LLM Citation Generation — Plausible References Attached to Clinical Claims
Fluent citations that look real, formatted correctly, generated alongside claims.
- 8–24: Citations per response (typical clinical-answer output)
- >95%: Format plausibility (looks like a real reference)
- <10%: Reviewer spot-check rate (without a dedicated tool)
- Title, journal, year, DOI: Fields per citation (all independently checkable)
Why citations get generated as fluent text
Citations are text tokens, not database lookups — fluency isn't truth.
Existence Verification — Does This Source Even Exist?
First gate: search title, journal, and DOI against real bibliographic records.
- DOI + title lookup: Existence check method (CrossRef, PubMed, journal index)
- 10–20%: Fabrication rate (typical) (depends on model and topic)
- Well-formed, unresolvable: False DOI patterns (looks valid, points nowhere)
- Seconds per citation: Check cost (automatable at scale)
A fabricated citation is a distinct failure mode
No paper, no journal issue, no DOI record — the source is invented.
Relevance Verification — Does the Real Source Support the Claim?
Second gate: read the real source and compare it to the attributed claim.
- 5–15%: Misattribution rate (typical) (of existing citations)
- Right topic, wrong finding: Common pattern (plausible but unsupported)
- Claim-to-abstract comparison: Check method (human or LLM-assisted read)
- Higher than existence check: Detection difficulty (requires reading the source)
Existing sources can still be the wrong evidence
Real paper, real journal — but it never claims what's attributed to it.
Classification — Fabricated, Misattributed, or Verified
Every citation lands in exactly one of three buckets after both checks.
- 3: Outcome classes (fabricated, misattributed, verified)
- Red: Fabricated stamp (source does not exist)
- Orange: Misattributed stamp (real source, wrong claim)
- Green: Verified stamp (real and accurate)
Two independent gates, one final label
Existence and relevance both pass before a citation earns "verified."
Outcome — A Distinct Failure Mode Beyond Fact-Checking
A fabricated citation can slip past reviewers who trust the reference list.
- Yes: Distinct from fact-checking (claim can be true, source fake)
- High: Reviewer trust risk (formatting looks authoritative)
- Independent source verification: Mitigation (every citation, every time)
- Caught failures ≠ 0: Net effect (checks find what reading misses)
Verification is a separate step from accuracy review
Checking the claim is true is not the same as checking the source is real.
A citation can be fabricated even when the underlying medical claim happens to be true — which is exactly why source verification has to run as its own independent check.