AI Is Already in the Courtroom — Just Not on Your Side

Yes — but mostly not on your side. AI already plays a role in criminal cases across the US, and it's overwhelmingly the prosecution and the system wielding it: DNA mixture interpretation, facial recognition, recidivism risk assessments, and predictive policing, according to a Cornell Law Review analysis.

Here's the part that should trouble every defender. Judges have often accepted claims that these AI systems should remain undisclosed in court — despite the constitutional protections criminal procedure is supposed to provide. Duke Law's Brandon Garrett and computer scientist Cynthia Rudin put it bluntly: unless AI is interpretable, the lawyers and judges who must use it can't detect the errors underneath. Their argument is that black box AI actually "performs predictably worse" in criminal justice settings, and they call for a right to "glass box" AI in every criminal case — with the burden on the government to justify keeping a system hidden from the defense.

So when someone asks whether anyone has used AI in court, the honest answer cuts two ways:

• The state uses opaque algorithms to generate evidence and assess defendants — often without the defense seeing the source code.
• AI-generated media is now destabilizing trials themselves. Defense attorney and former federal prosecutor Duncan Levin warns that jurors who know manipulation is possible may doubt even genuine video, audio, and texts — "the challenge is not just false evidence; it is persuasive false evidence."
• The defense bar is catching up institutionally. NACDL maintains a dedicated AI hub, including its "AI Playbook for Defenders," covering discovery organization, forensic challenges, and cross-examination building — with the explicit caveat that AI should be used without surrendering strategy, judgment, or control.

That NACDL framing matters, because it's the standard any defense-side AI tool should meet. The problem with prosecution-side AI isn't that it's AI — it's that it's a black box with no citations, no audit trail, and no way to test the output. A grounded evidence-review workspace like Evidence Copilot is built on the opposite premise: every AI answer cites the exact second of video or page of the PDF it came from, every output is marked for human review, and the tool never draws legal conclusions.

There's also a defensive angle Levin raises that's easy to miss: AI skepticism cuts both ways. It may make it harder for the government to authenticate digital proof, but it can just as easily make jurors doubt your exculpatory footage. He argues the prosecution should bear the primary burden of establishing authenticity — a position worth raising in your next motion in limine, and a reason your own evidence workflow needs provenance and audit trails you can actually show the court.

The takeaway: AI is in the courtroom already. The question is whether it arrives as an unchallengeable black box from the state, or as a transparent, cited, human-controlled tool on the defense side.

The Glass Box Standard: Why Interpretable AI Wins in Criminal Cases

When the government's algorithm says your client is dangerous, can you look inside it? Duke Law's Brandon Garrett and computer scientist Cynthia Rudin say you must be able to — and their Cornell Law Review article makes the case with unusual force.

Their argument starts with an empirical claim: black-box AI "performs predictably worse" in criminal justice settings. The intuition that opaque models are somehow more accurate is, in their view, backwards. "Unless AI is interpretable, decisionmakers like lawyers and judges who must use it will not be able to detect those underlying errors," they write. An unexplainable DNA mixture interpretation or facial recognition match isn't more scientific — it's just unauditable.

Garrett and Rudin also flip the secrecy burden. Judges have often accepted claims that AI should remain undisclosed in court, despite the constitutional protections criminal procedure normally provides. The authors argue the burden should rest on the government to justify keeping AI hidden from the defense — not on defendants to justify why they need to see it. That's how discovery works everywhere else. Proprietary algorithms shouldn't get a pass.

Here's the uncomfortable part for defenders: the same standard applies to our own tools. If you're using AI to work through discovery, every output should be traceable to its source — the exact second of bodycam footage, the exact page of the police report. Anything less is a black box you're building into your own case file, and cross-examination material for the prosecution if it ever surfaces.

What does a glass-box workflow look like in practice?

• Every AI-generated answer carries a clickable citation into the underlying evidence — a video timestamp or a PDF page — so you can verify it in one click.
• AI output is marked as AI-generated and pending human review, never presented as a conclusion.
• Originals stay immutable, with every redaction and edit producing a new derived file and an audit trail.
• Suggested inconsistencies — report versus bodycam, statement versus jail call — surface as review signals with both sides quoted, not as findings.

This is the design logic behind Evidence Copilot: a grounded evidence-review workspace where the AI agent answers questions only with citations to the exact source moment. It's also consistent with NACDL's professional norm that defenders use AI "without surrendering strategy, judgment, or control." Garrett and Rudin's point applies at both ends of the courtroom — interpretability isn't a nice-to-have. It's the difference between evidence you can stand behind and a number you can't explain.

