Predictive Algorithms in Sentencing: Are We Automating Bias?

Apr 24, 2018

By Linda Henry


See all of Our JDSupra Posts by Clicking the Badge Below

View Patrick Law Group, LLC

Although algorithms are often presumed to be objective and unbiased, recent investigations into algorithms used in the criminal justice system to predict recidivism have produced compelling evidence that such algorithms may be racially biased.  As a result of one such investigation by ProPublica, the New York City Council recently passed the first bill in the country designed to address algorithmic discrimination in government agencies. The goal of New York City’s algorithmic accountability bill is to monitor algorithms used by municipal agencies and provide recommendations as to how to make the City’s algorithms fairer and more transparent.

The criminal justice system is one area in which governments are increasingly using algorithms, particularly in connection with creating risk assessment profiles of defendants.  For example, COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) is a computer algorithm used to score a defendant’s risk of recidivism and is one of the risk assessment tools most widely used by courts to predict recidivism.  COMPAS creates a risk assessment by comparing information regarding a defendant to historical data from groups of similar individuals.

COMPAS is only one example of proprietary software being used by courts to make sentencing decisions, and states are increasingly using software risk assessment tools such as COMPAS as a formal part of the sentencing process.  Because many of the algorithms such as COMPAS are proprietary, the source code is not published and is not subject to state or federal open record laws. As a result, the opacity inherent in proprietary programs such as COMPAS prevents third parties from seeing the data and calculations that impact sentencing decisions.

Challenges by defendants to the use of such algorithms in criminal sentencing have been unsuccessful.  In 2017, Eric Loomis, a Wisconsin defendant, unsuccessfully challenged the use of the COMPAS algorithm as a violation of his due process rights.  In 2013, Loomis was arrested and charged with five criminal counts related to a drive by shooting.  Loomis maintained that he was not involved in the shooting but pled guilty to driving a motor vehicle without the owner’s permission and fleeing from police.  At sentencing, the trial court judge sentenced Loomis to six years in prison, noting that the court ruled out probation based in part on the COMPAS risk assessment that suggested Loomis presented a high risk to re-offend.[1]Loomis appealed his sentence, arguing that the use of the risk assessment violated his constitutional right to due process.  The Wisconsin Supreme Court ultimately affirmed the lower court’s decision that it could utilize the risk assessment tool in sentencing, and also found no violation of Loomis’ due process rights.  In 2017, the U.S. Supreme Court denied Loomis’ petition for writ of certiorari.

The use of computer algorithms in risk assessments have been touted by some as a way to eliminate human bias in sentencing.  Although COMPAS and other risk assessment software programs use algorithms that are race neutral on their face, the algorithms frequently use data points that can serve as proxies for race, such as ZIP codes, education history and family history of incarceration.[2]  In addition, critics of such algorithms question the methodologies used by programs such as COMPAS, since methodologies (which are necessarily created by individuals) may unintentionally reflect human bias.  If the data sets being used to train the algorithms are not truly objective, human bias may be unintentionally baked into the algorithm, effectively automating human bias.

The investigation by ProPublica that prompted New York City’s algorithmic accountability bill found that COMPAS risk assessments were more likely to erroneously identify black defendants as presenting a high risk for recidivism at almost twice the rate as white defendants (43 percent vs 23 percent).  In addition, ProPublica’s research revealed that COMPAS risk assessments erroneously labeled white defendants as low-risk 48 percent of the time, compared to 28 percent for black defendants.  Black defendants were also 45 percent more likely to receive a higher risk score than white defendants, even after controlling for variables such as prior crimes, age and gender. [3]ProPublica’s findings raise serious concerns regarding COMPAS, however, because the calculations used to assess risk are proprietary, neither defendants nor the court systems utilizing COMPAS have visibility into why such assessments have significant rates of mislabeling among black and white defendants.

Although New York City’s algorithmic accountability bill hopes to curb algorithmic bias and bring more transparency to algorithms used across all New York City agencies, including those used in criminal sentencing, the task force faces significant hurdles.  It is unclear how the task force will make the threshold determination as to whether an algorithm disproportionately harms a particular group, or how the City will increase transparency and fairness without access to proprietary source code.  Despite the task force’s daunting task of balancing the need for more transparency against the right of companies to protect their intellectual property, critics of the use of algorithms in the criminal justice system are hopeful that New York City’s bill will encourage other cities and states to acknowledge the problem of algorithmic bias.

