AI scribe hallucinations are not theoretical. They happen in real clinical notes, in real practices, every day.
A hallucination in an AI-generated clinical note is not always a dramatic fabrication. More often it is subtle — a diagnosis that contradicts a documented finding, a medication listed for a condition that was ruled out, a value that is physiologically impossible but formatted to look normal.
The physician who signs the note is legally attesting that its contents are accurate. That makes catching these errors before signing a professional and legal obligation. This article covers the three types of AI scribe hallucinations found in clinical notes, with real examples of each.
What Is an AI Hallucination in a Clinical Note?
In clinical documentation, an AI hallucination is any instance where the AI generates content that:
- Contradicts other documented facts in the same note
- Is physiologically or clinically impossible
- Contains temporal inconsistencies suggesting data was not generated from the current encounter
Unlike general AI hallucinations — where a model simply invents facts — clinical note hallucinations are dangerous because they are embedded in a medical record that drives treatment decisions, billing, and legal proceedings.
Ontario's Auditor General reviewed all 20 government-approved AI scribe vendors and found every one produced hallucinations, incorrect information, or missing content during procurement testing. This is not a problem limited to one platform or one specialty.
Type 1: Contradiction Hallucinations
A contradiction hallucination occurs when two statements in the same note cannot both be true. These are the most dangerous type because they look like a documentation error — easy to miss on a quick review.
Example 1 — Hemodynamic status:
A blood pressure of 58/40 with a heart rate of 128 is not hemodynamically stable by any clinical definition. The AI generated a standard phrase that contradicts the objective data it also documented.
Example 2 — Heart failure type:
An ejection fraction of 58% is normal — this is HFpEF, not HFrEF. This type of HFrEF vs HFpEF confusion is one of the most consequential AI scribe errors in cardiology — the treatment implications are completely different. Entresto is indicated for HFrEF. Prescribing it based on a hallucinated diagnosis creates medication risk.
Example 3 — Symptom contradiction:
The AI documented denial of symptoms in the history and then listed those same symptoms as the reason for the visit in the assessment.
Example 4 — Age contradiction:
The same patient is documented as two different ages in the same note.
Type 2: Implausibility Hallucinations
An implausibility hallucination is a value or finding that is physiologically impossible. The number exists in the note but cannot exist in a human patient. These are easier to catch than contradiction hallucinations because the value itself is the red flag — if you know what normal ranges look like.
Example 1 — Oxygen saturation:
Oxygen saturation cannot exceed 100%. This value was generated by the AI and has no clinical basis.
Example 2 — Kidney function:
Normal eGFR is 60–120. Values above 150 are physiologically implausible in adults and represent an AI-generated error.
Example 3 — HbA1c:
HbA1c values above 20% are not clinically meaningful and typically represent a lab error or, in AI-generated notes, a hallucinated value.
Example 4 — Blood pressure:
A blood pressure of 340/190 represents a hypertensive emergency. "No acute distress" cannot coexist with this value.
Type 3: Copied-Forward Hallucinations
A copied-forward hallucination occurs when the AI inserts data that does not reflect the current encounter — presenting findings in a way that suggests they are current when the note's own internal evidence contradicts this.
Unlike cross-visit errors that require access to prior records to detect, these hallucinations leave internal clues within the note itself.
Example 1 — Future-dated findings:
The AI referenced a visit that has not yet occurred. This is a clear internal inconsistency — the note itself reveals the temporal error.
Example 2 — Plan actions described as already completed:
The plan and the outcome of that plan appear in the same current note — an internal contradiction suggesting the AI mixed content from different time points.
Example 3 — Contradictory clinical status within the note:
The note presents both new onset symptoms and long-term stability as simultaneously current — an internal inconsistency detectable without any prior visit data.
Why These Errors Are Hard to Catch Manually
Physicians reviewing AI-generated notes are looking for clinical accuracy — not internal consistency. The natural tendency is to read the note as a narrative and assess whether it sounds right, not to cross-reference every value against every other statement.
This is not a failure of physician diligence. It is a limitation of human pattern recognition when reading familiar-sounding text quickly. AI hallucinations in clinical notes are designed — unintentionally — to sound correct. The AI produces fluent, clinically plausible language. The errors are embedded within that fluency.
How to Check for Hallucinations Before Signing
For contradiction hallucinations:
- Read the vital signs, then read the assessment — do they agree?
- Read the HPI symptom denials, then read the assessment — any conflicts?
- Check the diagnosis against the objective findings — do they match?
For implausibility hallucinations:
- Flag any SpO2 above 100%
- Flag any eGFR above 150
- Flag any HbA1c above 20%
- Flag any blood pressure above 300 systolic
- Flag any age that differs between sections
For copied-forward hallucinations:
- Look for future dates referenced as past or current findings
- Check whether plan actions are described as already completed in the same note
- Look for contradictions between "new onset" language and "longstanding stable" language in the same note
Automated Hallucination Detection
Manual review using the checklist above is effective but adds time to an already pressured workflow. VerifyChart automatically checks every AI-generated note for contradiction hallucinations and implausible values, and flags internal patterns that suggest copied-forward data — such as future-dated references or temporal inconsistencies within the note itself.
Each flag includes the specific text that triggered it and an explanation of why it was flagged — giving physicians the information they need to make a correction or confirm the finding is accurate.
No signup required. No credit card. Results in under 60 seconds.
The Bottom Line
AI scribe hallucinations are not rare. They occur across all major platforms and all specialties. The three types — contradictions, implausible values, and copied-forward patterns — each create different risks: wrong diagnosis, wrong treatment, and missed clinical changes.
The physician who signs the note carries the liability for its contents. Independent verification before signing is the professional standard that protects both patients and physicians.
Protect your license. Prove you verified.