You are three days from a deadline. A draft chapter cites a paper with a plausible author, a plausible journal, a plausible year, and a DOI that looks like every other DOI. Then you paste the DOI into a browser and land on nothing, or on a different paper entirely. The citation was generated, not found. The reference list now contains a claim that has no source behind it.
This is a repair problem before it is a writing problem. The goal is not to delete the sentence and move on. The goal is to separate three questions that a fake citation collapses into one: does the work exist, is the metadata correct, and does the source actually support the sentence you attached it to? Each question has its own check, and each check produces an artifact you can keep.
This guide walks through that repair for a realistic case: a master's student finishing a thesis chapter over winter break, with supervisor feedback asking for stronger sourcing in the literature review. The same steps apply to a course paper, a first manuscript, or a January project start where you are rebuilding a reference list from scratch.
Why fake citations survive a first read
A fabricated reference is built to pass the checks a tired reader actually runs. The author names sound like real researchers in the field. The journal title resembles a real venue. The year fits the narrative. The DOI has the right shape. None of that is evidence.
What makes the problem tractable is that real scholarly metadata is deposited and exposed in structured form. Crossref's REST API documents publisher-deposited metadata including titles, authors, venues, dates, identifiers, licenses, funding, ORCID and ROR identifiers, abstracts when deposited, and post-publication updates from trusted sources. That means a genuine reference usually leaves a trace in a metadata record you can inspect field by field, not just a string that looks right.
A DOI adds a second, separate signal. A DOI can be resolved through doi.org to the current destination registered for that identifier. Resolution helps locate a work, but the bibliographic fields and the source text still need to be checked for identity and claim support. In other words, a DOI that resolves is a starting point, not a verdict. And a DOI that does not resolve is a strong prompt to stop and verify everything about that entry.
Diagnose the entry before you touch the sentence
Start with the reference itself, not the paragraph around it. Copy the full entry into a scratch file and mark each field: authors, title, venue, year, volume, pages, DOI. Then run three checks in order.
Check 1: Does the identifier resolve? Resolve the DOI through doi.org and look at where you land. If the destination is a different title or a different author set, the identifier has been attached to the wrong work. If nothing resolves, treat the entry as unverified until you find a real record.
Check 2: Does a metadata record match field by field? Search the title and the first author together. Compare what comes back against your entry: exact title wording, author order, venue name, year, and identifiers. Small mismatches matter. A shifted year or a swapped author order often means the generated entry blended two real papers.
Check 3: Does the source support the sentence? This is the check most people skip. Even a real paper can fail to support the claim you attached to it. Read the abstract and, where the claim is specific, the relevant section. If the sentence says a method "outperforms" something, confirm the source makes that comparison.
Record the outcome of each check next to the entry. You now have a diagnosis rather than a suspicion.
Sort the failure into one of four cases
The repair differs by case, and conflating them wastes time.
- Fully fabricated. No record exists for the title and author combination. The entry must be replaced or the sentence must be re-grounded in a source you can find.
- Real work, corrupted metadata. The paper exists but the year, venue, authors, or DOI are wrong. Repair the fields against the metadata record.
- Real work, wrong claim. The citation is accurate but the sentence overstates or misreads the source. Rewrite the sentence to match what the source says.
- Two real works blended. The generated entry merged a real title with a real author from a different paper. Split it and decide which work you actually meant.
Write the case label beside each suspect entry. A reference list with ten labels is a work plan.
Rebuild the evidence base with real records
Once you know which entries are broken, you need real papers to stand behind the claims. This is where a scholarly search interface earns its place in the workflow. Use WisPaper search to look up candidate papers by topic or title, then open each result and confirm the metadata against the record before it enters your draft. If you prefer a public entry point first, the WisPaper Scholar Search page explains the paper-finding starting point.
For each replacement, capture the same fields you marked in the diagnosis step. Keep the resolved DOI, the exact title, and the author order. If the abstract is deposited, save the sentence or two that supports your claim. That saved sentence is what lets you answer a supervisor who asks, "Where exactly does this paper say that?"
