Most US hospitals cannot tell you which AI model wrote which sentence in the signed note.
The EHR vendor has not published hallucination rate, error type, severity, or risk disclosure for the summary and scribe features that draft clinical content. That gap is itself part of the legal record. Idem Labs reconstructs AI behavior in the medical record for in-house counsel and outside trial counsel.
AI-generated content is in the legal medical record. The audit trail often does not distinguish it from clinician edits.
Medical AI is producing content that ends up signed by clinicians and pulled into discovery. The hospital's informatics team cannot easily reconstruct what the AI did at the time of the encounter. The audit trail frequently fails to preserve which sentence was AI-drafted, which was clinician-edited, and which model version produced it.
Four threshold questions surface in any matter involving AI-drafted clinical content. Which model and version generated which sentence. What did the model omit from the underlying chart. Did the feature carry a Predictive DSI Applied Model Card under HTI-1. Where is the vendor's evaluation methodology, and what did it actually test for. Idem Labs answers all four on the matter you are working.
In May 2026, Ontario's Auditor General audited all 20 AI scribes approved for clinical use in the province. Every one had inaccuracies. Twelve captured the wrong drug. Seventeen missed important mental-health details. The audit is the strongest publicly available signal for the evidentiary record now sitting under every US hospital's AI deployment.
Which model and version generated which sentence?
What did the model omit from the underlying chart?
Did the feature carry a Predictive DSI Applied Model Card under HTI-1?
Where is the vendor's evaluation methodology, and what did it actually test for?
Reconstruction, review, and risk assessment for AI in the medical record.
A scoped engagement covering identification of the AI model and version active at a given encounter or across a feature deployment, structured testing against the underlying source records, omission and hallucination analysis, vendor disclosure inventory, and a written report sized to its use. For active matters, the report is suitable for expert designation and deposition. For in-house pre-incident work, it supports board risk reporting, vendor contract negotiation, regulatory inquiry response, and incident readiness planning.
Outside counsel engagements run under written conflict checks; defense and plaintiff retain Idem under separate engagement letters, and Idem does not take both sides of the same dispute. In-house engagements are typically structured under Kovel arrangement to preserve attorney-client privilege.
Who it is for
In-house General Counsel, Deputy General Counsel, and AVP Legal at health systems with material AI exposure. Trial attorneys handling medical malpractice matters involving AI-drafted clinical content. Particular fit when the matter or the deployment involves EHR-native summarization features (Epic Art, Oracle Clinical AI Agent, Meditech Expanse) or AI scribe output that has become part of the signed record.
Read the evidence.
Then let's talk.
The June 2026 white paper documents what EHR-native AI features are doing in the medical record and what your organization is on the hook for.