

Replace a Fake AI Reference Without Citing the Wrong Paper
To replace a fake reference, first make sure the work doesn't exist. Its title leads to no paper, but its wording names the topic: search those words with the key terms of your own sentence. Read a candidate's abstract against your sentence, narrow the sentence to what that paper supports, and update every citation. If nothing supports the claim, cut the claim. In PerfectRefs, Find a replacement searches by the fake title and lists up to three candidates; whether one supports your sentence is yours to judge.
The AI-generated list below was checked with PerfectRefs on November 12, 2026; the Crossref records, arXiv page, and library and journal-editor pages were read on November 17, 2026.
First, make sure the reference is fake
The search steps are in how to check if ChatGPT references are real and how to check references before you submit. A failed search isn't proof: a typo or a journal a database doesn't cover can hide a real paper. If the entry gives an issue and pages, look up what that issue prints there.
A paper with the same title and mostly the same authors and journal is the same paper with wrong details (hallucinated citations describes this form). Correct the entry from the record, then still read the abstract against your sentence. Replace only when the work isn't there at all: a made-up entry has no record to correct it from. For a citation that doesn't exist, Liberty University's library lists "Remove it from your reference list" and "Locate a real, verifiable source that supports your claim" (Miller, 2026).
Example: replacing one fake reference from an AI chatbot
The entry and the sentence it supported
In August 2026, we asked a popular AI chatbot, in Chinese, for a short literature review on knowledge graph construction with 10 references. PerfectRefs found 1 of the 10. Here is one of the other 9, rewritten in APA 7 (the chatbot used a numbered style):
Pons, J., Lee, S., & Wang, H. (2025). Multi-source fusion framework for domain knowledge graph construction. IEEE Transactions on Knowledge and Data Engineering, 37(4), 689–702.
The review cited it in this sentence (our translation):
Existing studies mostly use attention-based graph neural networks and multi-feature fusion algorithms for entity alignment, relation fusion and deduplication, and combine Bayesian optimization with ontology constraints to verify knowledge, which effectively improves the reliability of fusing multi-source heterogeneous data [6-7].
PerfectRefs didn't find [6] either; it needs the same checks before you drop or replace it. We follow [7].
Confirming it doesn't exist
- In Crossref's records for volume 37 of the journal, issue 4 starts at page 1508. Pages 689-702 fall in issue 2, across the end of an article on clustering and the start of one on data streams.
- A Crossref search with the title's words returns no article with that title, but four of its top five results are real papers on building knowledge graphs from several sources, among them a 2026 paper by Zhu and Hu. The PerfectRefs Existence check marked the entry Not found and named the same paper as the closest real record.
Same paper or a different one?
Zhu and Hu (2026) is real: its DOI resolves to Wiley's site. Its title shares words with the fake entry, but the authors, journal, year, and volume all differ. So it is a different paper, a replacement rather than a correction, and it must support the sentence on content.
Reading the abstract against the sentence
| Part of the sentence | What the abstract says |
|---|---|
| Graph neural networks with attention, for entity alignment | Yes, for one framework: "a multi-source entity fusion mechanism employing Graph Attention Networks (GATs) with composite similarity metrics for cross-source alignment" |
| Multi-feature fusion algorithms | Only the "composite similarity metrics" in that alignment step |
| Relation fusion and deduplication | Not mentioned |
| Bayesian optimization and ontology constraints for verification | Not mentioned; it verifies with "semantic, statistical, and rule-based methods" |
| "Improves the reliability" of fusion | Reports recognition and extraction scores, not fusion reliability |
| "Existing studies mostly use" | Describes its own framework, not what most studies do |
Rewriting the sentence and the entry
The paper supports part of the sentence, for one framework, so the sentence shrinks:
Zhu and Hu (2026), for example, align entities across sources with graph attention networks and composite similarity metrics, and verify the fused knowledge with semantic, statistical, and rule-based methods.
The reference list gets the real paper:
Zhu, E., & Hu, Y. (2026). A multi-source fusion and dynamic verification framework for scientific knowledge graph construction: Methods, evaluation, and applications. AI Magazine, 47(2), Article e70062. https://doi.org/10.1002/aaai.70062
A matching name isn't enough
Entry [9] in the same list was "EntGPT: A two-stage refinement model for knowledge graph alignment," by Ding, Zhang, and Liu; Crossref has no such article. A real EntGPT paper exists, by Yifan Ding and five coauthors, none named Zhang or Liu, but it is about entity linking (Ding et al., 2024). Swapping it in for the name would attach your sentence to a paper on another task.
