Which AI research tools make up citations, and how to catch one before you submit
Yes. Chat assistants can produce references that do not exist or contain wrong details, and a real source link does not prove that the paper supports your claim. Tools that retrieve records reduce invention, not the need to check. Before you submit, resolve every DOI, find the exact title in a scholarly index, and read the cited passage.
A real link is not the same as a supported claim
A fabricated reference points to a work that does not exist. An unsupported citation points to a real work that does not support the sentence. A half-real reference combines real details with invented ones: the author may exist while the title, year or pages are wrong.
Check existence and support separately. A DOI that resolves is evidence of a record, not evidence for your argument. A polished bibliography can still contain either kind of failure.
What the studies actually measured
These are different denominators. The chatbot study measures requested references; the proceedings study measures papers containing a hallucinated reference and counts those references. Neither is a current ranking of research tools.
| Study | Task and findings | What it does not establish |
|---|---|---|
| Cabezas-Clavijo & Sidorenko-Bautista, arXiv 2505.18059 (May 2025)Source 1 | Of 50 references requested per tool, using free versions, Copilot 100%, Perplexity 72% and Claude 64% were wrong or fabricated. | Wrong or fabricated is not the same as wholly invented. The result applies to the free versions and prompts tested in 2025, not today's models or all research tasks. |
| Denny et al., arXiv 2609.16574 (September 2026 preprint)Source 1 | 2.3% of SIGCSE TS 2026 papers had at least one hallucinated reference, a lower bound. The authors found 17 hallucinated references in 2026 versus 3 in 2025. | This is a proceedings audit, not a chatbot test. The percentage describes papers, not references; 17 and 3 count references, not papers. It is not a rate for all academic publishing. |
What to check in the tools you use
Use the dossier failure modes below as reminders, not measured error rates. Generating a bibliography, retrieving a paper record and summarising an uploaded PDF are different tasks. Each still needs a check against the original work.
ChatGPT
Dossier failure mode: Breadth can hide weak methodology
Open every original source, even when the answer includes browsing links.
Claude
Dossier failure mode: Research/web answers still need source rejection
Match each research claim to a passage, not just a linked document.
Gemini
Dossier failure mode: Search grounding can cite SEO/support/forum/stale pages
Apply the same source and passage checks to Deep Research reports.
Microsoft Copilot
Dossier failure mode: Copilot products have different data boundaries
Compare the title and authors with the original bibliographic record.
Perplexity
Dossier failure mode: Pro tier changes can degrade value for existing subscribers without clear migration
Follow each link and check whether it actually supports the sentence.
Consensus
Dossier failure mode: AI summaries can over-compress uncertainty
Cite the paper you read, not the assistant's answer text.
Elicit
Dossier failure mode: Pricing jumps quickly from free to Pro/Scale
Check extracted findings against the full text and its qualifications.
Scite
Dossier failure mode: Reference Check can miss non-DOI/difficult-format sources
Manually check sources that a reference checker does not resolve.
NotebookLM
Dossier failure mode: Audio Overviews can omit caveats
Locate the cited passage in the uploaded original, and check what your corpus omits.
How to check before you submit
Run the steps below for every reference you keep. Do not ask the same chatbot to certify its own bibliography. If you cannot locate a work or the supporting passage, mark the claim unresolved and do not cite it as verified.
Resolve the DOI
Open the DOI at doi.org. Does it lead to the same title and authors? A working link to a different paper fails this check.
Search the exact title if there is no DOI
Search the title in quotation marks in Crossref, OpenAlex, Semantic Scholar or Google Scholar. Check the publisher or library record too. No result is a reason to investigate, not proof that the work is fabricated.
Compare the bibliographic details
Compare authors, year, title, venue and volume/pages with the original record. Fix mismatches rather than preserving a plausible-looking citation from the chat.
Read the passage that supports your sentence
Open the full text and locate the evidence for the exact claim. Record the page or section and any limits on population, method or outcome. If you cannot access the text, do not treat the assistant's summary as verification.
Check for retractions or publisher notices
Check the publisher page and the Retraction Watch database for a retraction or notice affecting the work. Record what you found and when you checked.
Log the source and the claim
Save the verified record in the source ledger. In the claim table, connect your sentence to its supporting passage, page or section, and any caveat. Keep unresolved entries visibly unresolved.
Find and save sources before asking AI to summarise
Start with a scholarly index, open the record and save the verified work in your reference manager. Only then ask an assistant to extract or summarise a PDF you already have permission to use. Keep the claim, passage and source together rather than citing an answer paragraph.
What if you already submitted a bad reference?
Document the error, identify the affected sentence and contact your supervisor or editor. Explain what you checked and provide the correct source if one supports the claim. If no source does, propose removing or revising the sentence. Ask how to file a correction; do not quietly replace a submitted version.
VerdictPal's own results are not published
Citation Fidelity is a public protocol, not evidence for a tool ranking here. Human grading and sign-off are pending; this guide quotes no Lab scores or results.
Questions this guide answers
Every answer below is assembled from the dated fields on this page. Nothing is written separately for search.
Does ChatGPT make up citations?
Yes, it can. A convincing title, author list or DOI is not proof that a reference exists. Treat generated references as leads: locate the original record, compare the bibliographic details and read the supporting passage before you put the reference in a thesis or paper.
Why does AI make up sources?
Generating plausible text is not the same as checking a bibliographic database. An assistant can produce a citation-shaped answer without verifying the record or the passage. Retrieval helps you locate material, but you still need to confirm that the record is correct and that the work supports the claim.
Can Perplexity be wrong even when it shows links?
Yes. A link can lead to a real page while the answer misstates or overextends what it says. In the 2025 free-version test by Cabezas-Clavijo and Sidorenko-Bautista, 72% of 50 requested Perplexity references were wrong or fabricated. That is not a current error rate for all Perplexity searches.
Which AI never hallucinates references?
This guide establishes no such guarantee. A tool that retrieves real records can still misattribute a finding or omit relevant literature. Choose a workflow that makes checking possible: original records, accessible full texts and a claim table. Do not replace verification with a vendor label or an old study ranking.
How do I check whether an AI citation is real?
Resolve the DOI or search the exact title in a scholarly index, then compare authors, year, title, venue and pages. Read the full-text passage supporting your sentence and check for retractions. Save the record and passage together. A resolving DOI alone verifies neither the bibliographic match nor the claim.
Can I ask the same chatbot to check its references?
You can ask for leads, but its reply is not independent verification. Use the publisher or library record and the original text instead. If the assistant supplies a new DOI or a replacement reference, run the same checks on that new item rather than treating the correction as proof.
Is Deep Research safer for citations?
A longer report and more links do not establish citation accuracy. Treat each reference and each claim as a separate check: does the work exist, are its details correct, and does the cited passage support this sentence? The studies quoted here do not establish a current Deep Research reliability guarantee.
What if I already submitted a fabricated reference?
Contact your supervisor or editor with the affected sentence, the incorrect reference and your verification notes. Offer a supported correction or remove the unsupported claim. Ask which correction process applies to the submitted version. This is practical guidance, not a prediction of institutional sanctions or legal outcomes.