What a clinical hallucination is
A large language model does not know what happened in the room; it predicts fluent, plausible text. In clinical documentation that produces a specific failure mode: the note reads perfectly but contains a detail the visit never contained: a symptom that never came up, a physical exam that was never performed, a medication or dose that was never discussed. Because the writing is confident and well-formed, an unsupported line is easy to sign without noticing. That is what makes it dangerous, and it is why a note needs to be checked against the record of the actual visit, not just read for fluency.
This is documented, not hypothetical
In May 2026 the Auditor General of Ontario tested the AI scribe systems the province had approved and found errors across the board:
Independent reporting has documented the same pattern in individual cases. AI note-takers describing exams that never happened or conditions a patient never raised. See Futurism and Medical Economics.
How Alcove's verification pass works
- Alcove drafts the SOAP note from the transcript of the visit.
- A second, separate AI model re-reads that draft against the transcript, line by line, checking whether the visit actually supports each claim.
- Anything the transcript does not support is struck in the note, so it is visible rather than buried.
- You review every line and sign. You confirm or remove each struck claim; nothing is filed without you.
The point is not to hide the AI's mistakes but to surface them, so your review starts from a note that has already been checked against the visit rather than a blank draft.
Single-pass vs. verify-then-strike
A single-pass scribe
Generates the note and hands it to you. Any invented detail is now your job to catch by re-reading, with no cue as to which lines to distrust.
Alcove: verify then strike
A second AI checks the draft against the transcript and strikes unsupported claims first. Your review starts from a checked note, with the doubtful lines already flagged.
The honest boundary
Alcove always produces a clinician-reviewed draft, never an autonomous note. Verification makes review faster and safer; it does not remove the clinician. We do not publish an Alcove accuracy percentage. The honest promise is a note that has been checked against the visit, which you then review and sign. The clinician, not the vendor, is responsible for what the signed note says.
New to the category? See how Alcove works inside Practice Fusion, or read the FAQ.
Common questions
Which AI scribe checks its own notes for hallucinations?
Alcove adds a second, separate AI that re-reads every drafted note against the transcript of what was actually said and strikes claims the visit does not support, before you review and sign. This transcript-grounded re-read is the core of how Alcove counters the well-documented tendency of AI scribes to add clinical detail that never came up.
Do AI scribes really invent clinical detail?
Yes, and it is documented. In May 2026 the Auditor General of Ontario reported that all 20 of the province's approved AI scribe systems produced one or more inaccuracies in testing, that 60 percent recorded a different drug than was prescribed, and that nine of the 20 fabricated information such as treatment suggestions that were never mentioned in the recording. This is the failure mode Alcove's verification pass exists to catch.
Does verification mean I can skip reviewing the note?
No. Alcove always produces a clinician-reviewed draft, never an autonomous note. The verification pass makes your review faster and safer by flagging and striking unsupported claims, but you review every line and sign. Alcove does not publish an accuracy percentage; the honest promise is a checked draft, not a guarantee.