How AI Is Quietly Changing the Medical Record (and Your Discovery Requests)

By John Mahoney · August 2026 · 12 min read

For the entire history of medical malpractice litigation, one assumption has sat underneath every chart review, every expert report, and every cross-examination: the note is the physician's account of the encounter. It might be wrong, self-serving, templated, or copied forward — but it originated with a human who was in the room, and everything about how we litigate charts flows from that.

That assumption is now false in a large and growing share of American medicine, and most discovery requests haven't noticed. Surveys compiled by Ona Health's 2026 adoption analysis put physician AI use at 81% in 2026, more than double the 38% recorded in 2023, with ambient documentation the leading use case. At UCSF, roughly 70% of physicians use AI scribes in daily practice. The American Hospital Association has been showcasing system-wide ambient-scribe rollouts since early 2026, and EHR vendors have moved from integration to bundling — Epic announced native AI charting in February 2026, with athenahealth and others shipping equivalents. This analysis is vendor-neutral because it has to be: the issue is structural, not a defect of any one product.

If your requests for production still just say "any and all medical records," you are litigating a 2019 chart in a 2026 world. Here is what changed, why it matters to provenance, and precisely what to add.

What an AI-Drafted Note Actually Is

An ambient scribe workflow looks like this: a phone or room microphone records the visit; the audio is transcribed; a large language model generates a draft clinical note from the transcript; the clinician reviews (in theory), edits (sometimes), and signs. The signed note lands in the chart looking exactly like every note before it.

Three properties of that pipeline matter enormously in litigation:

1. The note is a generated summary, not a contemporaneous account. An LLM decides what was clinically salient. Reported error rates for modern ambient scribes run around 1–3% of content — low, until you multiply by thousands of encounters — and the failure modes are distinctive: hallucinated findings (exams documented that never occurred), critical omissions (the symptom the patient mentioned that never made the note), misattribution (the daughter's history recorded as the patient's), and plausible-sounding contextual misreadings. An Ontario government audit that tested 20 approved AI scribe platforms found errors in every single one. Malpractice insurers are already warning their insureds about exactly this.

2. There are now intermediate artifacts. Between the encounter and the signed note sit an audio recording, a transcript, and one or more AI drafts. Each is a contemporaneous record of what actually happened in the room — often more contemporaneous than the signed note. Whether they were retained, for how long, and under what policy varies by vendor configuration and health-system choice. We flagged this in our earlier piece on AI scribe discovery; a year later, the practical point has sharpened: if you don't ask for the intermediate artifacts by name, nobody will volunteer them, and retention windows are short.

3. "Review and sign" is where the liability story lives. The physician's signature now attests to a document the physician may have skimmed for eleven seconds. Whether that happened is not a matter of testimony — it is a matter of metadata. Edit-time, edit-count, and time-in-note data exist inside the EHR, and the emerging consensus in the defense bar itself is that the only real defense to an AI-documentation error claim is a verifiable log showing the clinician actually reviewed the contested content. What defends their case, when absent, builds yours.

Provenance: The New Front in Chart Litigation

Plaintiff lawyers already know charts get cleaned up; that is why audit trail discovery exists. AI documentation adds a second provenance question that audit trails alone don't answer: not just who touched the note and when, but which parts of the note did a human ever write, read, or verify at all?

Consider the practical consequences:

The note's evidentiary weight changes. A jury told that the "physician's note" contradicting your client's account was machine-generated and signed after a nine-second review hears that note very differently. Conversely, if the physician materially edited the draft, the edits themselves are a window into what the physician believed mattered — and what they wanted removed.
The draft-to-final delta is discoverable gold. If the AI draft documented "patient reports chest pain radiating to jaw" and the signed note says "patient reports discomfort," you have found the case. That comparison is only possible if the draft still exists — retention is the fight.
Hallucinations cut both ways. A fabricated normal finding ("cranial nerves II–XII intact" in a visit where no neuro exam occurred) is a documentation-integrity problem for the defense that no amount of witness rehabilitation fixes. But be prepared for the mirror image: defendants will argue an unfavorable chart entry was machine error, not admission. Either way, provenance decides.

The strategic frame: for decades, the chart was the defense's home turf — their document, their custodian, their narrative. AI documentation splits "the chart" into a chain of artifacts created by different actors (patient, machine, clinician), and every link in that chain is separately discoverable and separately impeachable.

What to Add to Your Discovery Requests

Adapt to your jurisdiction and case; the categories below assume a post-2023 encounter at a system of any size. Serve them alongside your standard med-mal discovery checklist, not instead of it.

