Live Deposition AI With Real-Time PubMed Lookups: A Med-Mal Cross-Exam Edge

May 10, 2026 · MedLegal AI Editorial · 7 min read

The defense expert, a board-certified pulmonologist, has just told you that the published literature shows initial-presentation D-dimer is "an unreliable rule-out for pulmonary embolism in patients over 50." She's cited a 2019 cohort study by name. You have maybe 90 seconds before your next question lands. Your fellow at the table has the deposition outline open and a Westlaw tab she can't search fast enough.

This is the moment most plaintiff cross-exams quietly lose ground — not because the lawyer is unprepared, but because the literature is doing its work in the witness's pocket and not in yours.

This article is about a specific workflow we built into MedLegal AI to close that gap, and a frank look at what it does, what it doesn't, and why no other legal-AI vendor we know of ships it today.

The 90-second problem

In a med-mal deposition, the expert witness's strongest move is the literature pivot. They cite a study, summarize it favorably to their position, and dare you to challenge the methodology. The honest plaintiff response requires three things in the same minute:

1. The actual paper. Title, authors, journal, year. PMID if you can get it.
2. The methods + sample size. Was it 200 patients or 200,000? Retrospective chart review or prospective trial? Single-center or multi-center?
3. The contradicting literature. What does the 2021–2024 follow-up work say? Did a meta-analysis come out since? Has any guideline body changed its recommendation?

That's a 30-minute paralegal job under normal circumstances, and you have 90 seconds while the expert is mid-monologue. The realistic options today are: (a) interrupt and redirect ("Doctor, we'll come back to that"), (b) take the citation as gospel and lose the round, or (c) burn a recess to research, which the defense knows means you weren't ready.

That sequence didn't require the lawyer to be a pulmonologist. It required the literature to be reachable in the lawyer's hand the moment the expert opens the door.

What the workflow actually does

MedLegal AI's Courtroom AI module runs a live transcription of the deposition through Deepgram, segments speakers (witness vs counsel), and pipes the witness's answer through a Claude-powered listener that does three things:

• Catches medical claims. Any time the witness asserts a diagnostic threshold, treatment efficacy, mortality rate, or guideline statement, the system extracts the claim as a structured assertion.
• Catches paper citations. If the witness names a study, journal, or author, the citation is captured.
• Triggers a live PubMed query. The extracted assertion + citation become a PubMed search built around the claim's clinical entity. Results return in roughly 6–12 seconds with title, authors, year, abstract, and a flag for whether the cited paper exists, supports the claim as stated, or contradicts it.

The output appears in a side panel labeled "live literature." It tags each result as supports witness, qualifies witness, or contradicts witness. The lawyer reads the side panel between questions.

The value is not "the AI gives you the answer." The value is "the AI eliminates the 30-minute paralegal step so you can stay in the room."

Why this matters more for plaintiff than defense

Defense counsel typically has the expert on retainer. They've prepped the witness on the literature ahead of time. They know which 6 papers will be cited and have rebuttals ready.

Plaintiff counsel almost never has this advantage. The expert is the defense's expert. The literature they'll cite is unknown until they cite it. Live lookup is asymmetric — it gives the plaintiff, who has been at the information disadvantage, the same real-time access to the medical canon that the witness has had since med school.

What it doesn't do

Three honest limits worth stating up front:

1. It's not a substitute for an expert of your own. The system flags when the witness has overstated a paper's conclusion or cited a paper that doesn't say what they say it says. That's a Daubert flag, not a clinical opinion. You still need your own expert to put the contradicting view into the record.
2. It depends on the witness naming the citation. If they vaguely reference "the literature shows," the system tries to map it to the most-likely paper given the clinical context, but that's a guess. Vague-citation answers are a separate cross-exam play (we have a different tool for those).
3. PubMed is not Westlaw. Open-access medicine ≠ paywalled subspecialty journals. About 11% of top-tier surgical journals are not in PubMed's open-access index. The system flags when a witness cites an out-of-PubMed source so you know to flag it for a follow-up.

Why no one else ships this

The legal-AI vendor landscape in 2026 is full of products that read documents, draft demand letters, and analyze chronologies after the fact. Live courtroom or deposition workflow is harder for three reasons that aren't immediately obvious:

• Latency. A research lookup that returns in 45 seconds is useless during a live exchange. Sub-15-second response time requires a custom retrieval index, not a general-purpose RAG.
• Audio reliability. Most deposition rooms aren't acoustically engineered. Speaker diarization that works on a podcast collapses when two attorneys talk over each other across a glass conference table.
• Asymmetric value. Defense firms, which buy the most legal-AI today, don't need this capability — they're prepped. Building it for the plaintiff side requires a vendor that has decided plaintiff is the customer. Almost no major legal-AI startup has made that bet.

The competitive landscape today (May 2026): Casetext / CoCounsel does excellent post-hoc deposition summary and cross-exam preparation. Supio is great at the chronology + demand-letter handoff. EvenUp dominates demand-letter generation in PI. Spellbook serves transactional, not litigation. None of them — to our knowledge — runs real-time medical-literature retrieval inside a live deposition. Filevine's Depo CoPilot, the closest analog, does live transcription but does not pipe the audio through a medical-claim extractor or PubMed retriever.

How to try it

The Courtroom AI add-on is available on any MedLegal AI subscription tier. Free trial is 14 days, no credit card. The base subscription includes the deposition preparation tools (cross-exam question bank, expert vetter, witness-search that finds Bates-cited statements across the case records you've uploaded). Live deposition transcription with real-time literature is the Courtroom AI add-on.

If you want to see what live lookup feels like before committing to a deposition, the free Daubert workup tool uses the same underlying retriever — paste a deposition transcript and it generates the same kind of cross-exam-ready analysis after the fact, without the live audio piece.