ACL 2026 citation audit
The paper exists.
Does it say that?
We checked every reference in 4,459 ACL 2026 papers, then sampled claim citations and read the cited work. The second measurement changed when we repeated it. That result belongs in the open, too.
Does the source exist?
Match the reference to real scholarly work.
Does it support the claim?
Read the paper and compare the cited text.
A real reference can still be attached to the wrong claim.
What the audit found
One answer held.
The other moved.
2 / 209,985
Confirmed nonexistent references
After matching, triage and manual review, just two references pointed to papers that do not exist. This full-corpus result is unaffected by the support-check replication failure.
One run was not enough.
The confirmed support-defect rate changed across independent samples of 100 papers. The first run's headline number does not reproduce, so we publish all three.
Citation-level rates after second-stage review. The two redraws pool to 5.90%; the original audited run was 0.95%.
The method
A finding has to survive its own challenge.
Automated checks generate leads, not verdicts. We make their limits visible and review the evidence before calling anything a finding.
Read the method in fullFind the work
Resolve each reference against scholarly indexes, then inspect unresolved entries.
Read the evidence
For claim citations, retrieve the cited paper and locate passages that bear on the claim.
Try to refute the flag
A second-stage reviewer checks whether a benign reading or retrieval error explains it.
Single-paper audit
Bring your own paper into focus.
Enter an arXiv ID to examine its references and claim citations. Results are leads for human review.