The most dangerous failure in an AI chart summary is not a fabricated fact. It is a real fact that was never mentioned.
When a summary invents something, the clinician can catch it by checking the chart. When a summary silently omits a contraindication or an abnormal lab, the summary reads as complete, nothing looks wrong, and the omission travels forward unchallenged. That asymmetry is the whole problem, and most of the attention has gone to the wrong half of it.
A hallucination announces itself; an omission hides
Hallucination is visible by definition. If a summary states a drug the patient has never taken, that claim can be traced. It either appears in the chart or it does not. A clinician who suspects something reads back. A grounding check can flag it automatically.
Omission leaves no trace to follow. The summary is internally consistent. The sentences that exist are accurate. There is simply no sentence about the penicillin allergy, or the creatinine trending upward, or the second medication added two days ago. The reader has no signal that anything is missing.
This matters clinically because the two failures push in opposite directions. A hallucination creates doubt. An omission creates false confidence. One slows the clinician down; the other speeds them past the gap.
A 2025 paper in npj Digital Medicine, "A framework to assess clinical safety and hallucination rates of LLMs for medical text summarisation" (Asgari et al.), treats omission as a distinct, safety-relevant error class alongside hallucination in clinical LLM summarization. The two failure modes are not equivalent, and omission deserves its own scrutiny. In practice it is the less-tested failure, because it is harder to measure.
Why the attention skews toward hallucination
The term "hallucination" entered common language first. It is a vivid concept, easy to explain, and it lines up with how people already distrust automated systems. Concern about AI making things up is intuitive.
Omission runs against intuition. People do not naturally distrust a summary for what it does not say. The cognitive default is to read a summary as a representative sample of the source. If it is fluent and the facts it does state are accurate, it reads as trustworthy.
That default is the mechanism of the failure. The summary does not look incomplete. It looks finished.
A verifier checks what is written, not what is missing
This is a structural problem, not a technical one. A grounding verifier works by checking whether the claims in a summary are supported by the source document. For each sentence in the summary, it asks whether the chart backs it up. If the answer is no, the sentence is flagged.
But a verifier can only evaluate sentences that exist. If the summary never mentions a finding, there is no sentence to check. The verifier sees nothing wrong because there is nothing to evaluate. You cannot check a sentence that was never written.
So omission has a verification blind spot that hallucination does not. A pipeline can be built to catch fabrications with reasonable reliability. Omissions, by construction, escape the same pipeline. The check confirms accuracy. It cannot confirm completeness.
Completeness needs a different question. Not "is what the summary says true?" but "does the summary cover what matters for this patient?" The second question is much harder to automate, and it is the one that decides whether a summary is safe to act on.
The omissions that carry clinical weight
Not all missing information is equal. A missing diet note is not a missing allergy. Clinical risk is unevenly distributed across the chart, and it concentrates in a few sections.
The high-stakes gaps tend to sit in the medication list, especially recent additions and dose changes; in allergy and adverse-reaction records; in the active problems that bear on the current visit; in the assessment and plan, where the clinical reasoning lives; and in lab values that are abnormal or trending. These sections share a property. A clinician who misses something in them may act on the gap, and the harm downstream is concrete rather than theoretical.
A summary that covers demographics and history accurately but drops a contraindication from the medication list is not a mostly-correct summary. It is a summary with a clinical risk embedded in what it left out.
Read it as a draft, confirm against the chart
The safest model treats an AI summary as a starting draft, not a final account. The summary narrows what needs review. It does not replace the review.
In practice, the clinician confirms the high-risk sections against the actual chart, not against the summary. If the summary mentions a medication, the chart is the source of truth for that medication's current dose, start date, and interaction profile. If the summary says no known allergies, that claim still gets a direct check.
This is less efficient than fully delegating to the summary, and the efficiency loss is the point. The cost of one more check on a high-risk section is low. The cost of acting on an undetected omission is not.
Source traceability lowers the friction. When each summary line links back to the section of the chart it came from, the clinician moves from summary to source in one step. The summary becomes a way to read the chart faster, not a replacement for reading it.
What Meddies is designed to do
Each line in a Meddies summary carries a source reference, so the clinician can trace it back to the chart directly. The summary is treated as a draft to confirm, not a document to trust without review. The design intent is to pull attention toward the high-risk sections, medications, allergies, and the assessment and plan, because those are where an undetected omission is most likely to matter.
This does not claim to remove omissions. No summarization approach can claim that honestly. What it does is lower the cost of catching one when it exists.
The underlying principle is simple. Completeness is a property of the chart. A summary is a pointer to the chart. Keep the pointer visible, use the chart to confirm what matters, and the summary stops being a place where a fact can quietly disappear.