Tags: #academic-integrity, #AI-hallucination, #AI-writing, #citation-verification, #desk rejection, #hallucinated citations, #journal submission, #pre-submission checklist

Save 3 weeks and protect your research priority. ACL 2026 desk-rejected 100+ accepted papers in 2026 for hallucinated references. The Topaz et al. Lancet audit found 1 in 277 papers cited a fabricated reference — a 12-fold increase from 2023. Every one of these errors is detectable before submission. Run the checklist below in 45–75 minutes. It takes less time than a desk rejection puts you back.

The Gatekeepers Are Watching

In early 2026, the ACL program chairs made an unprecedented move: they desk-rejected over 100 papers that had already been “accepted” after an automated reference screening system flagged hallucinated citations. The flagged papers were caught during final camera-ready checks, and all rejections required a full resubmission. Every paper that was rejected — regardless of scientific merit — never reached peer review.

Three days later, ICLR 2026 implemented automated reference extraction that checked every citation against multiple bibliographic databases. NeurIPS 2026 began desk-rejecting papers with two or more hallucinated references. ICML 2026 added “hallucinated references” as explicit grounds for desk rejection in its Peer Review Ethics page.

And in The Lancet that same month, a landmark audit by Topaz et al. found that 1 in 277 biomedical papers indexed in PubMed Central (early 2026) cited at least one fabricated reference — a 12-fold increase from 2023.

Here is the reality most researchers face now: journals and conferences in 2026 actively screen reference lists before manuscripts ever reach peer review, and failures can trigger desk rejections without any scientific evaluation. The gatekeepers are not just looking at content — they are auditing your citations, your formatting, and increasingly, whether your references are real.

This article gives you the complete pre-submission reference audit checklist that prevents those desk rejections. It covers the four major error categories journals catch, the self-audit tools that verify references in under an hour, and a structured workflow you can implement before your next submission.

Every error on this list is detectable before submission. That is why this checklist exists.


What Journals and Conferences Actually Check

Major publishers now integrate automated “technical check” systems at the moment of submission. These systems sit between your uploaded manuscript and the editorial assignment, running automated validation before any editor ever reads your paper.

Wiley, Springer Nature, and JAMA Network all use these systems to validate DOIs, detect retracted citations, flag hallucinated references, and verify internal consistency between in-text citations and your bibliography. If a system finds an error, the manuscript never reaches a human editor.

IEEE venues require standard journal abbreviations and DOI formats. Non-compliant references are flagged during the technical check.

The Lancet enforces a strict 50-reference limit. A single fabricated reference in a Lancet submission can trigger a referral to the submitting author’s institution’s research integrity office.

arXiv announced in May 2026 that authors who submit papers with unambiguous AI-generated hallucinated citations face a one-year submission ban — after which, all future submissions must first clear peer review at a reputable venue.

The question is no longer whether journals check references. The question is whether your reference list will survive that check.

Reference Limits at Major Venues

Journal / Venue Reference Limit
Nature ~30–50
Science ~40
The Lancet 50
NEJM 70 (articles), 5 (letters)
Cell 100
PLOS ONE No stated limit

The editorial office counts references before anything else. This check takes under a minute. If you exceed the limit, your paper is returned without review.


The Pre-Submission Reference Audit Checklist

This checklist covers four error categories that journals catch during their technical screen. Work through each section in order. The entire process takes about 45–75 minutes for a 50–70 reference paper — roughly one hour.

Section A: Formatting & Style Compliance

Confirm exact citation style

The single most common cause of desk rejection is wrong citation style. A psychology paper formatted in APA resubmitted to a clinical medicine journal using Vancouver style references is “almost certainly desk rejected on technical grounds” by an editorial assistant within 30 seconds. Check your target journal’s most recent published article and copy its reference format exactly.

Count references against the journal’s stated limit

Use the table above as your reference. Nature allows 30–50. The Lancet allows 50. If your paper has 52 references for The Lancet, it will be rejected. The editorial office counts this before anything else.

Verify DOI format matches journal requirements

APA 7th edition requires DOIs as full URLs (https://doi.org/…). AMA style requires a doi: prefix without the URL. Submitting the wrong format flags immediate editorial review. Check your journal’s author guidelines for the required format.

Check journal abbreviations

Vancouver style requires NLM standard abbreviations. IEEE requires IEEE abbreviations. Reference managers frequently import full journal names instead of standard abbreviations, and this mismatch is flagged during automated screening.

Section B: Citation Integrity

Resolve every DOI and confirm the correct paper

Go to doi.org and resolve every DOI in your reference list. Compare the resolved paper’s title and authors against your citation. If they don’t match, your citation is wrong — even if the DOI “worked.”

