The AI Governance Gap in Criminal Justice: Why Safeguards Must Match the Speed of Deployment

When Angela Lipps was arrested at gunpoint by U.S. Marshals in her Tennessee home last July, she had never set foot in North Dakota. A facial recognition algorithm had flagged her driver’s license photo as matching a suspect in a Fargo bank fraud case. Detectives treated that probabilistic output—a mathematical similarity score, not evidence—as probable cause. Lipps spent nearly six months in jail before her public defender produced bank records proving she was 1,200 miles away during the crimes.

Her case arrived in national headlines just as two major reports landed on the desks of criminal justice policymakers: the Council on Criminal Justice’s AI User Decision Framework (March 2026) and RAND’s companion AI Taxonomy for Criminal Justice (May 2026). Together, they deliver an uncomfortable diagnosis: the criminal justice system is adopting AI tools faster than it is building the guardrails to prevent exactly the kind of catastrophe that destroyed Angela Lipps’s life.

What Happens When AI Outpaces Accountability?

“The main thing is that governance is required immediately… but yet people are adopting these technologies faster than the conversations have developed,” said Kristin Warren, a RAND engineer who researches equitable policies and contributed to the taxonomy report. Her question—”What are the real human costs here?”—is no longer abstract.

The taxonomy catalogs AI applications across four sectors: law enforcement, courts, corrections, and community supervision. Its central finding is that current deployment disproportionately concentrates on high-stakes decisions—pretrial risk assessment, sentencing recommendations, law enforcement targeting—while neglecting lower-risk administrative applications where AI could deliver efficiency gains with far less potential for harm.

This is backwards. The report recommends that agencies start with workflow automation, scheduling, document management, and case tracking—applications where errors carry administrative rather than liberty consequences—before expanding to domains that directly affect whether someone goes to prison.

How AI Is Actually Being Used—and Misused—in Criminal Justice Today

The landscape of AI deployment in criminal justice extends far beyond facial recognition:

• Predictive policing algorithms analyze historical crime data to suggest where officers should patrol or who might commit crimes—tools repeatedly shown to amplify existing biases in policing data rather than correct them.
• Pretrial risk assessment instruments generate scores that influence bail and detention decisions, often without defendants or their attorneys knowing the variables, weights, or training data behind the score.
• Automated social media monitoring scans public posts for language patterns flagged as suspicious, feeding investigative leads that may lack context or cultural understanding.
• AI-assisted legal research tools draft motions and research memos for courts—including a pilot program in California’s Riverside and Los Angeles county courts using a tool called “Learned Hand” that is currently limited to civil matters but has a roadmap to expand into criminal cases.
• Gunshot detection systems trigger police responses before any call is made, sometimes generating false activations in communities already over-policed.

Each tool generates data that can enter criminal proceedings. Each carries documented failure modes that most judges, prosecutors, and defense attorneys were never trained to evaluate.

The “Learned Hand” Experiment: When AI Reaches the Criminal Courtroom

Perhaps no current development better illustrates the governance gap than the AI court assistant being tested in Southern California. The Riverside County Superior Court signed a $10,000 agreement with Learned Hand—a company that uses language models from Anthropic, OpenAI, and Google—to help court attorneys draft research memos and summarize motions.

Currently, seven civil and probate attorneys have access to the tool. But emails obtained by CalMatters reveal the court’s internal discussion about expanding into criminal cases, specifically targeting “things with the largest paper records”—death penalty habeas petitions and parole revocation proceedings.

Los Angeles County Superior Court has a $314,000 contract with a roadmap to test the tool in criminal, family, and probate divisions. Officials declined to describe their criteria for evaluating whether expansion to criminal cases is safe.

The implications are significant: an AI system trained on legal databases might produce plausible-sounding analysis that contains factual errors, misapplied precedent, or subtle biases invisible to overworked court staff. In a parole revocation proceeding—where someone’s return to prison hangs in the balance—an AI hallucination disguised as legal research could have devastating consequences.

