Why the 2025 Civil Rulings Matter to Your Criminal Practice

None of the 2025 decisions that will shape your 2026 practice were criminal cases. They came out of copyright litigation, product-liability MDLs, and FTC enforcement. But if you handle discovery in criminal court, the principles those courts endorsed are already working their way into what judges, prosecutors, and clients expect from you, and ignoring them is a mistake.

Start with the biggest one: AI-generated content is now treated as discoverable ESI. In an analysis of five consequential 2025 rulings, HaystackID's Phil Favro highlights In re OpenAI, where Judge Ona Wang ordered production of 20 million output logs, including user prompts and AI outputs, over relevance and privacy objections. Favro's takeaway applies to any litigant: preservation and production obligations "could include AI content." Criminal discovery has no direct Rule 26 analogue, but the direction is clear. When a case involves AI transcription, AI-assisted reports, or AI analysis of evidence, expect courts to take seriously whether that material exists, was preserved, and can be audited. That cuts both ways: it's a Brady argument when the prosecution used AI tools, and it's an ethics obligation when you did.

The second trend is production validation. In the Uber passenger sexual assault litigation, the Northern District of California approved a protocol letting the requesting party audit responsiveness calls, including sampling nonresponsive material, because evaluating social media evidence "may sometimes be a subjective question." Favro predicts validation questions "are not going away anytime soon, particularly with the advent of artificial intelligence." Translation for criminal practice: summaries of evidence are losing their shield. Judge Emmett Sullivan said it plainly in the Stevens prosecution context, calling discovery summaries "an opportunity for mischief and mistake." If you're relying on a prosecutor's index or your own unverified review notes, the 2026 expectation is that you can show your work.

That expectation collides with a hard reality. NACDL's research shows Brady violations usually come from administrative failures, not malice: unmemorialized witness debriefings, misfiled handwritten notes, Brady review delegated to non-lawyers. With 95% of cases ending in guilty pleas, suppression rarely surfaces at all. The defense answer is a systematic discovery review workflow that treats every intake as an audit, not a skim.

It also collides with ABA Formal Opinion 512, which imposes competency, confidentiality, supervision, and verification duties on any attorney using generative AI. Independently verifying AI output is now an ethical baseline, not a nice-to-have. That's the standard we built Evidence Copilot around: every answer cites the exact video moment or PDF page it came from, and every output is marked AI-generated and pending human review. A grounded discovery intake process, where originals stay immutable and every claim is traceable, is how you meet the validation expectations these civil rulings are normalizing.

The civil-criminal line is blurring. Build your 2026 review process as if a judge will ask you to prove it.

Brady Failures Are Administrative, Not Malicious

The scariest thing about Brady violations is how boring they usually are. When you dig into why favorable evidence never reached the defense file, you rarely find a prosecutor twirling a mustache. You find an unmemorialized witness debriefing, a handwritten note that got misfiled, or Brady review delegated to someone who didn't know what to look for.

That's the conclusion the NACDL's analysis of Brady failures reached: administrative sloppeness, not deliberate suppression, is what "created a fertile ground where Brady's mandate could be marginalized and disregarded." The numbers back this up. Of U.S. capital convictions reversed between 1973 and 1995, 16% involved prosecutorial suppression of favorable evidence. And a 2010 Innocence Project study found that 10% of 255 DNA exonerations involved allegations of suppressed evidence.

Here's the problem for your practice: those failures are invisible from your side of the file. You can't demand what you don't know exists. And because roughly 95% of cases end in a guilty plea, most violations never surface at all. There's no appellate court coming to catch the misfiled note in your client's case.

The practical takeaway is that Brady compliance on the defense side can't be a vibe. It has to be a workflow. Specifically:

• Organize the initial discovery dump immediately, before anything else, so nothing sits unindexed in a folder named "New Folder (3)".
• Run targeted Brady demands against specific evidence categories, rather than relying on a blanket request the state can answer with a form letter.
• Compare every new production against what you already have, so inconsistencies between the police report and the bodycam, or the statement and the jail call, surface early.
• Track named parties across the whole file, because the exculpatory detail usually hangs on a specific officer or witness.

That first step matters more than the rest. Everything downstream, from a discovery review workflow to a Brady motion, depends on the dump being organized into a case file you can actually search. Evidence Copilot built its discovery intake process around exactly this: get the bodycam, the 911 audio, the police report PDFs, and the jail calls into one indexed, searchable place on day one.

The point isn't that prosecutors are villains. Most aren't. The point is that the system produces Brady failures without anyone intending one, and the only reliable counter is a review process that assumes the file is incomplete. Trust the workflow, not the production.

