CRIMINALYTICS · EDITION 4 · AI IN CRIMINAL JUSTICE
Where algorithms meet accountability. Where data meets due process.
Welcome to our 1,700+ subscribers.
Each week Criminalytics covers one big idea at the intersection of AI and criminal justice. This week, we’re looking inside the courtroom — and what we found should concern everyone.
THIS WEEK’S BIG IDEA
60% of Federal Judges Are Using AI. Nearly Half Have Had Zero Training.
In March 2026, Northwestern University and the Sedona Conference published the first-ever random-sample survey of federal judges and their use of artificial intelligence. The study surveyed 502 randomly selected bankruptcy, magistrate, district court, and court of appeals judges across the country.
The headline finding: more than 60% of responding judges reported using at least one AI tool in their judicial work.
They’re using it for legal research. Document review. Summarizing case records. Some are even using it to draft correspondence.
Here’s the problem: nearly half of those same judges said they received no AI training from their court administration.
Let that sink in.
The people who decide bail. Who evaluate evidence. Who sentence defendants to years or decades in prison. They’re adopting one of the most powerful technologies of our time — and many are doing it without formal guidance, without standardized protocols, and without a shared understanding of how these tools actually work.
THE TRAINING GAP IS A GOVERNANCE GAP
This isn’t a story about judges being reckless. Most are trying to be more efficient in courts drowning in caseloads. The study found that judges primarily use AI for legal research (30%) and document review (15.5%) — exactly the kind of time-consuming work that AI can genuinely improve.
But efficiency without literacy is dangerous.
Consider what happened in an Illinois courtroom last year. Judge Jeffrey Goffinet discovered that a legal brief filed before him cited a case that didn’t exist. He searched two legal databases. He even walked to the courthouse library — a room he hadn’t visited in years — to check the physical volume. The case wasn’t there. It was an AI hallucination, submitted as legal authority.
Goffinet later said something that should be a wake-up call for the entire justice system: “People are going to use [AI], and the courts are not going to be able to be a dam across a river that’s already flowing at flood capacity.”
He’s right. But you don’t solve a flood by ignoring it. You build infrastructure.
THE POLICY PATCHWORK
The Northwestern survey revealed a judiciary without consensus:
- 25% of judges formally permit AI use in their chambers
- 20% formally prohibit it
- 17.6% discourage it without a formal ban
- 24.1% have no official policy at all
That means more than 40% of federal chambers lack clear AI governance. If you’re a defense attorney filing a motion, you may have no idea whether the judge reviewing your brief used AI to research the legal issues, summarize the record, or draft preliminary analysis — and there’s no rule requiring them to tell you.
Meanwhile, Congress introduced the Research and Oversight of AI in Courts Act of 2026 in March — a bill that would create a task force to study AI speech-to-text and automatic speech recognition in federal courts. It’s a start. But studying speech-to-text while judges are already using large language models for legal reasoning is like commissioning a report on bicycle safety while everyone’s already driving cars.
RULE 707: THE RIGHT IDEA, STILL IN PROCESS
There is one genuinely promising development. The U.S. Judicial Conference has proposed Federal Rule of Evidence 707, which would require AI-generated evidence to meet the same reliability standards as expert testimony. Under the proposed rule, any machine-generated evidence offered at trial — whether it’s an AI-enhanced surveillance video, a predictive analytics model, or an algorithmic risk assessment — would need to demonstrate that it’s based on sufficient facts, produced through reliable methods, and reliably applied to the case.
The public comment period closed in February 2026. The Evidence Rules Committee was scheduled to vote on it in May 2026. If ultimately adopted, the rule wouldn’t take effect until December 2027 at the earliest.
Rule 707 is important. But it addresses what happens when AI-generated evidence reaches the courtroom. It doesn’t address the far more common scenario: judges and attorneys using AI tools every day to research, summarize, and analyze — with no evidentiary standard, no disclosure requirement, and no training mandate.
THE REAL QUESTION
The Stanford Policy Lab released a white paper in March 2026 that named the core problem precisely: AI capabilities and products are being developed and deployed in criminal justice without sufficient understanding of how they work, attention to their failure risks, or analysis of how they implicate constitutional and other values.
Judges, prosecutors, police officers, and probation officers are adopting complex technologies without the technical background to assess their design, limitations, or failure modes. Vendors market directly to practitioners who lack both the time and expertise for meaningful independent evaluation.
The paper argued that good guidelines are necessary but not sufficient — what’s needed is a governance entity that can translate rapidly evolving technical realities into actionable advice aligned with constitutional requirements.
I agree. And I’d add one thing.
You can’t audit what you don’t understand.
That’s not a critique of judges. It’s a critique of a system that asks them to evaluate tools they were never trained to evaluate. It’s a critique of court administrations that haven’t prioritized AI literacy. And it’s a critique of a legal profession that’s adopting AI faster than it’s building the institutional capacity to govern it.
WHAT SHOULD HAPPEN NEXT
For court administrators: AI training should be mandatory, not optional. The Northwestern study found that among the 39% of judges who were offered training, nearly 74% attended. The appetite is there. The infrastructure isn’t.
For legislators: Stop studying narrow use cases and start building comprehensive governance. The Research and Oversight of AI in Courts Act is fine — but it needs to be version 0.1, not the whole plan.
For the legal profession: Develop clear disclosure standards. If a judge or attorney uses AI in any capacity that touches case outcomes, the other side should know. Transparency isn’t a burden — it’s a constitutional imperative.
For all of us: Pay attention. The judiciary is one of the few institutions where individual decisions can deprive someone of their liberty. If AI is going to be part of how those decisions get made — and it already is — then the people making those decisions need to understand what they’re using.
THE BOTTOM LINE
The Northwestern survey didn’t reveal a scandal. It revealed something more concerning: a slow-motion governance failure playing out in courtrooms across America. Judges are adopting AI because it’s useful. They’re doing it without training because nobody’s providing it. And they’re doing it without consistent policies because the system hasn’t caught up.
This is exactly how the problems we’ll read about in five years begin — not with a dramatic failure, but with a thousand quiet adoptions that nobody governed.
Pramod Kunju is the Founder and CEO of Nakunj Inc., an AI strategy and data analytics consulting firm. He is the author of AI in Criminal Justice and writes Criminalytics to bridge the gap between technologists who build AI and the justice professionals who use it.
Subscribe to Criminalytics on LinkedIn: linkedin.com/newsletters/criminalytics-ai-in-criminal-justice/
Read past editions at criminalytics.com
Leave a Reply