Interrogatories / deposition-on-written-questions topics

Identify every AI or automated documentation tool (ambient scribe, transcription, note-generation, coding-assist, chart-summarization, or clinical decision support) in use in the department(s) that treated the plaintiff during the relevant period, including vendor, product name, version, and deployment dates.
Identify the policies governing clinician review, editing, and signature of AI-generated documentation, and any training materials provided to clinicians on those tools.
State the retention configuration for encounter audio, transcripts, and AI-generated drafts, and whether such artifacts exist for the plaintiff's encounters.
Identify any error reports, quality audits, or incident tickets concerning the documentation tools during the relevant period.

Requests for production

For each encounter at issue: the audio recording, verbatim transcript, and every AI-generated draft version of each clinical note, in native format with metadata.
The complete audit trail / access log for the plaintiff's chart — including note-creation source (dictation, ambient tool, template, manual), edit history with timestamps, time-in-note data, and signature events. (Our audit-trail request guide covers the system-specific report names; pair it with the AI-specific items here.)
Vendor contracts, BAAs, and configuration documents for the documentation tools, including any settings governing draft retention or deletion.
Policies, protocols, and committee minutes concerning adoption, validation, and monitoring of AI documentation tools.

Preservation letter — send it first

Because vendor-side retention of audio and drafts can be measured in days or weeks, your spoliation groundwork now starts pre-suit. The preservation letter should name encounter audio, transcripts, AI draft versions, and tool-usage logs explicitly, and should go to the provider before the records request tips them off to the theory. If those artifacts were routinely destroyed after a proper preservation demand, you have an adverse-inference argument that didn't exist in the paper-chart world.

Don't forget the consent trail

Ambient scribes record the encounter, and most health systems obtain some form of patient consent — a signage notice, an intake form checkbox, or a verbal acknowledgment the tool itself logs. That consent documentation is discoverable and doubly useful. First, it tells you the tool was in use for your client's visits even when the notes don't say so (most signed notes carry no AI attribution at all). Second, in two-party-consent states, a recording made without documented consent is its own issue — and a provider's position that "no recording was retained" sits awkwardly next to a consent form promising the patient the audio would be handled under a specific retention policy. Request the consent artifacts alongside the recordings themselves.

One more regulatory hook worth knowing: HIPAA's Security Rule audit-control requirement (45 C.F.R. § 164.312(b)) is why the metadata you are asking for must exist, and the federal push for algorithm transparency in certified health IT gives you a straight-faced answer to "that information is proprietary." You are not asking for the vendor's model weights. You are asking what the tool wrote, what the clinician changed, and whether anyone looked — operational records of the defendant's own documentation process.

Deposing on AI Use

The deposition sequence writes itself once the produced metadata is in hand. The themes:

Establish the pipeline. "Walk me through how this note came to exist." Most physicians will freely explain the scribe workflow — it is not a secret, and they often like the tool.
Establish the attestation. "When you signed this note, you were attesting to its accuracy?" Yes is the only available answer.
Confront with the metadata. Time-in-note, edit count, signature timestamp. "The audit trail shows the note was open for fourteen seconds before you signed. What did you review in fourteen seconds?"
Walk the delta. Where drafts survive: "The AI draft said X. The signed note says Y. Who changed it, and why?"
Lock in the policy gap. "The health system's policy required verification of AI-generated content before signature. Sitting here today, you cannot tell me you verified this content — correct?"

Note the posture: you are not arguing AI is bad medicine. You are arguing the signature is the physician's, the duty to verify is the physician's, and the record shows the duty was not discharged. That framing survives whatever the defense says about the tool's accuracy, and it keeps a technology-friendly jury on your side.

What This Means for Your Chart Review

All of this raises the bar on the plaintiff side too. A chart that mixes human-authored notes, AI-drafted notes, copy-forward blocks, and template text — across hundreds or thousands of pages — is beyond honest manual review at intake pricing. Contradictions between the note and the rest of the record (meds, orders, nursing flowsheets, vitals) are precisely the signature of generated text that nobody verified, and finding them requires reading everything against everything. That is a machine's job; deciding what it means is yours. It is also why AI-era charts make the economics of manual merit review worse every year.

Read the chart the way you'll litigate it

MedLegal AI cross-checks every note against meds, orders, labs, and flowsheets — every finding linked to the exact Bates page — and flags the internal contradictions that mark unverified documentation.

Analyze a case →

The Bottom Line

The medical record quietly stopped being a single human-authored document, and discovery practice has to stop pretending otherwise. Ask what tools drafted the chart. Demand the intermediate artifacts before retention windows eat them. Get the metadata that shows whether anyone actually read what they signed. The firms that update their templates this year will spend the next decade impeaching charts their opponents still take at face value.