Why You Can't Scrub Through Discovery Anymore

A client hands you a USB drive with 40 hours of bodycam. You have a preliminary hearing in nine days. Ten years ago that discovery would have been a police report; today it's a stack of files named X13qrt9__x3455 that you're expected to review on top of everything else on your desk.

The volume problem is not anecdote — it's measured. Public defenders handled 45x more digital evidence in 2024 than in 2016, according to industry reporting on defender workloads. And in Virginia, where 80% of criminal defense cases involve digital evidence, 73% of public defenders surveyed by the state Indigent Defense Commission said reviewing bodycam footage sometimes crowded out other case work entirely (StateScoop). Virginia defenders carry up to 100 cases at a time; the footage keeps arriving anyway.

So the real question isn't whether to transcribe your discovery video. It's how to get a searchable, citable transcript without burning the hours you don't have. Scrubbing through a three-hour file looking for the two minutes that matter is the task that eats your week. As one R Street Institute analyst put it, the goal is to "find the two minutes you're looking for without having to scrub through a three-hour video file."

What actually works looks like this:

• Every video file gets a timestamped, diarized transcript — so "search the footage" becomes "search the text."
• Speakers are labeled correctly, because a transcript that merges the officer's statement with the civilian's is worse than none at all.
• Every quote you pull for a cross-exam question or a suppression motion traces back to the exact second of video it came from.
• A human reviews the machine's work before anything gets relied on — because the machine gets things wrong in ways that matter.

That last point isn't caution for its own sake. A Northeastern study of 176 bodycam videos found AI transcription identified an average of 3.3 participants per incident versus 4.8 in human-edited versions, and dropped roughly 25,000 words across the sample — words that might contain the explanation or warning that explains why an encounter escalated (Northeastern Global News). Automated transcription can roughly cut review time in half, but only if the review step stays human.

That's the workflow Evidence Copilot is built around: upload the discovery dump, get a word-level timestamped transcript, and let an evidence agent answer questions with citations that jump to the exact video moment — with every AI output marked pending human review, because the file prep is the machine's job and the judgment is yours.

What AI Transcription Gets Right — and Where It Breaks

Before you trust an AI transcript with anything that matters — a suppression motion, a cross-exam, a Brady review — it's worth knowing exactly where the technology holds up and where it quietly falls apart. The honest answer: AI transcription is genuinely good at getting most of the words down fast, and genuinely bad at keeping track of who said them.

The clearest evidence comes from a Northeastern study of 176 body camera videos covering roughly 23 hours and 73 incidents. Researchers found the AI identified an average of 3.3 participants per transcript, compared with 4.8 in human-edited versions — and produced about 25,000 fewer words across the sample. The AI "struggled to keep the speakers straight when they switched turns," according to the university's research team. In practice, that means an officer's statement and a civilian's response can get merged into one speaker, and words get dropped entirely.

Those dropped words aren't cosmetic. The researchers noted that omitted text could mean losing an explanation, a warning, or a response that helps establish why an encounter escalated — exactly the material you'd build a cross-examination or a suppression argument around.

Calli Schroeder of the Electronic Privacy Information Center put the stakes plainly: "An AI transcription or summary may be 98% accurate, but that 2% could completely change who is at fault and what occurred in an encounter." For defense work, that 2% is often the whole case — a consent statement, a Miranda warning, a retracted identification.

That's why the workflow rule matters more than the tool you pick. Shellie Solomon of Justice & Security Strategies framed it well: think of AI as the investigative assistant who prepares the file — the humans decide what the details mean. AI gets you a searchable, timestamped draft in minutes instead of hours. It does not get you a record you can cite in court without reading it.

When you run transcription in a workspace like Evidence Copilot, that division of labor is built into the output: every transcript is marked AI-generated and pending human review, speaker labels are one-click renamable across every segment, and every AI answer cites the exact video second it came from — so your review happens against the source, not against the AI's summary of the source.

A few things worth checking on every AI transcript before you rely on it:

• Speaker attribution, especially at turn changes — the study's clearest failure point
• Missing words and short exchanges, which account for the word-count gap
• Anything you plan to quote in a motion, brief, or cross-exam outline
• Segments where multiple people talk over each other or audio is degraded

The takeaway isn't "don't use AI transcription." With public defenders handling 45x more digital evidence in 2024 than in 2016, manual review of every bodycam file isn't realistic anymore. The takeaway is that AI prepares the file; you decide what it means — and any workflow that skips the second half isn't ready for a criminal case.

Step-by-Step: Upload and Transcribe a Full Video in Evidence Copilot

The Northeastern diarization study found AI transcripts identified an average of 3.3 speakers per bodycam video versus 4.8 in human-edited versions — which is exactly why the review step in your transcription workflow matters as much as the transcription itself. Here's the full Evidence Copilot workflow, from upload to cited transcript.

Step 1: Create a case. Start a new case in your workspace and name it for the matter — preliminary-hearing prep, suppression motion, whatever you're building toward. Each case gets its own evidence set, activity feed, and members with assigned roles, so your investigator and law clerk see the same files you do.

Step 2: Upload the discovery dump. Drag in bodycam video, interrogation footage, jail calls, 911 audio, police reports, and scanned documents. Files up to 2 GB upload through the web interface, or you can forward attachments to the case's dedicated inbound email address — useful when discovery arrives as a sprawling email chain from the prosecutor's office.

Step 3: Let the transcription run. Evidence Copilot uses Deepgram Nova to generate a transcript with speaker diarization and word-level timestamps. Every "Officer 1" and "Subject" segment is tied to the exact moment it occurred, not just a rough paragraph.

