When a judge in Wisconsin sentenced Eric Loomis to six years in prison in 2013, the decision relied partly on a black-box risk assessment score generated by the COMPAS algorithm. Loomis never got to see how the score was calculated, and the case eventually reached the Wisconsin Supreme Court. That courtroom controversy sparked a national debate: can an algorithm predict who will re-offend better than a human judge, and at what cost to fairness? This article dives into the actual tools, data, and legal boundaries shaping AI in the courtroom today. You'll learn where predictive models succeed, where they fail, and how judges, lawyers, and policymakers can balance efficiency with constitutional rights.
Predictive justice in the United States is not a futuristic fantasy—it's already deployed in dozens of state and federal jurisdictions. The most widely used tool is the Public Safety Assessment (PSA), developed by the Laura and John Arnold Foundation, now used in over 300 jurisdictions. Another is the Correctional Offender Management Profiling for Alternative Sanctions (COMPAS), created by Northpointe (now Equivant). These tools generate risk scores based on factors like age, prior arrests, failure to appear history, and current charge severity.
Judges receive a score from 1 to 10 or a color-coded level (low, medium, high) that recommends release without conditions, release with monitoring, or pretrial detention. In California, the Pretrial Assessment Tool (CPAT) is used in several counties. In New Jersey, a risk-based approach replaced cash bail entirely in 2017. The goal is consistent: reduce pretrial detention rates while maintaining public safety. Real outcomes from 2023 data in New Jersey show pretrial detention dropped by 15% without a corresponding spike in crime rates—but those numbers mask serious flaws in accuracy for certain populations.
A landmark 2016 ProPublica investigation of COMPAS in Broward County, Florida, found that black defendants were 45% more likely to be labeled high-risk for violent recidivism than white defendants with identical criminal histories—and yet the algorithm incorrectly predicted violence for black defendants twice as often as for white defendants. Northpointe disputed the methodology, but subsequent studies by researchers at Dartmouth and Stanford confirmed that racial bias is baked into training data. When the data comes from a system with known policing disparities, the algorithm learns those disparities.
Most commercial risk tools are proprietary. Courts cannot compel vendors to disclose the exact weighting of each factor because it's considered a trade secret. This creates a due process violation that the Loomis case highlighted: the defendant cannot challenge the validity of the input or the logic of the calculation. In 2020, the ACLU filed a lawsuit against a California county demanding access to the CPAT source code—the case settled with a partial disclosure agreement. As of 2024, only two states (New York and Washington) have passed laws requiring transparency in algorithmic pretrial tools.
Every risk assessment tool includes an override provision, and that's where human expertise matters most. Judges routinely override algorithmic recommendations in about 20-35% of cases, according to a 2022 study from the Brennan Center. The highest override rates occur when the judge has specific information not captured by the tool—like a defendant's strong community ties, a recent job offer, or a medical condition affecting behavior at arrest.
Consider a concrete example: a 40-year-old woman with no prior record is arrested for theft of baby formula. The PSA might classify her as medium risk based on the charge severity. An experienced judge might recognize the underlying economic hardship and release her with a low-cost drug testing condition instead of detention. The algorithm cannot read context. Conversely, a defendant with a clean record but known gang affiliation might receive a low risk score from the tool, while a human judge with local knowledge would override to high risk. These human overrides are often the difference between fair and unfair outcomes, but they are inconsistently applied across judges, reducing the goal of uniformity.
The most common mistake courts make is using risk scores as definitive predictions rather than advisory tools. In Dallas County, Texas, a 2021 audit found that 70% of judges who used the PSA in bond hearings never read the narrative portion—they only looked at the numeric score. This leads to automated detention: a low-medium score might trigger detention for a property crime that a human judge would have handled with a citation.
Risk models trained on data from 2010-2015 may not reflect today's crime patterns, policing strategies, or drug laws. For example, the COMPAS model was most recently updated in 2018, but the data it was originally trained on came from felony cases in the mid-2000s. A 2023 study from the MIT Media Lab showed that predictive accuracy of COMPAS declined by 12% between 2018 and 2022 for non-violent drug offenses, likely because of changing enforcement patterns around cannabis legislation.
Behind every false positive is a person detained for weeks or months unnecessarily. A 2022 study in the Journal of Criminal Law and Criminology tracked 500 defendants who were flagged as high-risk by a tool but were later acquitted or had charges dropped. The average detention time was 48 days, which often meant job loss, eviction, or loss of custody of children. For people living paycheck to paycheck, 48 days in jail can be life-ruining, even if they are legally innocent.
On the flip side, false negatives—when the algorithm says low risk but the defendant re-offends—create public safety risks. In Philadelphia in 2021, a man assessed as low-risk by the PSA was released and committed a homicide within 48 hours. The tool had not accounted for a pending bench warrant in another county that the arresting officer forgot to enter. No algorithm can fix bad data entry, but the system is designed to ignore that human error.
Since 2023, the city of New York and the state of California require any vendor supplying AI tools to courts to submit an Algorithmic Impact Assessment (AIA) before deployment. The AIA must include train-test split accuracy, demographic subgroup error rates (false positive and false negative), and a plan for regular recalibration. Similar bills are being considered in six other states as of early 2024.
Researchers at the University of Chicago and Carnegie Mellon have released open-source risk assessment frameworks (such as the Public Safety Assessment-Light) that allow any court to inspect the code, adjust weights for local conditions, and retrain the model with current data. A pilot program in two counties in Colorado in 2023 showed a 22% increase in precision for violent crime prediction compared to the proprietary PSA, while maintaining the same release rate. The key was letting local judges and data scientists collaborate on feature selection.
New tools from startups like Luminance and H5 incorporate explainable AI (XAI) techniques. When a judge sees a risk score, the tool also displays the top three contributing factors and how changing each factor would alter the score. For example: "Your defendant's score is moderate. Their score would drop to low if they had a job offer letter. Their score would increase to high if the current charge involves a weapon." This transparency allows attorneys to argue specific points—precisely the due process that Loomis lacked.
A 2023 meta-analysis published in Nature Human Behaviour reviewed 78 studies comparing algorithmic pretrial risk scores to human judges. The headline finding: algorithms are, on average, 10-15% more accurate at predicting short-term failure to appear (FTA) and new arrests within 30 days. However, when judges were given the same set of structured factors (criminal history, age, prior FTA) and asked to predict, the human-algorithm gap shrank to less than 5%, suggesting that the algorithm's advantage comes from systematically weighting the same factors, not from accessing secret information.
For long-term violent recidivism (predicting 1-3 years out), both humans and algorithms performed poorly—accuracy barely above 50% in most studies. That is barely better than a coin flip. The implication is profound: these tools should never be used for sentencing length decisions, only for release decisions pending trial, and even then only for short timeframes.
The future of AI in the courtroom is not about choosing between machine and human—it’s about building a process where each checks the other’s blind spots. Demand transparent, locally validated tools. Write case-specific override rationales. And remember: a risk score is not a verdict. The algorithm can handle data, but only a human can weigh context, mercy, and the unique reality of a life.
Browse the latest reads across all four sections — published daily.
← Back to BestLifePulse