 


[1] State v. Loomis, 881 N.W.2d 749 (2016)

[2] Hudson, L., Technology Is Biased Too. How Do We Fix It?, FiveThirtyEight (Jul. 20, 2017), https://fivethirtyeight.com/features/technology-is-biased-too-how-do-we-fix-it/.

[3] Julia Angwin et al., Machine Bias, ProPublica (May 23, 2016), https://www .propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing.

OTHER THOUGHT LEADERSHIP POSTS:

San Francisco Says The Eyes Don’t Have It: Setting Limits on Facial Recognition Technology

By Jennifer Thompson | On May 14, 2019, the San Francisco Board of Supervisors voted 8-1 to approve a proposal that will ban all city agencies, including law enforcement entities, from using facial recognition technologies in the performance of their duties.

NYC’s Task Force to Tackle Algorithmic Bias: A Study in Inertia

By Linda Henry | In December, 2017 the New York City Council passed Local Law 49, the first law in the country designed to address algorithmic bias and discrimination occurring as a result of algorithms used by City agencies.

U.S. Lawmakers Want Companies to Check their Bias

By Linda Henry | Although algorithms are often presumed to be objective and unbiased, technology companies are under increased scrutiny for alleged discriminatory practices related to their use of artificial intelligence.

The Weight of “GDPR Lite”

By Dawn Ingley | In June, 2018, California’s legislature took the first steps to ensure that the state’s approach to data privacy was trending more closely to the European Union’s General Data Protection Regulation (GDPR), the de facto global industry standard for data protection. Though legislators have acknowledged that further refinements to the California Consumer Privacy Act (CCPA) will be necessary in the coming months, its salient requirements are known.

The ABA’s Valentine’s Gift to Same-Sex Couples: Formal Opinion 458 Requires Judges to Perform Marriages

By Jennifer Thompson | On Valentine’s Day, the American Bar Association (ABA) Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 485, entitled “Judges Performing Same-Sex Marriages,” stating that judges may not decline to perform marriages for couples of the same sex.

The Intersection of Artificial Intelligence and the Model Rules of Professional Conduct

By Linda Henry | Artificial intelligence is transforming the legal profession and attorneys are increasingly using AI-powered software to assist with a wide rage of tasks, ranging from due diligence review, issue spotting during the contract negotiation process and predicting case outcomes.

Follow the Leader: Will Congressional and Corporate Push for Federal Privacy Regulations Leave Some Technology Giants in the Dust?

By Dawn Ingley | On October 24, 2018, Apple CEO Tim Cook, one of the keynote speakers at the International Conference of Data Protection and Privacy Commissioners Conference, threw down the gauntlet when he assured an audience of data protection professionals that Apple fully supports a “GDPR-like” federal data privacy law in the United States.

Yes, Lawyers Too! ABA Formal Opinion 483 and the Affirmative Duty to Inform Clients of Data Breaches

By Jennifer Thompson | Developments in the rules and regulations governing data breaches happen as quickly as you can click through the headlines on your favorite news media site.  Now, the American Bar Association (“ABA”) has gotten in on the action and is mandating that attorneys notify current clients of real or substantially likely data breaches where confidential client information is or may be compromised.

GDPR Compliance and Blockchain: The French Data Protection Authority Offers Initial Guidance

By Linda Henry | The French Data Protection Authority (“CNIL”) recently became the first data protection authority to provide guidance as to how the European Union’s General Data Protection Regulation (“GDPR”) applies to blockchain.

D-Link Continues Challenges to FTC’s Data Security Authority

By Linda Henry | On September 21, 2018, the FTC and D-Link Systems Inc. each filed a motion for summary judgement in one of the most closely watched recent enforcement actions in privacy and data security law (FTC v. D-Link Systems Inc., No. 3:17-cv-00039).  The dispute, which dates back to early 2017, may have widespread implications on companies’ potential liability for lax security practices, even in the absence of actual consumer harm.