ChatGPT Made Up Citations
You asked for sources, you got a clean list with authors, journals, years, and DOIs, and none of it exists. This is one of the most damaging failure modes in AI-assisted writing, because fabricated citations look more credible than real ones. Real references are messy: odd page ranges, awkward titles, inconsistent formatting. Generated ones are tidy, because they were assembled from patterns rather than retrieved from a database. Peer-reviewed testing has put fabrication rates for bibliographic citations high enough that spot-checking is not sufficient: measured rates have run from roughly 18 to 29 percent even on stronger models, and considerably higher on older ones. The rule that follows is simple and non-negotiable: verify every citation individually before it goes anywhere near a submission. This page covers a verification routine that takes under a minute per reference, why the fabrication happens, and how to get the model to help you instead of inventing. This is general study and research guidance, not academic-integrity or legal advice; your institution's rules on AI use always take precedence.
Affiliate link, we may earn a commission at no extra cost to you.
Symptoms You Might See
An exact-title search in Google Scholar returns no results
The DOI does not resolve, or resolves to a completely unrelated paper
Real authors are paired with journals they have never published in
The paper is cited as being in a journal that did not exist that year
Page numbers or volume numbers do not match the journal's actual structure
The reference list looks unusually uniform and tidy for a real bibliography
Common Causes
Citations are generated, not retrieved
Without a live search or a connected database, the model produces text that statistically resembles citations in its training data. Author names, journal titles, and years are assembled because they look right together, not because any record was consulted. Nothing in the process checks that the paper exists.
The request implied sources must exist
Asking for 'five peer-reviewed studies supporting X' makes producing five items the goal. If five suitable real papers are not available to it, the model still produces five, because the instruction was to produce five.
Browsing was not actually used
Even in a model that can search, a given answer may be produced without a search being run. Answers with live retrieval carry links you can click; answers without it read the same but are unverifiable, and the difference is not always visibly signposted.
Paywalled and niche literature is thinly represented
Obscure subfields, non-English sources, and paywalled journals appear less in training data, so the model has fewer real examples to draw on and interpolates more. Fabrication rates rise sharply the more specialized the request.
Formatting requests invite invention
Asking for a fully formatted APA or Vancouver reference list encourages complete-looking entries. Missing fields such as volume, issue, and pages get filled in to satisfy the format rather than left blank.
Step-by-Step Fixes
Search the exact title in Google Scholar
When to try: First, for every citation, without exception
Paste the full title in quotation marks into Google Scholar, PubMed, or Semantic Scholar. A real paper appears immediately. No result on an exact-title search is the strongest single signal that the reference is fabricated. This takes about fifteen seconds per citation.
Resolve the DOI
When to try: Whenever a DOI is provided
Paste the DOI into doi.org. A genuine DOI lands on the paper's publisher page. A fabricated one either fails to resolve or opens something unrelated, which is equally disqualifying. Note that a resolving DOI attached to the wrong title is still a bad citation.
Check that the author actually published in that journal
When to try: When the title search is ambiguous or returns near-matches
Look up the author in Google Scholar or ORCID and scan their publication list for the title and venue. Hallucinated references very commonly pair real, prominent researchers with journals they have never appeared in, which is why author recognition alone is not verification.
Ask for links rather than formatted references
When to try: At the point you ask, not after
Request 'a URL for each source that I can open' instead of a formatted bibliography. A model that cannot retrieve real links will more often say so, whereas a formatting request can always be satisfied with invention. Then open each link.
Turn on search and confirm it ran
When to try: For any research task involving sources
Use the search or browsing mode explicitly and check that the answer carries clickable citations to real pages. Then open them, because a link being present is not the same as it supporting the claim. Retrieval-backed answers are substantially more reliable, but they still need checking.
Give the model your sources instead of asking for sources
When to try: For anything going into a submission or publication
Find papers yourself through your library database or Scholar, upload or paste them, and ask the model to summarize, compare, and cite only what you provided. Explicitly instruct it not to introduce any reference that is not in the supplied material. This eliminates the failure mode rather than detecting it.
Permit an honest 'I do not know'
When to try: In every prompt that asks for sources
Add: 'If you are not certain a source exists, say so and leave it out. Do not produce a reference you cannot verify. An incomplete list is better than an invented one.' Removing the pressure to fill a quota measurably reduces fabrication.
Run a final pass over the bibliography before submitting
When to try: Immediately before submission
Before anything is submitted, check every reference in the list against a database one at a time, including any that came from an earlier draft. Fabricated citations that survive into a final version are treated as a serious integrity issue regardless of how they got there, and 'the AI wrote it' is not accepted as a defense.
How to Prevent This
Verify each citation as it enters the draft, never in one batch at the end
Collect sources yourself and let the model work only with what you supply
Prefer requests for openable links over requests for formatted reference entries
Keep a record of where each source came from so a final check is fast
When to Contact Support
There is no support ticket that fixes this, because generating plausible text is what the model does when it has no retrieval to draw on. If you are facing an academic-integrity process over citations that turned out to be fabricated, your route is your institution's appeals procedure and your supervisor rather than the AI vendor. Gather your drafting evidence, such as document version history and research notes, as early as you can. This page is general information, not legal or academic-integrity advice, so follow your institution's published policy and any advice from your student union or academic advisor.
Frequently Asked Questions
Why does ChatGPT invent citations that look completely real?▼
Because it is generating text that matches the pattern of a citation rather than retrieving a record. Author names, journals, and years are combined because they fit together statistically. Nothing in that process verifies the paper exists, and the result looks more polished than a real bibliography precisely because it was assembled rather than collected.
How do I check whether a ChatGPT citation is real?▼
Search the exact title in quotation marks in Google Scholar, PubMed, or Semantic Scholar. If nothing comes back, treat it as fabricated. If a DOI was given, paste it into doi.org and confirm it resolves to that exact paper. Finally check the author's own publication list for that title and journal. Under a minute per reference.
Does turning on web search stop the problem?▼
It reduces it substantially but does not eliminate it. With retrieval, answers carry links to pages that exist. You still need to open each one and confirm it actually supports the claim it is attached to, because a real link can be cited for something it does not say.
How common are fabricated references?▼
Common enough that spot-checking is unsafe. Peer-reviewed testing of bibliographic citations found fabrication in a substantial share of generated references, with rates that remain material even on stronger recent models and that climb further in specialized or non-English literature. Assume every unverified citation is suspect until you have checked it.
What if fabricated citations already made it into my submitted work?▼
Raise it yourself as early as possible rather than waiting to be asked, and bring your drafting evidence such as version history and research notes. Institutions treat proactive disclosure differently from a discovered fabrication. This is general information rather than advice on your specific case, so check your institution's policy and talk to your supervisor or student union.
Sources and verification
Product behavior and limits can change. These primary sources were used to verify this guide.
Other Common AI Errors
ChatGPT Not Working
Service Issues
🚦Rate Limit Exceeded
Usage Limits
📏Context Length Exceeded
Input Limits
🔥Model Overloaded
Service Issues