have been raised about racial disparities.
A data scientist auditing a lending model computes both calibration and false positive rates for each group. As with COMPAS, equalizing one breaks the other because the groups' base rates differ.
Researchers following the 2018 Dartmouth study show that a simple model using only age and number of prior convictions predicts rearrest about as accurately as the full COMPAS score.
风险与防护栏
将存在风险视为科幻小说，同时能力复合。
混淆了表面产品安全与高度自治下的对准。
只给非英语和非专业观众留下低质量的资源。
实施路线图
单独的产品危害、误用和失控/失调风险。
询问哪些证据会改变您对时间表和严重性的看法。
比起营销主张，更喜欢主要来源和具体评估。
确定一条行动路径：职业、政策、资金或技能——而不仅仅是意识。
不断探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the COMPAS and Bias in Recidivism Algorithms quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常见问题
What is COMPAS and Bias in Recidivism Algorithms?
COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) is a commercial risk assessment tool from Northpointe, now Equivant, that scores how likely a defendant is to reoffend. It became the central case study in algorithmic fairness after a 2016 ProPublica investigation found that its errors fell unevenly by race. The debate that followed showed that common definitions of fairness can mathematically conflict.
Among defendants who did NOT reoffend, what did ProPublica find about Black defendants compared with white defendants?
ProPublica found a higher false positive rate for Black defendants. Among people who did not reoffend, Black defendants were nearly twice as likely to have been labeled higher risk.
What was the core of Northpointe's defense of COMPAS?
Northpointe argued predictive parity, also called calibration: a given score corresponds to a similar reoffense rate regardless of race.
Under what condition does the impossibility result say calibration and equal error rates cannot all hold?
Kleinberg, Mullainathan and Raghavan, and separately Chouldechova, showed the conflict arises whenever base rates differ and prediction is imperfect.
What did the Wisconsin Supreme Court hold in State v. Loomis?
The court allowed COMPAS to be used alongside other factors, with written cautions about its limits, but ruled it could not be the deciding factor.
Is race an input to COMPAS?
Race is not an input. Disparities can still arise through correlated inputs and through differences in base rates.
继续学习
相关指南
为此主题精选的更多指南