For a deeper dive on this verification step, see our guide on how to check if a reference is real.

Screen all references against retraction databases

Check every citation against the Retraction Watch Database and PubMed. Use the PubMed retraction filter (Retracted publication[pt]). For a detailed workflow, see our guide on checking if a paper has been retracted.

Papers that cite multiple retracted works may be returned before review. The CrossMark programme (which includes most major publishers) queries retraction status automatically during processing.

Verify numbering sequence

For Vancouver, AMA, and IEEE styles, any gap, duplication, or out-of-sequence reference numbering is “an immediate red flag.” This error almost always occurs during revision when new references are added and renumbering is forgotten.

Cross-reference in-text citations against bibliography

Every in-text citation must match its corresponding bibliography entry. Name spelling differs, year is wrong, or reference number points to the wrong entry — “none of these are hard to create, and none are hard to catch with a systematic check.”

Section C: AI & Fabrication Detection

Run the four-database verification workflow

This workflow catches virtually all hallucinated citations and takes 20 minutes to 2 hours depending on your reference list size:

  1. Extract your reference list to a flat format (BibTeX or RIS from your reference manager; plain text if you don’t use one)
  2. Flag any reference whose title you do not recognize — this is where hallucinations are highest risk
  3. Resolve every DOI at doi.org and compare the destination paper’s title and authors against your citation
  4. Verify every non-DOI reference in PubMed or Crossref (title search + first-author surname)
  5. Cross-check escalated references in OpenAlex and Google Scholar
  6. For preprints, verify both the preprint AND any published version
  7. Spot-check in-text claims against sources (verify 10% of references, or 100% for systematic reviews)
  8. Document the audit — save your verification log as evidence

Why four databases? Each catches different failure modes. PubMed catches biomedical-specific patterns. Crossref catches DOI mismatches. OpenAlex catches humanities and conference papers. Google Scholar catches under-indexed sources. A title-only Google Scholar lookup misses fabrication patterns 2 and 3; a DOI-only Crossref lookup misses pattern 1.

If you’re unsure whether a reference is real or AI-generated, see our guide on spotting Frankenstein citations.

Check for preprint-to-publication mismatches

If you cited an arXiv preprint but the same paper has since been published in a peer-reviewed journal for two years, that suggests a stale literature review. Clinical journals explicitly prohibit citing preprints as primary evidence for key claims.

Spot-check in-text claims

Open a sample of 5–10 references and verify that the cited source actually supports the claim you made. This catches the “misattributed quote” problem: the paper exists, the DOI is correct, but the paper doesn’t actually say what you claim it says.

Section D: Final Pre-Submission Checks

Run an automated reference checker pass

Use one of the tools listed below (Citely, Paperpal, or ProofreaderPro.ai) for a final automated scan. These catch the errors that manual review misses.

Document the audit

Save your verification log and any tool-generated reports. This serves as evidence that you conducted a good-faith audit, which is useful if a question arises post-submission.

Check self-citation proportion

Some journals state in author guidelines that self-citation should not exceed a specified percentage. A 60-reference paper with 20 self-citations will attract editorial scrutiny. Review your target journal’s author guidelines.


Tools for Automated Verification

You don’t need to verify every reference manually. Several tools automate this work:

Scholar Sidekick (scholar-sidekick.com/tools/bibliography-audit)

  • Free, no signup required (up to 25 references per run)
  • Parses BibTeX, RIS, CSL-JSON, or Word .docx manuscripts
  • Verifies citation identity AND checks retraction status
  • Generates an evidence report (HTML + JSON receipt) for documentation
  • Measured 87.3% fabrication recall on blind holdout

Citely (citely.ai/reference-checker)

  • Batch compares references against CrossRef records
  • Catches wrong years, author name substitutions, journal title mismatches
  • 1–2 minutes for a 60-reference list (vs. 2–3 hours manually)

Paperpal Reference Checker (edit.paperpal.com)

  • Validates citation mismatches, incorrect DOIs, retracted studies, outdated references
  • Integrated into broader document health check (grammar + plagiarism + AI review + references)

ProofreaderPro.ai (proofreaderpro.ai/ai-proofreader)

  • Runs the full four-database audit in one pass
  • Highlights hallucinated and mismatched references
  • Exports a downloadable verification log

Crossref Public API — Use the free Crossref API at api.crossref.org/works/{DOI} to check references.