When a claim needs more than one source, or when you are rebuilding a whole section, a longer-running literature task may fit better than one-off lookups. The Scholar Agent area is built for longer research tasks including literature review, inspiration, and reproduction use cases. Use it when the job is "map what exists on this question," not "find this one title."
Screen the whole BibTeX file, not just the entry you caught
One fake citation is rarely alone. If a draft was assembled with generated references, others are likely in the same file. Screening the entire bibliography is faster than discovering them one at a time during a final read.
TrueCite is built for BibTeX citation screening: it checks BibTeX against real academic databases to flag hallucinated references. Export your .bib file, run it through the screening step, and treat the flagged entries as the input to the four-case sort above. This turns a spot check into a sweep.
A practical order of operations:
- Export the current BibTeX file from your reference manager.
- Screen the file and collect the flagged entries.
- Label each flagged entry with one of the four cases.
- Repair metadata for real works; replace fabricated ones with verified records.
- Re-check that every repaired entry still supports its sentence.
Keep verified papers where you can find them again
The repair is only durable if the verified papers live somewhere you will return to. Scattering PDFs across a downloads folder is how the same problem reappears next semester.
Use My Library to save or upload papers and organize them in the WisPaper library. As you verify each replacement source, add it there with its corrected metadata. When the same paper is cited in two chapters, you pull from one verified copy instead of re-deriving it.
If you want to interrogate the papers you have collected rather than re-read each one, Library QA lets you ask questions based on papers uploaded to My Library or added from search results. That is useful when you need to confirm whether a collected paper addresses a specific sub-question before you cite it.
Turn the repair into a project you can hand over
A thesis chapter repaired over winter break becomes next semester's starting point. Treat the cleaned reference list and the verified paper set as project assets, not disposable drafts.
The Research Projects area is where you can create a new project and keep the work together. Put the repaired chapter, the verified bibliography, and your case labels in one place. When your supervisor asks for the revised section in January, you are handing over a traceable set of sources rather than a cleaned-up guess.
If the next step after repair is generating directions for the following chapter, Idea Discovery takes research topics, background, references, and experimental constraints to generate research ideas. Feed it the verified references you just repaired so the ideas are grounded in sources that exist.
What "ready" looks like for the corrected deliverable
Readiness is not a feeling that the draft looks fine. It is a set of observable conditions on the artifacts you produced.
- Every DOI in the reference list resolves to the work it is attached to.
- Every entry's title, authors, venue, and year match a metadata record field by field.
- Every in-text claim has a saved supporting sentence or passage from the cited source.
- Every flagged entry from the BibTeX screen has a case label and a resolution.
- The verified papers are stored in one library, not scattered across folders.
- The repaired chapter and bibliography sit together in one project.
When those six conditions hold, you can answer the supervisor's sourcing question without reopening the whole draft. That is the real test.
Transferable lessons for the next project
Verify at the point of entry, not at the point of submission. A citation checked the moment you add it costs seconds. The same citation checked the night before a deadline costs a rewrite of the surrounding paragraph.
Separate existence from support. A resolving DOI proves a work exists. It says nothing about whether the work backs your sentence. Keep the two checks distinct so neither masks the other.
Screen the file, not the entry. Generated references cluster. A BibTeX-level screen catches the ones you did not notice.
Keep the evidence sentence. The single most useful artifact from this repair is the saved passage that supports each claim. It converts a citation from a decoration into a defensible link.
Make the repair reusable. A verified library and a labeled project mean the next chapter starts from clean ground instead of from a fresh pile of unverified references.
Your next concrete action
Open your current bibliography file and export it. Screen it for hallucinated references, then label every flagged entry with one of the four cases. Before the day ends, resolve the DOI on each flagged entry and record where it lands. That single pass tells you whether you are fixing a handful of fields or rebuilding a section, and it gives you the first artifact your supervisor will want to see.