How to search for a replacement
The fake title still names the topic, so start there and let your sentence narrow the search:
- Take the topic words from the fake title and add the key terms of your sentence (method, task, population, outcome). Search them in Crossref Search, Google Scholar, or your library's databases.
- For a specific finding, cite the study that reported it; authors "should provide direct references to original research sources whenever possible" (International Committee of Medical Journal Editors [ICMJE], n.d.).
- Read each abstract against your sentence, part by part as in the table above, then the section that backs it. The Liberty page puts it plainly: "You should only cite sources that you have actually obtained and read" (Miller, 2026).
A quote or number taken from the fake source has no source either. Take it from the new paper, or delete it.
Update everything that pointed to the fake reference
- Replace the entry in the reference list, formatted from the new paper's record.
- In author-date styles, change every citation of the fake entry: in APA 7, (Pons et al., 2025) becomes (Zhu & Hu, 2026). Search for the fake surname. The list is alphabetical, so move the entry to where Zhu belongs (see how to order references).
- In numbered styles, the number stays the same if the new paper takes the fake entry's place. If the new paper is already in your list, cite its existing number and delete the fake entry. If your style numbers by order of first citation, every later number moves up by one, in the list and in the text.
Then check in-text citations against the reference list.
If you can't find a paper that supports the claim, delete the claim. ICMJE says authors "should be able to attest that the references cited support the associated statement" (ICMJE, n.d.).
Finding a replacement in PerfectRefs
In PerfectRefs, a reference whose final result is Not found has a Find a replacement button at the bottom right of its card. On the free allowance, a reference marked "Needs a deeper check" gets the button only if Deep check still can't find it.
Find a replacement starts where the search above starts: it searches by the original title and shows up to three candidates, the real papers closest to the original reference. Under them a line reads: "These are the closest matches to the original reference. Check that each one supports the claim in your text." You paste only your reference list, so it never sees your sentence; reading each candidate against it is still your step. Open one to read its abstract, then click Swap in at the bottom of the side panel. The confirmation reads: "This is a different paper, not a correction of the original. Read its abstract and make sure it supports the claim in your text." For author-date styles and MLA 9, it also tells you what to change the in-text citation to, or reminds you to update it; for numbered styles, the number stays the same.
Every candidate has already been checked and completed against its record, so the new reference takes the original's place with nothing left to fix; in a Style check it is marked Verified. Its card has links to show the original or undo the swap. In the report from a Style check export, it carries a note that it was swapped in, with the original text.
It uses premium credits: 3 per search, whether or not it finds anything, and 2 more when you swap a paper in, once per reference. If a data source fails before the search finishes, you aren't charged. Undoing doesn't refund the credits, and you can swap in another candidate without paying again.
FAQ
What should I do with a fake citation I can't replace?
Delete the entry, and delete or rewrite the claim it supported. Deleting only the entry leaves the claim without a source and, in author-date styles, leaves in-text citations pointing to nothing.
References
- Ding, Y., Poudel, A., Zeng, Q., Weninger, T., Veeramani, B., & Bhattacharya, S. (2024). EntGPT: Entity linking with generative large language models. arXiv. https://doi.org/10.48550/arXiv.2402.06738
- International Committee of Medical Journal Editors. (n.d.). Preparing a manuscript for submission to a medical journal. Retrieved November 17, 2026, from https://www.icmje.org/recommendations/browse/manuscript-preparation/preparing-for-submission.html
- Miller, R. (2026, August 5). What are AI-generated or "hallucinated" sources? Jerry Falwell Library, Liberty University. https://askusatthelibrary.liberty.edu/faq/436066
- Zhu, E., & Hu, Y. (2026). A multi-source fusion and dynamic verification framework for scientific knowledge graph construction: Methods, evaluation, and applications. AI Magazine, 47(2), Article e70062. https://doi.org/10.1002/aaai.70062
Original article by PerfectRefs. To republish it, credit PerfectRefs and link to this page. Do not edit it or use it commercially.