Step 4: Fix the speaker labels — once. This is where the diarization research matters. AI systems "struggled to keep the speakers straight when they switched turns," per the Northeastern researchers. In Evidence Copilot, you rename a speaker one time and the change propagates across every segment in the transcript — no find-and-replace marathon across a three-hour interrogation.

Step 5: Search and cite. The AI evidence agent answers questions across the full transcript and returns citations that seek the exact video second. Ask "when does the officer first mention consent to search?" and you get a clickable citation that plays the moment, not a paraphrase you have to hunt down.

Step 6: Review before you rely on it. Every AI output in Evidence Copilot is marked AI-generated and pending human review, and the originals stay immutable. That posture reflects what transcription researchers and ethicists keep saying: a transcript that's 98% accurate can still misstate the 2% that changes the encounter. Treat the machine output as a working draft, then verify the passages you plan to quote.

For context on why this workflow pays for itself: a Virginia Indigent Defense Commission survey found 73% of public defenders said reviewing bodycam footage sometimes prevented them from completing other case work, and a 2021 pilot with 100+ attorneys and staff suggested automated transcription could cut evidence-review time roughly in half. The transcription isn't the verdict — it's the file preparation that gets you to the two minutes you actually need.

Review, Verify, and Use the Transcript the Right Way

The transcript is done, but you're not. AI transcription gets you searchable text in minutes — not a record you can put in front of a judge. Treat what comes out of the pipeline as a working draft that you, the person with a law license, verify against the actual footage.

The reason is empirical, not cautionary boilerplate. A study of 176 body camera videos (~23 hours across 73 incidents) found AI transcripts identified an average of 3.3 participants per video versus 4.8 in human-edited versions, and produced roughly 25,000 fewer words across the sample. Those dropped words aren't typos — they can be the warning, the explanation, or the response that explains why an encounter escalated. As Calli Schroeder of EPIC put it, a transcript can be 98% accurate, but that 2% "could completely change who is at fault and what occurred."

So build review into the workflow as a distinct step:

• Spot-check speaker labels against the raw footage. Diarization is the known weak point — AI "struggled to keep the speakers straight when they switched turns," so confirm who said what before you rely on any attribution.
• Listen for omissions. Play the segments around key statements and confirm nothing was dropped — especially overlapping speech, muffled audio, and speech off-camera.
• Verify every quote you plan to use. Word-level timestamps let you jump to the exact second and confirm the quote before it goes into a cross-exam outline or a suppression motion.
• Check context. Automated transcripts put "a layer of interpretation between the recording and the record," as one R Street Institute analyst observed — watch for automation bias in your own review.

That last point matters most when the transcript feeds a motion to suppress. If your consent argument or your stop-duration calculation rests on a quoted line, the citation needs to resolve to the actual moment in the recording, not to a paragraph of AI text. This is where Evidence Copilot's citation model earns its keep: every AI answer carries a clickable citation that seeks the exact second of video, and every AI output is marked AI-generated and pending human review. You click, the video seeks to that moment, and you confirm the quote with your own ears before it goes anywhere near a filing.

One boundary to keep clean: this workflow produces internal work product, not the official record. Court reporters and certified transcripts still govern when you need a transcript for the record itself — AI handles evidence review and case prep, and human court reporters handle certification. The two are complementary, not substitutes.

The payoff for this discipline is real. Automated transcription has been shown to deliver roughly 50% time savings in defender workflows — but only when paired with verification you can defend. AI prepares the file; you decide what the details mean. If you want to see the citation-to-second workflow on your own footage, book a 20-minute walkthrough at evidencecopilot.com/book, or start on the Solo plan at evidencecopilot.com/signup.

Choosing a Tool Without Getting Burned

Every transcription vendor publishes an accuracy benchmark, and none of them matter for your cases. Speechmatics' own guidance to law firms is blunt: test transcription systems against your own representative legal audio rather than trusting vendor scores. So before you commit to anything, run a real file through it — a jail call with crosstalk, a bodycam with wind noise, an interrogation where two people talk over each other.

The stakes are real. A Northeastern study of 176 body camera videos found AI identified an average of 3.3 speakers per transcript versus 4.8 in human-edited versions, and dropped roughly 25,000 words across the sample. As EPIC's Calli Schroeder put it, a transcript that is 98% accurate still leaves a 2% that "could completely change who is at fault." Merged speaker turns aren't a cosmetic error — they can merge an officer's command with your client's response.

The free DIY stack — local Whisper plus NotebookLM — is where most cost-conscious defenders start, and it genuinely works for one short file. It breaks at case scale:

• No diarization QC — Whisper's speaker labels can't be renamed and propagated, so you can't verify who said what.
• No citations — an AI summary can't point you to the exact video second for impeachment.
• No redaction — nothing for bystander faces, minors, or PII before discovery turnover.
• No audit trail — no record of what was reviewed, by whom, or what the AI touched.

Evidence Copilot exists as the alternative for defenders who need those four things. Every AI answer carries a clickable citation into the exact second of video; redaction is frame-accurate and never alters the original; and ingest auto-signals flag candidate inconsistencies — a police report that doesn't match the bodycam — before anyone asks. Every AI output is marked pending human review, which matches what the Northeastern researchers conclude anyway: AI prepares the file, the human decides what it means.

Pricing is published, not gated behind a sales call: Solo at $39/user/month with 15 media-processing hours, Firm at $99/user/month, and a PD Office tier at $750/attorney/year — roughly 38% below JusticeText's reported ~$1,200, with overage hard-capped at 3x included hours so runaway intake can't overbill. Book a 20-minute walkthrough at evidencecopilot.com/book, or start on Solo at evidencecopilot.com/signup and run it against your own audio first.