Courts are right to approach artificial intelligence with caution. The record in high-stakes criminal justice applications is mixed at best. Predictive policing programs in Los Angeles and Chicago were defunded after oversight reviews found limited effectiveness and evidence of biased targeting. In Plainfield, New Jersey, an analysis found that a commercial crime-forecasting platform’s predictive success rate fell below 0.5 percent. At least six people, all Black, have been wrongfully arrested following facial recognition misidentifications documented by the Innocence Project. Research published in Science Advances found that COMPAS, one of the most widely deployed recidivism risk assessment tools, achieved roughly the same predictive accuracy as people with no criminal justice expertise.
That record justifies caution. What it does not justify is applying that caution without distinction.
A May 2026 taxonomy of AI applications in criminal justice, produced by the RAND Corporation for the Council on Criminal Justice, provides the framework courts have been missing. The report organizes AI applications by what they actually do within justice institutions and assigns each category an equity risk rating and a transparency level. When you read those ratings alongside the deployment data, a pattern emerges that reframes the conversation entirely.
A taxonomy built on function, not vendor claims
The RAND/CCJ taxonomy covers five broad categories of AI in criminal justice: risk assessment and prediction, surveillance and identification, analysis and decision support, operations and case management, and training and services. Within each category, applications are rated for equity risk (high, medium, or low), transparency (explainable, partially explainable, or black box), and automation level.
The ratings for high-stakes applications align with the cautionary record. Pretrial risk assessment tools are rated high equity risk and are partially explainable. Recidivism prediction is high equity risk, partially explainable by a black box. Predictive policing is high equity risk, black box. Facial recognition is high equity risk, black box.
The ratings for high-stakes applications align with the cautionary record. Pretrial risk assessment tools are rated high equity risk and partially explainable, meaning their reasoning can be examined but not fully traced. Recidivism prediction is high equity risk, ranging from partially explainable to what researchers call a black box: a system in which data goes in, a result comes out, and no one can fully account for which factors drove the output or how they were weighted. Predictive policing and facial recognition are both rated high equity risk and black box. With those tools, neither the court nor the defendant can see the methodology, challenge the reasoning, or verify whether similar cases were treated consistently.
The same report that documents serious governance failures in predictive policing and risk assessment explicitly classifies court reminder systems as the lowest-risk, most transparent category of AI in the justice system.
The underutilization finding
What makes the RAND/CCJ report particularly useful for court administrators is its discussion of adoption patterns. The taxonomy shows that more sophisticated, higher-risk AI has concentrated in criminal justice functions, where governance is weakest, and the consequences for individuals are most severe. At the same time, low-risk administrative tools, such as scheduling, notifications, and reminders, remain underutilized relative to their documented potential.
The report identifies this gap as a specific policy problem. In its conclusions, Opportunity 1 is titled “Understanding and unlocking efficiency in underutilized low-risk applications.” The authors observe that governance ambiguity may be causing agencies to avoid administrative AI tools not because those tools are genuinely risky, but because court administrators lack a clear framework for distinguishing them from higher-risk applications. Recommendation 3 follows directly: provide explicit governance clarity for low-risk administrative functions so that agencies can adopt these tools with confidence.
In other words, the same document that calls for restraint in predictive policing and risk assessment also explicitly calls for expanded adoption of court reminders and administrative scheduling tools.
The evidence already exists
The case for court reminders does not rest on the RAND taxonomy alone. The taxonomy classifies them as low-risk and explainable, but the evidence base has been building for years before anyone drew that governance map.
A 2020 randomized controlled trial published in Science, conducted by ideas42 across more than 800,000 summonses in New York City, found that text reminders reduced failure-to-appear rates by 21 percent. Combined with a redesigned summons form, that figure reached 36 percent. A 2023 meta-analysis in Criminology & Public Policy synthesized 12 studies involving more than 79,000 participants and found that reminders significantly reduced failure to appear in 11 of the 12 cases studied, with reductions ranging from 11 to 61 percent.
Pew Charitable Trusts reported in 2025 that despite this evidence base, 19 states still have no statewide court reminder program. States using opt-in enrollment reach as few as 2 percent of eligible court users. Auto-enrollment states reach 72 to 90 percent.
The (Un)warranted initiative at ideas42, which tracks real-world outcomes across partner jurisdictions, has documented more than 125,000 prevented missed court dates and an estimated $357 million in system-wide savings. In Tarrant County, Texas, the warrant rate dropped from 9 percent in 2017 to 2.6 percent in 2025 while court volume grew by roughly 50 percent over the same period.
That combination, a low-risk rating from RAND and years of consistent outcome data, describes a tool courts should have more confidence adopting, not less.
What governance clarity looks like in practice
The RAND taxonomy does not tell courts which tools to adopt. It is explicitly designed as a descriptive framework, not a prescriptive one. But it provides something practitioners have needed: a way to talk about AI in criminal justice without treating all applications as equivalent.
For a court administrator considering a reminder program, the taxonomy offers a useful anchoring point. The tool you are evaluating is rated as having low equity risk and is explainable. Its outputs are structured data: messages sent to a known number associated with a specific case record. Errors are detectable and correctable. No feedback loop amplifies bias over time. There is no predictive inference about future behavior. There is no algorithmic opacity to hide behind.
Courts should ask hard questions about facial recognition and recidivism tools: what data trained the model, what demographic disparities exist in error rates, and how affected parties can contest the outputs. These are legitimate governance questions. They are not the right questions for a system that sends a text message to someone with a court date next Tuesday.
The RAND/CCJ taxonomy did not create that distinction. It documented it. Courts that have applied the same level of caution to both categories have operated without the governance map the field now has.
Sources
Pew Charitable Trusts. (2025, May). States underuse court date reminders.
About Greg Shugart
Director of Government Relations
Greg Shugart brings over 30 years of public sector experience to the eCourtDate team, with a background in court administration, criminal justice reform, and government operations. He previously served as Criminal Courts Administrator for Tarrant County, Texas, where he led statewide-recognized initiatives in pretrial modernization, court communications, and system efficiency. Greg now contributes to eCourtDate’s strategy and partnerships, helping agencies implement technology that improves access, compliance, and trust in the justice system.