AI in Medical-Record Review: The "Misgrounding" Problem the Hallucination Warnings Miss

By John Mahoney · June 2026 · 7 min read

Every med-mal practice is now being pitched AI that turns boxes of medical records into a chronology in minutes. The promise is real. So is the risk — and the risk everyone warns about is the wrong one.

The headline fear is hallucination: the model invents a case, a quote, or a study that does not exist. By now most firms have a rule against filing AI-drafted briefs without checking the citations. Good. But in record review, the dangerous failure is quieter, and your "did it make up a case?" check sails right past it.

Disclaimer: This article is for informational purposes only and is not legal advice. Professional-responsibility rules and the standard for relying on technology are jurisdiction-specific; confirm the rules that govern you.

The failure mode isn't fabrication — it's misgrounding

The quiet failure isn't an invented citation. It's a real one that doesn't say what the AI claims it says.

The tool writes: "On 3/14 the patient denied chest pain (p.212)." Page 212 is a real page of the chart. But it never mentions chest pain — or it says the opposite. The fact looks sourced. The page number is right there. Nothing trips your fabrication check. And it slides into your chronology, your expert's report, or your motion, until the one person motivated to read page 212 finds it: opposing counsel.

There's a name for this: misgrounding — a citation that exists but does not support the statement attached to it. It is the quiet way AI loses cases. Independent testing has repeatedly found that leading legal-AI tools produce unsupported or incorrect outputs a meaningful share of the time — widely reported in the 17–33% range — even when marketed as "hallucination-free." The operative point isn't the exact percentage; it's that the error rate is not zero, the errors are plausible on their face, and in record review they hide inside citations that look legitimate.

Why record review is where it bites hardest

A chronology or causation summary is built from thousands of pages. You cannot re-read the whole chart to check every line — that defeats the point of the tool. So the practical question for any AI is not "is it fast?" It's "can I verify it without re-reading the record?" If a tool can't show you the exact page behind each fact, every assertion it makes is a hypothesis you're taking on faith. On contingency or on a defense panel, that's a liability with a deadline.

Verification is now a duty, not a nicety

The professional-responsibility ground has shifted to meet this. ABA Formal Opinion 512 (2024) makes clear that the duty of competence extends to understanding the AI tools you use and verifying their output; candor and diligence do the rest. Several state bars have issued parallel guidance. The throughline: a lawyer remains fully responsible for every factual assertion, regardless of which tool produced it. "The software said so" is not a defense to a Rule 11 problem — and the running tally of AI-hallucination sanctions cases is now well over a thousand.

For the work product itself, verifiability is also a strategic asset. A chronology you can defend line-by-line survives a deposition. A black-box summary you can't trace does not — and it puts your expert's and your own credibility on a fact you can't stand behind.

A verifiability checklist for AI in record review

The goal isn't to avoid AI. It's to insist that AI in record review be verifiable by construction, so checking is a fast click instead of a re-read. Five working rules:

1. Demand page-level citations on every assertion. Not "the records show" — the specific Bates page (ideally the line) for each fact, quote, and date. If a tool can't cite to the page, it can't be relied on for record review.
2. Spot-check for misgrounding, not just fabrication. Pull a sample of the AI's citations and confirm the cited page actually supports the statement. You're testing grounding, not existence — and you weight the sample toward the facts your case depends on.
3. Treat uncited assertions as unverified. Anything offered without a traceable source is a hypothesis. Flag it; source it manually or drop it.
4. Watch the gaps the AI surfaces. A good review names what's missing — the absent record, the un-obtained study — as clearly as what's present. Those are subpoena targets; surface them before the deposition, not after.
5. Keep the human name on the work product. The chronology and the analysis still carry your name — or your nurse-consultant's credentials. The tool removes the mechanical labor; it does not remove your responsibility, and it shouldn't remove your ability to defend every line.

What it means for each side of the table

• Plaintiff attorneys — screen and work up more cases without staking a theory on a fact you can't trace. The efficiency is margin; the traceability is protection.
• Defense attorneys — an unverifiable AI summary is asymmetric risk: a misgrounded fact in your chronology or your expert's premises is a self-inflicted impeachment exhibit. Verifiable review is the way to capture the speed without inheriting the failure mode.
• Legal nurse consultants — the tool collapses the 20–40 hours of record-flipping into the first-pass timeline, with every entry traceable, so the chronology still carries your name and you can defend every line in a deposition. It supercharges the judgment only you provide; it doesn't replace it.
• Expert witnesses — an opinion is only as strong as its weakest factual premise. Verifiable, page-cited inputs are what let your conclusions survive cross.

Bottom line

AI belongs in the med-mal toolkit. But adopt it the way you'd adopt any new source of factual assertions in litigation: make it show its work. Insist on page-level citations, test for misgrounding rather than just fabrication, treat the uncited as unproven, and keep human judgment — and the human name — on the output. Verifiability isn't a constraint on using AI well. It's the whole point.

Questions? Contact us at [email protected] or (856) 979-6525