Patent Pending — Published Article Reference Verification
NO AI  •  Fully Deterministic  •  No False Positives

One paper.
5 references pointing to the wrong science.

SciPubAudit verifies every reference in a published scientific article against more than 800 million canonical source records. Paste a PubMed Central ID. Every reference is checked field by field — title, authors, journal, volume, year, pages, DOI. The most dangerous error — Type AI, where a DOI resolves to a completely different paper — is caught automatically. No AI. Every result is traceable to its source.

Run a Free Audit Manuscript Mode → Journal Integration

800M+

Canonical Source Records

0

AI in the Verification Pipeline

7

Citation Error Types Detected

SciPubAudit uses no artificial intelligence or machine learning. Every verdict is produced by deterministic comparison of canonical bibliographic records against independently reproducible rules. The same input always produces the same output.

No AI or ML
Fully Deterministic
No False Positives
Every Result Traceable
The citation integrity crisis

Science has a reference problem. It’s accelerating.

Fabricated and distorted citations now appear in journals across every discipline. They are correctly formatted, attributed to real researchers, and invisible to conventional peer review.

12× Increase

Rise in papers containing at least one fabricated reference — from 1 in 2,828 (2023) to 1 in 277 (2026).

Ponce-de-León et al., The Lancet (May 2026). 2.5 million biomedical papers screened.

6.5× Higher Rate

Unverifiable references in AI-assisted papers relative to human-authored controls, across 56 journals in JAMA Network.

Topaz et al., JAMA Network Open (2026). Pre-/post-AI comparison across 56 journals.

1.23%

Unverified reference rate in post-2024 PNAS articles — 31% higher than the pre-AI baseline of 0.94%.

Evidite PANDA Study (2026). 2,345 PNAS articles, 269,603 references. RR 1.31, p = 2.2×10−13.

110,000+

Publications with invalid AI-generated references identified in the scientific literature in 2025 alone.

Naddaf & Quill, “Hallucinated Citations Are Polluting the Scientific Literature,” Nature 652 (2026): 26–29.

“Hallucinated citations are correctly formatted and attributed to real researchers—making them indistinguishable from legitimate references to the human eye.”

Naddaf & Quill, Nature 652 (2026)
Type AI — the identity-substitution error

The DOI resolves. To the wrong paper.

A Type AI error is the most dangerous citation failure: the identifier is technically valid, but the DOI resolves to a completely different scientific work. The asserted title, authors, and findings are attributed to a paper that says something else entirely. No existence-based tool catches it.

Real examples from a 2026 audit of PMC13415537 — a Nature Communications paper on ion-sieving membranes containing 5 Type AI errors (7.1% of 70 references). Each DOI resolves to a real paper on an unrelated topic.

Type AI • Ref #3
Asserted citation
Research on ion separation membranes — relevant to the paper’s claimed mechanism.
10.1038/s41467-023-44064-7
DOI resolves to
Liu et al., “Bionic artificial skin for continuous pressure and temperature monitoring” — a paper on wearable sensors and bionic skin. Nothing to do with ion-sieving membranes.
Evidite finding: DOI resolves to a real, indexed paper with zero content overlap. The citation would pass any existence check. Only field-level verification catches it.
Type AI • Ref #52
Asserted citation
Research on graphene oxide membrane selectivity supporting a claim about charge-based ion exclusion.
10.1038/s41467-018-07894-4
DOI resolves to
Heeter et al., “Spatially heterogeneous trends in Pacific Northwest streamflow...” — a geophysical study of Pacific Northwest precipitation and climate events.
Evidite finding: Title match score: 0.00. Author match: 0.00. The DOI is from a different journal, a different year, and a completely different scientific domain.
Type AI • Ref #57
Asserted citation
Research on ion sieving at the sub-angstrom scale, cited to support claims about membrane selectivity.
10.1038/nature24035
DOI resolves to
Hirschi et al., “Structural basis for competency of the mammalian TRPML1 channel” — structural biology of a lysosomal ion channel in mammals.
Evidite finding: DOI is a valid Nature article DOI. The paper exists. It has nothing to do with the asserted claim or the citing paper’s subject matter.

Why conventional tools miss this: Any tool that checks only whether a DOI resolves to a real paper will report all three citations as “verified.” SciPubAudit verifies the identity of the resolved paper against every field of the asserted citation — and flags the mismatch.

800M+

Canonical source records in the Evidite Stacks database

7

Citation error types identified and classified

0

AI or machine learning in the verification pipeline

1–3 min

Typical audit time for a full published article

How it works

Deterministic. No AI. No machine learning.

References are fetched from structured JATS XML — not scraped, not parsed from PDF. Every field is compared against canonical records in parallel. The verdict is reproducible and traceable.

1

Paste a PMC ID

Enter any PubMed Central identifier. SciPubAudit fetches the structured JATS XML record directly from NCBI — the authoritative source.

2

Extract every reference

All references in the article are parsed from the XML. No OCR. No PDF heuristics. Every field is captured exactly as it appears in the published record.

3

Verify against 800M+ records

Each reference is matched against the Evidite Stacks database — 800 million canonical records from Crossref, PubMed, OpenAlex, and Semantic Scholar. Field by field, in parallel.

4

Receive a complete audit

Every reference receives a verdict: Verified, Partial Match, Unverified, or Type AI. Full field-level diff. CSV export with complete provenance.

The interface

Paste a PMC ID. Get a complete audit.

Every reference is displayed with its verdict, field-by-field comparison, and source provenance. Export the full audit as CSV.

SciPubAudit audit results interface showing reference verdicts with field-level comparison
Try SciPubAudit Free
Who it’s for

Editors. Authors. Anyone accountable for the record.

SciPubAudit serves two different constituencies with the same tool and the same standard of evidence.

Journal Editors

Run any published article through SciPubAudit before or after peer review. Identify reference errors that reviewers cannot catch. Build a systematic integrity standard into your editorial workflow.

  • Paste a PMC ID — audit runs automatically
  • Catch Type AI errors invisible to DOI existence checks
  • Field-level discrepancy report for every flagged reference
  • CSV export for editorial records and correspondence
  • Enterprise integration for journal management systems
Discuss Journal Integration

Article Authors

Audit your own published paper before a correction becomes necessary. AI writing tools introduce errors that authors do not intend and cannot easily detect. SciPubAudit is the fastest way to know what you actually cited.

  • Audit your article the day it goes live in PMC
  • Verify that all DOIs resolve to the papers you intended to cite
  • Identify field errors introduced by citation managers or AI tools
  • Document your reference accuracy as a professional standard
  • Free first audit — no subscription required to start
Audit Your Article Free
Run your first audit. Free.

Every published article in PMC
can be audited in under three minutes.

Paste a PMC ID. No credit card. No setup. The first audit is free. If the results matter, you’ll know immediately.

Run a Free Audit Enterprise Inquiry