Introduction | A first draft generated by GPT most easily hides a landmine not in the writing, but in the references—fabricated titles and fabricated arXiv numbers. This article explains how to verify dozens of citations thoroughly in one pass.
1. How Costly Are Hallucinated Citations?
In a GPT first draft, it is all too common for a few nonexistent papers to slip into the references: neat formatting, plausible-looking numbers, and even authors and years that seem reasonable.
What is the cost? One sentence from a reviewer—“this reference does not exist”—and the credibility of the entire manuscript collapses. By then you have already cited it in a dozen places, and checking dozens of citations one by one is an enormous amount of work.
So there are only two choices: verify everything thoroughly before submission, or have a reviewer point it out to your face.
2. Why “Correct Citation Format” Does Not Mean “Real Citation”
Many people think that as long as the format is set to IEEE / ACM, there is no problem. In fact, format is only appearance—a perfectly formatted citation can still point to a nonexistent paper.
There is only one way to determine whether it is real: take the DOI / arXiv number and “resolve” it to see whether it lands on a real publication. If it cannot be resolved, it is a hallucination.
3. Four Steps to Fully Verify Citations
Step 1: Batch-Resolve DOI / arXiv Links
Submit the citations in the first draft in batches to the citation verification module of MuchSci Paper Sailor, which resolves each DOI and arXiv number one by one. This step is automatic; you do not need to open a browser for each one.
Step 2: Remove Fake Citations
Entries that fail resolution are automatically removed. After this step, you are left with only references that actually exist.
Step 3: Check Whether In-Text cite Markers Match the References
Having entries alone is not enough—citations in the text that are missing from the list, and entries in the list that are never cited in the text, are both serious flaws. Paper Sailor performs a three-way cross-check:
| Category | Meaning | How to Handle |
| Used | Cited in text and present in list | Keep |
| Missing | Cited in text but not in list | Add an entry, or delete this citation sentence |
| Unused | Present in list but not cited in text | Add a citation, or delete the entry |
All three categories must be reduced to zero or handled cleanly before citations are truly closed-loop.
Step 4: Export BibTeX / RIS and Import into Zotero
Verified citations can be exported with one click as BibTeX or RIS files, imported directly into Zotero for management, or used in the submission system.
4. A Complete “Manuscript Rescue” Walkthrough
Suppose you have received a reviewer comment: “Reference [17] does not exist”:
1. Batch-resolve every DOI / arXiv in the entire manuscript.
2. Review the resolution results: which ones do not resolve to real publications.
3. Remove all fake entries.
4. Three-way cross-check: handle Used / Missing / Unused category by category.
5. Replace in-text citations: for deleted entries, replace them with real references or delete the sentence.
6. Export BibTeX / RIS and import into Zotero.
7. Run a final check to confirm zero hallucinations.
5. Common Pitfalls
• Fixing only the one that was called out. The reviewer only mentioned [17], but that does not mean the other entries are real—the entire manuscript must be checked.
• Checking dozens of citations manually. Your eyes will glaze over, and there is no way to guarantee resolution accuracy. Batch resolution is made for exactly this scenario.
• Looking only at the list and not checking the main text. Mismatches between in-text cite markers and the list will also be caught.
• Not exporting after verification. If you verify and then walk away, you will have to recheck when revising later—export to BibTeX / RIS and manage everything in Zotero.
6. Conclusion
References are the part of a paper most prone to being taken for granted, and also the part reviewers are most likely to spot-check. The citation verification module of MuchSci Paper Sailor turns “batch resolution → removal of fake citations → three-way cross-check → export” into a pipeline; all you need to do is run it once.
Pre-submission self-check:
• Have all DOIs / arXiv numbers in the manuscript been resolved?
• Have all fake entries been removed?
• Have Used / Missing / Unused all been handled?
• Have you exported BibTeX / RIS and imported it into Zotero?