Algorithmic Sentencing Gains Ground in Criminal Courts
|

Algorithmic Sentencing Gains Ground in Criminal Courts

ByJon Dykstra

Around the world, courts are testing AI tools that suggest sentence lengths, rank offender risk, and summarize aggravating factors. When algorithms begin whispering in a judge’s ear, the line between assistance and automation blurs and one of law’s most consequential decisions hangs in the balance: how long someone should lose their freedom.

Contents hide
When Algorithms Meet the Bench
Parallel Paths Worldwide
Due Process and the Algorithmic Veil
Can Judges Delegate Judgment?
Opening the Black Box
When Accuracy Masks Discrimination
Toward Human-Centered Justice
Sources

When Algorithms Meet the Bench

In the United States, sentencing algorithms such as COMPAS have been used for over a decade to inform pretrial and sentencing decisions. Judges in states including Wisconsin, New York, California, and Florida may view algorithmic risk scores, though courts maintain those numbers are advisory. In State v. Loomis (2016), the Wisconsin Supreme Court upheld COMPAS’s use but warned it must never be the “determinative” factor in sentencing. The ruling preserved judicial discretion but validated an uneasy alliance between human conscience and coded logic.

Yet COMPAS is not the most widely used tool. The Public Safety Assessment (PSA), developed by Arnold Ventures, has been adopted across more than 40 jurisdictions encompassing over 15 percent of the U.S. population, including entire states like New Jersey, Arizona, Utah, and Kentucky. Unlike COMPAS, the PSA is free, requires no interview, and makes its algorithm publicly available. It predicts three pretrial outcomes: failure to appear, new criminal activity, and new violent criminal activity. While designed for pretrial decisions rather than sentencing, it represents the broader trend toward data-driven criminal justice.

At the federal level, the Bureau of Prisons developed PATTERN (Prisoner Assessment Tool Targeting Estimated Risk and Needs) as mandated by the First Step Act of 2018. PATTERN assesses both general and violent recidivism risk to determine eligibility for early release and program placement.

The tool uses 15 variables spread across gender-specific models and has been subjected to annual review by the National Institute of Justice. While PATTERN demonstrates relatively high predictive accuracy across racial groups, researchers continue working to address remaining issues of differential prediction that affect Black and white defendants at different rates.

This distinction between pretrial risk assessment and sentencing tools matters. Pretrial instruments evaluate whether someone will appear for court or commit crimes while on release. Sentencing algorithms, by contrast, may influence the length and type of punishment after conviction. Both raise due process concerns, but sentencing carries the added weight of determining how long someone loses their freedom.

Parallel Paths Worldwide

Beyond American borders, the movement accelerates. China’s Supreme People’s Court issued guidance in 2022 encouraging the “in-depth integration” of AI with judicial work. The Shanghai “206 system”, developed through collaboration between Shanghai courts, the procuratorate, public security bureau, and AI company iFLYTEK, has been deployed across criminal proceedings in that city.

The system encompasses similar case recommendations, evidence authentication, and can detect inconsistencies in testimony. According to research published in 2019, all pending criminal cases in Shanghai were being entered into the system. The tools promise consistency and speed but invite a deeper question: can justice remain individual when decisions are driven by pattern recognition?

The United Kingdom and Australia have taken more cautious approaches. The UK’s Offender Assessment System (OASys) and Offender Group Reconviction Scale (OGRS) combine algorithmic risk assessment with human evaluation for sentencing and parole decisions, though independent researchers have not been permitted access to OASys data for external validation since its introduction in 2001.

The Courts and Tribunals Judiciary issued guidance in December 2023 approving certain AI uses for judges, including summarizing large bodies of text, but maintaining strict human oversight. Australian courts have similarly explored AI for case management and administrative efficiency while questioning whether tools like COMPAS would be appropriate in their criminal justice context.

Due Process and the Algorithmic Veil

Sentencing has long balanced statutory guidelines with judicial intuition. Introducing AI shifts that balance toward data-driven uniformity. Critics raise a structural concern: algorithms trained on historical sentencing data inevitably inherit racial and socioeconomic bias. ProPublica’s 2016 analysis found that COMPAS incorrectly classified Black defendants as high-risk at nearly twice the rate of white defendants, while white defendants were more likely to be incorrectly classified as low-risk. When these outputs inform punishment, discrimination becomes automated, and accountability becomes obscured.

The proprietary nature of many algorithms compounds the problem. COMPAS’s methodology remains protected as a trade secret, preventing independent scrutiny by defendants, attorneys, or judges. As one analysis in the Harvard Law Review noted, the Wisconsin Supreme Court’s acceptance of this opacity means “that an individual may be sentenced, in part, based on an algorithm no one in the courtroom fully understands.”

Can Judges Delegate Judgment?

Judicial ethics frameworks have begun to catch up. In July 2024, the American Bar Association’s Standing Committee on Ethics and Professional Responsibility released Formal Opinion 512, its first guidance on generative AI use in legal practice. While focused primarily on lawyers rather than judges, the opinion emphasizes duties of competence, transparency, and human review. These principles extend naturally to judicial use of AI tools: automation without responsibility undermines justice.