  • Free, no signup needed
  • Returns JSON with canonical title, authors, year
  • Can be automated with a simple script to check an entire reference list
  • Catches fabrication and wrong-paper-match patterns in seconds per reference

How Long It Actually Takes — A Researcher’s Workflow

Dr. Emily Carter, a biomedical informatics researcher with 10+ years of publishing experience and 50–80 references per paper, reports finding 2–4 issues per manuscript using the workflow below:

Step Time Description
Export reference list as plain text 5 min Raw text, no formatting or embedded fields
Automated DOI + metadata verification 1–2 min Citely, Paperpal, or similar tool
Manual review of flagged references 10–20 min Open DOI link, compare, fix
Retraction check 5–10 min Retraction Watch + CrossRef metadata
Formatting verification 15–20 min Match 3–4 references against journal’s published papers
In-text citation cross-check 10–15 min Every in-text citation matches bibliography entry
Total 45–75 min About an hour for 50–70 references

She notes: “Before I started using automated tools for Step 2, the manual DOI verification alone took 2–3 hours. The automated step reduced total workflow from half a day to about an hour.”

For multi-author papers, she recommends assigning one researcher per section to run the verification pass independently — this reduces the per-person time investment while increasing coverage.


Consequences of Skipping the Checklist

The consequences of skipping this checklist are not theoretical:

Desk rejection is the most immediate risk. Desk rejection rates at high-impact journals routinely exceed 50–70%. Citation and formatting errors account for a significant proportion of these technical desk rejections. The research may be perfectly suited to the journal, but the paper is returned without review.

One researcher (Dr. Carter) reported that a Q1 journal desk rejection “cost me a revision cycle and, because another group published similar findings in the interim, probably cost me the priority claim.” Weeks are lost. Months can be delayed.

Post-publication retraction carries heavier consequences. An Elsevier Q1 paper was retracted over 38 fake citations. Mass Springer Nature book retractions occurred when publishers could not verify cited sources. Retracted papers receive permanent “retracted” status and watermark across Scopus and Web of Science — the retraction is permanent.

Institutional consequences are severe. The Lancet and BMJ report single fabricated references to the author’s institution research integrity office. arXiv issues one-year submission bans for unambiguous hallucinated citations.

Resnik and Hosseini (2026) argue in Accountability in Research that hallucinated citations may constitute research misconduct when citations function as data in scholarly papers — meaning “using GenAI output without checking it for veracity” is reckless, not careless.


2026 Policy Changes You Need to Know

The Topaz et al. Lancet Audit

Published in The Lancet in May 2026, this audit of 2.5 million biomedical papers indexed in PubMed Central found that 1 in 277 papers (first 7 weeks of 2026) cited at least one fabricated reference — a 12-fold increase from the 2023 rate of 1 in 2,828. The greatest uptick began mid-2024, coinciding with AI tool adoption. Over 98% of papers with small numbers of fake references (1–2 per article) historically received no editorial action — this is changing.

Conference Policy Shifts

  • ACL 2026: Automated system + human review → 100+ desk-rejected accepted papers (2026.aclweb.org/acl_statement/)
  • ICLR 2026: Automated database cross-reference + three human review rounds (StrictCite)
  • NeurIPS 2026: Informal threshold of 2+ hallucinated references for desk rejection (StrictCite)
  • ICML 2026: Hallucinated references listed in Peer Review Ethics as desk-rejection ground (StrictCite)
  • arXiv: One-year ban for unambiguous hallucinated citations (May 2026)

JCR 2026 Changes

Clarivate’s Journal Citation Reports 2026 introduced changes to uphold research integrity, excluding retracted citations from impact factor calculations (Clarivate blog). This affects how journals are evaluated and incentivizes more rigorous reference verification.


Bottom Line

Here is what I think most researchers should do: build reference verification into your pre-submission workflow as a non-negotiable step. Not a “maybe,” not an “after review,” not an “if there is time.” A dedicated step, same weight as proofreading or formatting.

The trade-off is clear:

  • Do the checklist: 1 hour of your time
  • Get desk-rejected: weeks or months of lost research priority

The common mistake is assuming that because the science is sound, the references will be fine. That assumption costs more papers every month than any editor wants to admit. The ACL 2026 event alone — 100+ accepted papers rejected in a single automated sweep — proves that reference verification is no longer optional.

Every error on this list is detectable before submission. The tools exist. The policies are active. The data is clear.

Here’s what I’d do: run the four-database workflow (Step 8 above) on your next manuscript before you hit submit. It takes about an hour. The alternative is losing weeks to a desk rejection that no amount of good science can undo.

This article was written by an AI-assisted writer and reviewed by our editorial team. For the full methodology, see our about page.