What the CCJ Framework Actually Recommends

The Council on Criminal Justice’s User Decision Framework is not anti-technology. It presents a structured five-phase process for agencies to evaluate, procure, and monitor AI tools—a practical checklist rather than a blanket prohibition. Its key recommendations:

1. Explicit prohibitions on algorithmic determinations of guilt, sentencing, and charging. AI can inform but must never decide.
2. Mandatory judicial disclosure when AI recommendations influence court outcomes.
3. Defense access to information about all AI tools used in their clients’ cases—directly challenging the trade secret claims that currently shield many algorithms from scrutiny.
4. Independent expert testing and validation before deployment, not after harm occurs.
5. Public disclosure of basic information about tools being considered, piloted, or deployed.

Indiana’s Supreme Court AI Governance Committee issued guidance late in 2025 requiring judges to consider transparency, human verification steps, and vendor accountability. Montana’s Fourth Judicial District now requires disclosure of AI use in all court filings, including the specific tool used and certification that all citations have been manually verified. California is considering legislation to prevent arbitrators from delegating decision-related tasks to AI entirely.

These are encouraging signals, but they remain islands of governance in an ocean of unregulated deployment.

The Deeper Problem: AI as Accelerant for Existing Structural Failures

The Lipps case is instructive not because facial recognition technology is inherently evil, but because it exposed a failure that predates AI entirely: investigators who skip basic due diligence when they believe they have a shortcut to closure.

As the ElcomSoft analysis noted, facial recognition software produces “mathematical probabilities, not definitive facts.” It is designed as an investigative lead—”closer to an unverified tip phoned in by an anonymous informant than to actual evidence.” The technology itself did not jail Angela Lipps for six months. Institutional failures did: detectives who treated a probability as certainty, a prosecutor who filed charges without corroboration, a judge who signed an arrest warrant based on an algorithm’s output, and a system that held a grandmother without bail for four months before extradition.

AI did not create these failures. It accelerated them. And that acceleration effect is precisely what makes the governance gap so dangerous: tools that enable faster decisions also enable faster mistakes, at a scale that human-only systems could never achieve.

Where AI Could Help Instead of Harm: The Low-Risk Opportunity

Both reports identify a significant untapped opportunity: redirecting AI capabilities from decision-making to oversight and accountability. Instead of using AI to predict who will commit crimes or determine bail amounts, agencies could deploy it to:

• Audit existing decisions for bias patterns—identifying systemic racial or socioeconomic disparities in sentencing, bail, or parole outcomes.
• Monitor compliance with court-ordered conditions in community supervision, where current GPS electronic monitoring systems already generate massive datasets that could benefit from intelligent pattern analysis.
• Detect anomalies in case processing—flagging delays, lost documents, or procedural irregularities that delay justice for defendants.
• Automate administrative workflows—scheduling, docket management, form generation—that consume court resources without directly implicating liberty interests.
• Improve defense access—AI tools that help underfunded public defenders search case law, identify precedent, and prepare motions could help level a playing field that currently tilts heavily toward well-resourced prosecution offices.

The RAND taxonomy specifically notes that community supervision—probation, parole, and electronic monitoring—represents a domain where AI-assisted analytics could improve outcomes by identifying risk factors, optimizing supervision intensity, and reducing both false alarms and missed interventions. The key distinction is between AI that makes decisions about liberty and AI that improves the quality of information available to human decision-makers.

What Criminal Justice Agencies Should Do Right Now

The CCJ framework and RAND taxonomy converge on a practical sequence:

1. Inventory existing AI use. Many agencies do not know what AI tools their staff are already using—including consumer-grade tools like ChatGPT for report writing or legal research.
2. Establish governance before expanding deployment. This means written policies, designated oversight personnel, documented decision-making processes, and clear lines of accountability when tools fail.
3. Start low, go slow. Deploy AI in administrative functions first. Demonstrate safety and accuracy in low-stakes applications before allowing any use in proceedings that affect liberty.
4. Mandate human-in-the-loop requirements. For any AI output that could influence a criminal justice outcome, require documented human review, independent verification, and the ability to challenge the tool’s reasoning.
5. Build challenge mechanisms. Defendants and their attorneys must have the right to know when AI tools influenced their case—and the ability to examine those tools’ methodology and accuracy.

As Kristin Warren put it: “I’m not saying that innovation should slow down or people shouldn’t use these AI tools, but you have to match the speed of safeguards to the stakes of the decision being made and how much risk is involved, particularly as it pertains to someone’s freedom.”

The alternative—continuing to deploy powerful tools into a system already prone to error, bias, and institutional inertia—is not innovation. It is negligence dressed in the language of progress.