How the Defense Bar Is Actually Adopting AI (and the Authenticity Fight Ahead)

Ask ten defenders whether they've used AI in court and you'll get shrugs. Ask whether AI is already in their cases, and the answer is yes — just not on their side of the v.

AI is increasingly used in criminal cases in the US and globally: DNA mixture interpretation, facial identification, recidivism risk scores, predictive policing. And per a Cornell Law Review analysis by Brandon Garrett and Cynthia Rudin, judges have often accepted the government's claim that these systems should stay undisclosed — a black box the defense never gets to open. Their argument is blunt: black box AI "performs predictably worse" in criminal justice, and unless AI is interpretable, the lawyers and judges relying on it can't even detect its errors.

The defense bar isn't waiting for a ruling to catch up. NACDL now maintains a dedicated AI resource hub — covering facial recognition in the courtroom, AI and due process, and scientific evidence — anchored by The AI Playbook for Defenders. It's a structured prompt system for the work you actually do: organizing discovery, testing theories, prepping witnesses, challenging forensics, drafting motions, building cross. The norm it establishes matters more than the prompts: use AI without surrendering strategy, judgment, or control. Assistive workflow, never decision-maker.

But adoption cuts into a second fight most defenders haven't fully priced in: authenticity. Duncan Levin — former federal prosecutor and Harvard Law lecturer — argues that AI-generated media threatens to destabilize the criminal trial itself by undermining confidence in bodycam video, text messages, photos, and geolocation data. Jurors who know manipulation is possible may now doubt genuine evidence. "The challenge is not just false evidence; it is persuasive false evidence," he writes. Traditional authentication tools — witness identification, chain of custody, metadata — were never designed to answer whether an artifact was "born of reality."

That cuts both ways. AI skepticism may make it harder for the government to convict on digital proof — and harder for a defendant to persuade a jury with authentic exculpatory video.

For defenders, the practical takeaway is concrete: your own evidence-review workflow needs provenance and audit trails. When you're prepping a suppression motion off bodycam and a jail call, every AI-assisted answer should cite the exact video moment or PDF page it came from, originals should stay immutable, and AI output should sit marked as pending human review. That's the standard we built Evidence Copilot around — grounded citations, human review on every output, no legal conclusions from the machine.

Garrett and Rudin call interpretability the accuracy answer, not opacity. Levin predicts AI will force the system to become more sophisticated — "a healthy pressure, even if it is an uncomfortable one" — through doctrinal change and provenance tools. The defense bar adopting AI on its own terms, with citations and audit trails intact, is exactly that pressure working in your favor.

A Practical Grounded-AI Review Workflow You Can Defend

The difference between an AI workflow you can defend and one that gets you a talking-to from the bar comes down to one thing: can you show exactly where every output came from? Scholars writing in the Cornell Law Review argue that "glass box" AI — interpretable, citable, auditable — outperforms black-box systems precisely because lawyers and judges can't detect errors in what they can't see. Build your review workflow that way.

The workflow itself is straightforward. Upload the discovery dump — bodycam video, interrogation footage, jail calls, 911 audio, police reports, scanned documents — and get timestamped, diarized transcripts plus OCR text from documents with no text layer. That turns hours of unsearchable media into something you can actually query.

Then let the system surface what you'd otherwise miss. Every new file gets compared against earlier evidence in the case, so inconsistencies — report narrative versus what the bodycam actually captured, a statement versus the jail call audio — appear as review signals before you even think to ask. From a signal showing both sides' quotes, you can draft cross-exam and suppression questions where every question is tied to a quoted source. The tool drafts questions only; no credibility conclusions, and everything is saved as attorney-review notes.

What keeps this defensible:

• Every AI answer carries a clickable citation to the exact second of video or page of the PDF, so you can verify before you rely.
• Every output is marked AI-generated and pending human review — the tool never makes legal conclusions, consistent with NACDL's guidance that AI be used without surrendering strategy, judgment, or control.
• Originals are immutable; redactions create new derived files, and every action is audited.
• Nothing auto-applies — AI-suggested redaction regions and auto-sync proposals file as suggestions needing your review.

That audit trail matters more than it used to. As Duncan Levin has argued, AI-generated media is undermining confidence in digital evidence generally, and traditional authentication doctrines weren't designed to determine whether an artifact was "born of reality." A documented, citation-first chain from question to source moment is your answer when opposing counsel asks how you prepared.

Grounded tools like Evidence Copilot are built around exactly this structure — citations to the source moment, human-review gating, immutable originals — because that's the only version of AI review that survives scrutiny in a courthouse.