Dragon Copilot 6.0 and the Next Test of AI Record Transparency

Vendor documentationUpdated 10 Nov 2026No local deployment inspected

A dated companion to When the AI Scribe Starts Shaping the Medical Record. The original investigation and its earlier observations are unchanged; this report adds the November developments beside them.

What the sources establish

Microsoft’s Dragon Copilot 6.0 release documentation, updated November 10, describes Show AI. Its nursing simulation is a US public preview on supported web and iOS versions, disabled by default and enabled by administrators. [1]

The interface version table identifies Show AI, custom-template version history and immediate adaptation to feedback as 6.0 additions. Documentation does not establish which release or features a particular organization has activated. [2]

Current Microsoft support material says an additional ambient recording generates a version from the combined transcript without carrying forward prior edits or style changes; earlier versions can be restored. This is supporting workflow documentation, not a separately established November release event. [3]

Original analysis

For an investigation of medical-record integrity, new controls create a concrete opportunity to test reviewability. The evaluation should acknowledge improvements and ask what each control actually enables a user to establish. A documented feature can change the appropriate question without settling every concern about the final record.

Begin with a short availability record. Name the product, release, user role, EHR integration, enabled features and test date. “We use Dragon” is too broad to determine which controls a particular clinician can access. Likewise, the nursing preview does not establish that a physician has a Haiku practice environment. A role-specific inventory avoids both outdated criticism and exaggerated assurances.

A synthetic test could begin with a clearly stated symptom and an intentionally uncertain assessment. Inspect the first draft, correct an ambiguity and add a second recording. Then ask whether the correction survives, how the newest version is presented and whether the reviewer can identify the difference before transfer to the EHR. These are proposed tests, not findings that the product fails them.

A second test could follow a deleted or rejected candidate through the applicable workflow. Examine the saved note and, where relevant, the problem list, patient instructions and other downstream records. First establish which component performs each transfer. If a local EHR configuration or another application changes the information, evidence should identify that step rather than automatically attributing it to the scribe vendor.

Keep a visible editing aid distinct from a complete account of provenance. Highlighting may help a reviewer notice a change. A fuller investigation may also need the source of a statement, the identity and timing of its reviewer, the version transferred and its later destination. The evaluation should record what the interface shows, what logs can establish and what remains unavailable.

The strongest response to criticism is that the product creates a draft for a clinician to review. That may accurately describe the intended function. The appropriate test is whether the actual deployment preserves that division of work. Examine the available review steps, the clarity of the interface and the consequences of accepting, correcting or rejecting content. Document successful controls alongside any reproducible difficulty.

Review should also consider ordinary work conditions. A feature that is available can still require training or an adjusted sequence of actions. Conversely, a confusing first attempt may improve after appropriate instruction. A fair evaluation records the training provided, repeats the defined task and distinguishes a reproducible behavior from an isolated user misunderstanding.

The next public addition to the DAX–Dragon–Epic investigation should therefore be a dated release-and-workflow record. Preserve earlier observations with their original versions and dates. Link UNFAITHFUL’s question of whether the record represents the encounter to a repeatable review of the current workflow. No local tenant or clinical record was inspected for this article. Its contribution is a testable method for examining the documented changes before drawing conclusions about a particular deployment.