Federal legislative efforts are also underway. The Algorithmic Accountability Act of 2023, reintroduced in Congress by Senators Ron Wyden and Cory Booker alongside Representative Yvette Clarke, would require companies deploying automated decision systems in critical areas including criminal justice to conduct impact assessments and report findings to the Federal Trade Commission.

The bill specifically targets systems affecting access to housing, credit, education, and legal processes. While it did not advance in the 118th Congress, its reintroduction signals growing bipartisan concern about algorithmic accountability in high-stakes decision making.

The judiciary itself is mobilizing oversight mechanisms. In July 2024, the Council on Criminal Justice launched a national Task Force on Artificial Intelligence, chaired by former Texas Supreme Court Chief Justice Nathan Hecht. The 15-member task force includes AI developers, researchers, police executives, civil rights advocates, and formerly incarcerated individuals. Over 18 months, the group will develop consensus principles for safe and ethical AI use across law enforcement, courts, corrections, and community organizations.

Several state chief justices have established parallel initiatives, including California Chief Justice Patricia Guerrero’s AI Task Force in May 2024 and Georgia Chief Justice Michael Boggs’ committee in October 2024, both focused on safeguarding judicial integrity while exploring AI benefits.

Internationally, the EU AI Act, which entered into force in August 2024, takes a harder line. Article 5 prohibits AI systems that assess or predict the risk of a natural person committing a criminal offense based solely on profiling or personality assessment. The prohibition does not apply to AI used to support human assessment based on objective, verifiable facts linked to criminal activity, but it establishes a clear boundary: fully automated criminal justice decisions are forbidden.

Still, many judges welcome AI as a clerical tool. Sentencing platforms can compile prior cases, compute ranges under federal guidelines, or flag outliers for human confirmation. Used properly, they enhance consistency; misused, they create the illusion of objectivity. The challenge is not whether to use AI, but when its assistance becomes deference.

Opening the Black Box

Legal technologists now argue for algorithmic disclosure akin to expert testimony. If an AI tool influences a sentence, defendants should know the model’s name, data sources, and error rate. This approach parallels the Daubert standard for scientific evidence, translating judicial scrutiny from laboratories to codebases. Without such transparency, appeal courts face a blind record: a sentence recommended by something no one fully understands.

The PSA represents one model of transparency: its nine risk factors and algorithm are publicly available, and independent researchers continuously evaluate and validate it across jurisdictions. Studies show it does not exacerbate racial disparities, though questions remain about local accuracy when a tool calibrated on national data is applied to specific communities.

Some jurisdictions are beginning to mandate disclosure. Stanford Law School’s Pretrial Risk Assessment Tools Factsheet Project provides standardized audits of major risk assessment tools, offering stakeholders a mechanism to evaluate and compare these systems. The project frames transparency as both procedural fairness and public education. If the public is to trust AI in court, it must first see how the system reasons.

When Accuracy Masks Discrimination

The algorithmic bias debate centers on a mathematical impossibility: a tool cannot simultaneously be well-calibrated (accurate across all groups) and satisfy equalized odds (equal false positive and false negative rates across groups) when base rates of recidivism differ between populations.

COMPAS’s defenders point to calibration: the tool predicts recidivism at roughly the same accuracy for Black and white defendants. Critics emphasize error distribution: the tool fails differently, producing more false positives for Black defendants and more false negatives for white defendants.

As scholars writing in the Columbia Human Rights Law Review observed, “given the state of crime and policing in the United States, a criminal sentencing algorithm cannot boast both predictive accuracy and equalized odds.” The choice of fairness metric becomes a moral decision that algorithms cannot make, yet when buried in code, it masquerades as objectivity.

Research on Virginia’s sentencing algorithm revealed unintended consequences of human-machine interaction. A 2019 Washington Post analysis of the state’s risk assessment system found that after its adoption in 2002, defendants younger than 23 were 4 percentage points more likely to be incarcerated and received 12 percent longer sentences than older peers. Black defendants also faced a 4 percentage point increase in incarceration likelihood compared with otherwise equivalent white defendants. Rather than reducing disparities, the algorithm amplified judicial tendencies to deviate from its recommendations in ways that disadvantaged already vulnerable groups.

Toward Human-Centered Justice

Whether AI should recommend sentences may ultimately depend less on technology than on philosophy. The criminal justice system was designed to express moral judgment, not statistical prediction. Allowing a machine to suggest punishment risks turning justice into compliance management. Yet banning AI entirely would ignore its potential to expose inconsistency and guide reform.

The middle path now forming in court committees worldwide insists on human supremacy with algorithmic transparency. AI can advise, never decide. It must be explainable, auditable, and subject to challenge. Its limitations must be disclosed, its biases acknowledged, and its predictions verified against human judgment informed by the full context of a defendant’s life.

As sentencing becomes another frontier of machine reasoning, the judiciary’s task is to remember that fairness cannot be computed; it must be reasoned. The algorithm may whisper, but the judge must answer and take responsibility for what follows.

Sources


This article was prepared for educational and informational purposes only. It does not constitute legal advice and should not be relied upon as such. All cases, statutes, and sources cited are publicly available through court filings, government websites, and reputable legal publications. Readers should consult professional counsel for specific legal or compliance questions related to AI use in their practice.

See also: When Machines Decide, What Are the Limits of Algorithmic Justice?

More…

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *