AI Scribe Discovery in Medical Malpractice: Why Every Post-2024 Complaint Should Ask

Roughly 42% of US medical groups now run an ambient AI scribe during patient encounters. The scribe produces a draft note. The physician edits and signs. The draft and the signed note frequently diverge — and the draft is rarely preserved. Most med-mal complaints filed today don't ask for it. They should.

The quiet shift in how patient encounters are documented

Three years ago, when a doctor saw a patient, the resulting clinical note was — give or take a dictation pass — a record of what the doctor wrote. Today, in a meaningful share of US medical groups, the resulting note is a record of what an AI drafted and the doctor edited.

That's not a metaphor. It's an architectural change in clinical documentation, and it happened fast. KLAS data, AMA member surveys, and Elion Health's vendor census all converge on roughly the same number for early 2026: about 42% of US medical groups are running an ambient AI scribe of some kind, up from under 10% at the end of 2023. The five vendors that dominate the market — Abridge, Microsoft DAX Copilot, Suki, Nuance/Microsoft, and Augmedix — together count most major health systems among their customers.

The clinical workflow looks like this: the physician opens the encounter, the ambient scribe listens (with patient consent, in theory), the AI produces a structured draft note seconds after the visit ends, and the physician reviews, edits, and signs. The signed note is the one that lives in the EHR. The draft typically doesn't.

The dual-record problem

That divergence — between what the AI heard and structured into a draft, and what the physician edited and signed — is the single most important fact for plaintiff attorneys to absorb about this technology.

Two records exist. They are records of the same encounter, generated minutes apart, by different processes. They can disagree.

They can disagree about the chief complaint. They can disagree about the review of systems — about whether the patient denied chest pain or never said anything about it at all. They can disagree about the assessment, about the differential the physician initially considered, about the workup the AI suggested and the physician rejected, about the disposition and the return precautions and what the patient was told.

In ordinary cases, that disagreement doesn't matter. The physician edits the AI draft to match clinical reality, signs an accurate note, and the draft is reasonably treated as a working artifact — like the back-of-the-envelope calculation an engineer throws away after producing the final design.

In malpractice cases, that disagreement is potentially case-deciding. If the AI captured the patient's complaint of progressive dyspnea on Day 1, and the physician's signed note documents a chief complaint of "fatigue," and the patient died of a pulmonary embolism on Day 4, the draft is no longer a working artifact. It's the contemporaneous record of what the patient actually said.

What plaintiff complaints filed today miss

A modest informal review of recent med-mal complaints suggests that the standard discovery package — medical records, audit trails, billing records, custodian declarations — does not, today, ask for AI scribe artifacts. The interrogatories and requests for production track templates that predate ambient scribes. They ask for the chart. They don't ask whether anything contributed to the chart that isn't in the chart.

This is a fixable gap. The discovery requests required to surface AI scribe artifacts are not exotic. They are extensions of the kind of metadata and audit-log discovery that EHR-savvy plaintiff firms have been doing for a decade. They require:

• A direct interrogatory asking whether any AI scribe was used and identifying it by manufacturer, model, and version.
• A direct interrogatory asking whether any audio of the encounter was retained, where, and under what retention policy.
• A request for production seeking all draft, intermediate, and pre-edit versions of any provider note in connection with the plaintiff's care.
• A request for production seeking the BAA, master service agreement, and configuration documentation for the AI scribe vendor — because retention is a contractual variable, not a vendor default.
• A parallel third-party subpoena to the vendor itself, because the vendor often retains records the hospital does not.

Our free AI Scribe Discovery Checklist includes paste-ready language for each of those requests, with the five major vendors mapped to their parent companies, EHR integrations, retention defaults, and subpoena addresses.

Sharp Rees-Stealy and the spoliation theory

The first prominent filed pleading raising this theory is the Sharp Rees-Stealy AI Scribe Litigation, a putative class action filed in the Southern District of California in January 2026. The plaintiffs allege that the health system's deployment of an ambient AI scribe produced systematically inaccurate notes — that what the AI captured of the encounter and what the patient experienced of the encounter were one thing, and that what was signed into the chart was another, and that the drafts demonstrating the difference were either never preserved or purged on a short retention window.

The case is in early pleadings as of April 2026 and has not produced merits rulings. Whether it survives a motion to dismiss, whether the class is certified, whether the alleged divergence pattern can be proven on classwide proof — all of that is still ahead. But the pleading itself is a marker. It frames the dual-record problem in spoliation terms. And spoliation is a plaintiff's-bar specialty.

The spoliation theory, in short: once foreseeability of suit attaches — typically once the patient or family raises a complaint, requests records, or retains counsel — the defendant has a duty to preserve relevant evidence. If the AI draft is relevant evidence (and the dual-record argument is that it is) and the defendant's vendor configuration causes drafts to be auto-purged on a 24-hour or 30-day window, then a defendant who fails to halt that purge after foreseeability attaches has spoliated. The remedies range from adverse-inference instructions to outright sanction.

This theory is not yet settled law. Whether the AI draft qualifies as evidence the defendant had a duty to preserve will depend on jurisdictional rules of discovery, the timing of any litigation hold, and the technical specifics of the configuration. But the theory is plausible enough that plaintiff firms working malpractice in 2026 should be developing the factual record to support it: serving preservation letters that name AI scribe artifacts explicitly, deposing IT custodians about retention defaults, and getting the BAA on the record.

The ABA, OCR, and the regulatory backdrop

Three sources of authority bear on the discoverability of AI scribe artifacts.

ABA Formal Opinion 512 (2024) addresses lawyers' duties when using generative AI. Its commentary anticipates that the same dual-record problem will arise on the healthcare side and that opposing counsel will need to discover AI inputs and outputs separately from final work product. The opinion does not bind healthcare defendants, but plaintiff briefs cite it for the persuasive proposition that AI drafts are independently discoverable.

HHS / OCR HIPAA guidance has confirmed that AI scribe vendors are HIPAA business associates and that audio, transcripts, and drafts are PHI subject to the BAA's retention and access provisions. This is the foundation for asserting that those records exist within the covered entity's HIPAA "designated record set" and are accessible — both by the patient via a HIPAA right-of-access request and by litigants via discovery.

State medical boards (California, Texas, Florida, New York among them) have issued informal guidance requiring physicians to review and attest to AI-generated documentation before signing. This is a standard-of-care benchmark. A physician who signed an AI draft without meaningful review is, on the boards' own articulation, doing something the profession does not endorse. That's a deposition question.

What plaintiff attorneys should be doing now

Three things, immediately, for any med-mal case involving care delivered in 2024 or later:

One. When you send your preservation letter, name AI scribe artifacts explicitly. List the leading vendors. Demand preservation of audio, transcripts, drafts, audit logs, and vendor correspondence. The retention windows on these artifacts are often very short; a generic "preserve all relevant records" letter may not be enough to put an IT department on notice that the AI scribe configuration needs to be paused.

Two. Add the AI scribe interrogatories and requests for production to your standard medical discovery set. The volume is small — three or four interrogatories and three requests for production — and it doesn't displace anything. It runs in parallel.

Three. When you depose the treating physician, ask. Did you use an AI scribe? Which one? How did you review the draft? What changes did you make? The answers — and the body language, and the cross-examination opportunities they create — are some of the cleanest deposition material available in modern med-mal practice. It costs you nothing to ask. The downside of not asking, in a case where the AI draft would have helped your client, is real.

The edge case that's about to become a problem

The hardest case for the plaintiff bar is the configuration where the AI draft is auto-purged within 24 hours of physician sign-off. Several enterprise contracts default to exactly that retention. In those cases, the draft is gone before any litigation hold can attach — often before the patient has even left the building.

That's not a discovery problem; it's an evidence-preservation problem. The remaining proof is the audit-log entry showing a draft existed plus circumstantial proof of the configuration. The question — likely to be litigated in 2026 and 2027 — is whether such a configuration, deployed in a regulated industry where suit is foreseeable, itself constitutes a form of pre-litigation spoliation. Courts have not answered yet. Plaintiff briefs are starting to ask.

The bigger picture

The arrival of ambient AI scribes is the largest change in clinical documentation since the EHR mandate. It is happening faster than any prior technology adoption in healthcare. It is happening with very limited regulatory friction, no FDA premarket review, and very limited public attention.

From the plaintiff bar's perspective, that's both a problem and an opportunity. It's a problem because every contemporaneous record of what actually happened in the encounter is now mediated through a system that can be misconfigured, can be over-edited, can hallucinate, and can purge its own drafts. It's an opportunity because — for the cases where the draft is preserved or the retention window is long enough to hit a litigation hold — there is now, for the first time, a record of the encounter that is more complete than the physician's signed note, that was not authored by the defendant, and that can corroborate or contradict the patient's testimony.

The plaintiff firms that adapt their discovery practice to this technology in 2026 will have an advantage over the firms that don't. The firms that don't will, in a few years, have signed off on settlements and verdicts in cases where the answer was sitting in a vendor's cloud storage all along.