The Ethics of AI in Discovery: ABA Formal Opinion 512

You've probably heard a colleague say some version of this: "I'm not putting client files into an AI tool until someone tells me it's allowed." Fair. But the guidance you're waiting for already exists. ABA Formal Opinion 512, issued July 29, 2024, applies the existing Model Rules to generative AI, and it gives defense attorneys a workable test for any tool that touches case evidence.

The opinion breaks the analysis into four duties. Competency under Rule 1.1 means you either understand a tool's benefits and risks or you consult someone who does, and you independently verify what it produces. Confidentiality under Rule 1.6 means the tool has to protect client data, including preventing your input from training the model. Supervision under Rules 5.1 and 5.3 means the firm sets policy, trains staff, and vets vendors. Client communication under Rule 1.4 means you consider telling the client you're using AI on their case.

• Competency: verify the output yourself; don't trust a black box.
• Confidentiality: no training on client data.
• Supervision: policies, training, vendor vetting.
• Communication: talk to the client about AI use.

Here's where that lands in practice. A tool that generates a clean-sounding answer with no way to trace it back to the source fails the competency duty on its face. You can't independently verify a conclusion that arrives without a citation. Compare that to a grounded discovery review workflow where every AI answer carries a citation into the exact second of bodycam video or page of the PDF, and every output is marked AI-generated and pending your review.

That distinction matters even more in 2026. Generative AI has moved from novelty to regular use in document review, displacing TAR, according to Quinn Emanuel's analysis of AI in defensive discovery. And courts are increasingly willing to require production validation, a trend one e-discovery expert predicts will intensify "particularly with the advent of artificial intelligence," per HaystackID's review of 2025 decisions. If a court can sample your responsiveness calls, you need an audit trail, not a vibe.

The practical takeaway: you don't have to trust legal tech. You have to be able to audit it. Tools built for defense work, like Evidence Copilot's discovery intake by email, keep originals immutable and mark every AI output for human review, which is what Opinion 512 asks of you in the first place.

Implementing a Grounded Review Workflow

The gap between a discovery dump and a usable case file is where most review hours disappear. A grounded workflow closes that gap by tying every claim you make to a specific second of video or a specific page of a PDF, so nothing in your file rests on memory or a paraphrased note.

Start with intake. Forward the whole dump, attachments and all, into a single case file where bodycam video, 911 audio, jail calls, and scanned police reports get processed into searchable text. A structured discovery intake process matters because the raw materials are inconsistent: some PDFs have no text layer, video files arrive with meaningless filenames, and nothing is indexed. Once transcripts carry word-level timestamps and scanned documents are OCR'd, you can actually search across the entire case instead of scrubbing through footage by hand.

The core of the workflow is comparison. Set up a discovery review workflow that puts the police report side by side with the bodycam footage it describes, then look for the discrepancies that matter at a suppression hearing or trial:

• A report that describes "consent to search" while the bodycam captures an ambiguous or absent answer.
• Timeline claims in the report that don't match the timestamps on the video or the 911 call audio.
• Witness statements that shift between the report, the call audio, and later jail calls.
• Missing Miranda warnings on footage of what was clearly a custodial interrogation.

Tools can surface candidate inconsistencies automatically as new files come in, comparing each one against earlier evidence in the case. But treat those signals as leads, not conclusions. ABA Formal Opinion 512 requires attorneys to independently verify AI output before relying on it, and that duty fits this workflow naturally: every flagged inconsistency arrives with both sides quoted and cited, so verification means clicking through to the source rather than starting from scratch. That verification step is also where gen-AI review earns its keep over older TAR methods; as Quinn Emanuel's practice report notes, generative AI's advantage is the narrative reasoning it provides behind each call, which makes human review and validation practical rather than aspirational.

From there, build outward. Convert verified inconsistencies into timeline events, named-party records for each officer and witness, and cross-exam drafts where every suggested question is tied to a quoted source. When you draft the cross, the standard is simple: if a question references the report, it cites the page; if it references the footage, it seeks to the exact second. That discipline pays off in the courtroom and in the file itself, because a review memo built this way documents your work product and your diligence.

The stakes justify the rigor. A NACDL analysis found that 16% of reversals of U.S. capital convictions from 1973 to 1995 involved prosecutorial suppression of favorable evidence, and that Brady violations typically stem from administrative failures rather than malice. The same administrative logic applies on the defense side: a systematic workflow catches what memory and good intentions miss.

If you want to see the workflow end to end, Evidence Copilot offers a 20-minute walkthrough at evidencecopilot.com/book, or you can start on the Solo plan at evidencecopilot.com/signup.