Algorithmic tools promise to make criminal justice decisions more consistent, evidence-based, and efficient. Risk-assessment instruments use statistical models to predict the likelihood that an individual will reoffend or fail to appear for trial. Predictive policing tools use historical crime data to forecast where crimes are likely to occur. Facial recognition technology compares images against databases to identify suspects. Each of these applications raises distinct questions about accuracy, bias, transparency, and accountability.

Risk-assessment instruments have been used in pretrial, sentencing, and parole decisions for decades. Modern instruments typically use a small number of factors — criminal history, age, employment, and sometimes neighborhood characteristics — to produce a risk score. The COMPAS instrument, used in many jurisdictions, was the subject of a widely cited ProPublica analysis that found it produced false positives at higher rates for Black defendants than for white defendants. The instrument's developer disputed this finding, and subsequent research has shown that the debate involves fundamental tradeoffs between different definitions of fairness that cannot all be satisfied simultaneously.

Algorithmic bias can arise from multiple sources. Training data that reflects historical patterns of discriminatory enforcement will embed those patterns in the model's predictions. Features that are correlated with race — such as neighborhood, employment history, or prior arrests — can produce racially disparate predictions even when race is not explicitly included. The choice of outcome variable matters: an instrument trained to predict rearrest will reflect policing patterns as much as actual behavior.

Predictive policing tools use historical crime data to identify locations or individuals at elevated risk of future crime. Place-based tools — such as PredPol (now Geolitica) — generate maps of predicted crime hotspots that departments use to direct patrol resources. Individual-based tools — such as Chicago's Strategic Subject List — score individuals based on their predicted risk of involvement in gun violence. Research on place-based tools has found modest crime-reduction effects in some contexts, but critics argue that directing patrol to algorithmically identified areas can create feedback loops that amplify existing disparities.

Facial recognition technology compares a probe image against a database of reference images to identify potential matches. Accuracy varies substantially by demographic group: studies by the National Institute of Standards and Technology (NIST) have found that many algorithms produce higher false-positive rates for Black women, Asian individuals, and older adults than for white men. False positives in a law enforcement context can lead to wrongful arrests. Several documented cases of wrongful arrest based on facial recognition misidentification have prompted legislative responses in some jurisdictions.

Transparency is a persistent challenge with algorithmic tools. Many commercial risk-assessment instruments are proprietary, meaning that defendants and their attorneys cannot examine the algorithm that produced their risk score. In State v. Loomis, 881 N.W.2d 749 (Wis. 2016), the Wisconsin Supreme Court upheld the use of COMPAS at sentencing but noted that defendants must be informed that the instrument cannot be used as the determinative factor in sentencing. The court's reasoning has been criticized for failing to address the fundamental due process concern that defendants cannot challenge a score they cannot examine.

Human oversight is essential for algorithmic decision-making in criminal justice. Research on automation bias — the tendency to defer to algorithmic recommendations even when human judgment would produce better outcomes — suggests that simply providing a risk score does not guarantee that decision-makers will use it appropriately. Training, clear guidance on the appropriate weight to give algorithmic recommendations, and accountability for decisions that deviate from or defer to those recommendations are all necessary components of responsible implementation.

The legal framework governing algorithmic tools in criminal justice is still developing. The Fourth Amendment's reasonableness requirement applies to searches conducted using technology, but courts have not yet fully addressed how it applies to predictive policing or facial recognition. The due process clause requires that defendants have a meaningful opportunity to challenge evidence used against them, but the application of this requirement to proprietary algorithms is contested. Several states and cities have enacted legislation restricting or banning the use of facial recognition by law enforcement.

Angwin, J., Larson, J., Mattu, S., & Kirchner, L. (2016). Machine bias. ProPublica. https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing

Chouldechova, A. (2017). Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data, 5(2), 153–163.

National Institute of Standards and Technology. (2019). Face Recognition Vendor Test (FRVT) Part 3: Demographic Effects. https://doi.org/10.6028/NIST.IR.8280

State v. Loomis, 881 N.W.2d 749 (Wis. 2016).

Lum, K., & Isaac, W. (2016). To predict and serve? Significance, 13(5), 14–19.

Garvie, C., Bedoya, A., & Frankle, J. (2016). The Perpetual Line-Up: Unregulated Police Face Recognition in America. Georgetown Law Center on Privacy & Technology. https://www.perpetuallineup.org