Virginia Courts’ Algorithmic Recidivism Risk Assessment Failed to Lower Incarceration Rates

Virginia courts’ use of algorithmic predictions of future offending risks were found by researchers failing to reduce incarceration rates, showed racial and age disparities in risk scores and its application, and neither exacerbated or ameliorated historical racial differences in sentencing.

Risk classification

• Primary risk domain: 1 Discrimination & Toxicity
• Primary risk subdomain: 1.1 Unfair discrimination and misrepresentation

The algorithmic risk assessment tool introduced and exacerbated unfair sentencing outcomes and systemic discrimination against young and Black defendants based on demographic characteristics.

Additional risk subdomains

• 5.1 Overreliance and unsafe use: Judges relied on the algorithmic scores to rationalize harsher sentences or to shield themselves from public backlash, using the tool in critical situations without adequate training.
• 7.3 Lack of capability or robustness: The algorithm struggled because predicting future recidivism is extremely difficult, and the tool failed to perform reliably or effectively in predicting actual risk.

Causal factors

• Entity: AI
• Intent: Unintentional
• Timing: Post-deployment

The negative outcomes, including increased racial and age disparities and lack of incarceration reduction, were unexpected and unintended consequences of deploying the algorithmic risk assessment tool.

EU AI Act risk tier

• Risk tier: 2 High Risk

High Risk: The system is used in the administration of justice and law enforcement, specifically to assist judges in sentencing decisions, which has significant implications for fundamental rights and liberty.

AI system and alleged parties

• AI system: None named
• AI purpose: Recidivism Prediction; Judicial Decision Support
• Behaviour type: Assistant
• Alleged developer: Virginia Department of Criminal Justice Services
• Alleged deployer: Virginia courts
• Alleged harmed parties: Virginia young offenders, Virginia convicted felons, Virginia Black offenders

Harm severity

Severity ratings are based on this scale.

Highest direct severity in any category: Substantial. Severity is scored from Negligible to Catastrophic in each harm category, for harm the reports describe as caused directly or indirectly by the AI system.

• Physical: direct Negligible, indirect Negligible
• Infrastructure: direct Negligible, indirect Negligible
• Property: direct Negligible, indirect Negligible
• Financial: direct Negligible, indirect Negligible
• Environmental: direct Negligible, indirect Negligible
• Malicious content: direct Negligible, indirect Negligible
• Differential treatment: direct Substantial, indirect Substantial
• Civil rights: direct Substantial, indirect Substantial
• Democracy: direct Negligible, indirect Negligible
• Privacy: direct Negligible, indirect Negligible
• Psychological: direct Negligible, indirect Negligible
• Epistemic: direct Negligible, indirect Negligible
• Child sexual exploitation and abuse: direct Negligible, indirect Negligible

Differential treatment

Reported: Yes, the report explicitly describes differential treatment based on race and age.

Directly caused: The risk assessment tool directly assigned higher risk scores to younger defendants and Black defendants, leading to longer sentences.

Indirectly caused: Judges were more likely to follow leniency recommendations for White defendants than for Black defendants, resulting in Black defendants being 4 percentage points more likely to be incarcerated and receiving 17 percent longer sentences.

Inferred additional harm: Systemic bias in the judicial system was reinforced and legitimized by the algorithmic scores, leading to widespread unequal sentencing outcomes across Virginia.

Civil rights

Reported: Yes, the report describes impacts on defendants' liberty and fair trial rights through biased sentencing.

Directly caused: Defendants under 23 were 4 percentage points more likely to be incarcerated and received 12 percent longer sentences due to the algorithm's heavy weighting of age.

Indirectly caused: Black defendants faced a 4 percentage point increase in incarceration rates and 17 percent longer sentences due to the combined effect of higher scores and harsher judicial application of those scores.

Inferred additional harm: Thousands of individuals may have had their civil liberties and rights to fair, unbiased sentencing compromised over the years of the tool's statewide deployment.

People affected

• Occurrences reported: 1
• People reportedly harmed: 5000
• People reportedly exposed: 50000

Potential causes

Management

• Inadequate Risk Assessment: Management failed to anticipate unintended impacts on vulnerable groups.
• Conflicting System Objectives: Tool design goals conflicted with judges' individual sentencing criteria.
• Lack of System Accountability: Management did not correct documented racial and age disparities.

Technology

• Age-Biased Risk Scoring: Algorithm heavily weighted age, inflating scores for young defendants.
• Poor Recidivism Predictability: The tool failed to accurately predict future crime or recidivism rates.
• Flawed Sex Offender Scoring: Scoring enabled lower sentences for rapists by shielding judges from blame.

Data Inputs

• Biased Demographic Proxies: System used age and marital status as proxies, driving systemic disparities.
• Biased Historical Data: Input data reflected historical racial disparities in the justice system.

Human Factors

• Lack of Judicial Training: Judges received almost no training on how to interpret risk scores.
• Algorithmic Rationalization: Judges used the tool to justify decisions they already wanted to make.
• Asymmetric Risk Aversion: Judges feared public backlash unless protected by a low-risk score.

Process and Methods

• Unchecked Judicial Discretion: Judges could ignore or apply risk recommendations without accountability.
• Lack of Diversion Programs: Some jurisdictions lacked infrastructure to support rehabilitation paths.
• Inconsistent Implementation: The tool was deployed without standardizing how judges should use it.

Regulatory Environment

• Advisory-Only Policy: Advisory nature of the tool meant compliance was entirely optional.
• Abolition of Discretionary Parole: State policy changes forced reliance on algorithms to manage prison capacity.

Information quality

• Classification confidence: High
• Reason for confidence: The reports are based on a rigorous academic working paper analyzing tens of thousands of felony convictions in Virginia, providing clear statistical evidence of the algorithm's impact on sentencing, racial disparities, and age disparities. The roles of both the AI system and the human judges are explicitly detailed.

Virginia's deployment of algorithmic sentencing tools led to unintended racial and age disparities in criminal sentencing. While causing localized civil and human rights concerns, the incident has negligible national security implications and is manageable through domestic legal and policy reforms.

• Overall national security impact: Minor
• Scope: Single nation
• Primary target: United States
• Alleged perpetrator: Unknown

Threat characteristics

• Imminence: Long-term. Represents an ongoing, systemic policy and technology integration issue spanning decades, rather than an active, sudden crisis.
• Autonomy: Human-supervised. The AI provided risk scores to assist judges, who retained final sentencing authority and ultimate decision-making power.
• Novelty: Established threat. Algorithmic bias and issues with recidivism prediction tools are well-documented, long-standing challenges in criminal justice.

Impact by dimension

• Physical security: Negligible. No physical systems, critical infrastructure, or kinetic targets were affected by this judicial algorithm.
• Information security: Negligible. No intelligence compromise, classified data theft, or foreign disinformation campaigns occurred.
• Sovereignty: Minor. Affected state-level judicial sentencing processes in Virginia, but did not threaten overall state authority, constitutional processes, or federal government functions.
• Economic security: Negligible. No strategic technology theft, financial system attacks, or loss of national AI competitive advantage.
• Societal stability: Minor. Caused systemic bias and unequal sentencing for thousands of defendants, representing a localized civil rights concern, but did not trigger national-scale instability.