If the references an AI gave you cannot be found in any database, the odds are you did not search wrong — it fabricated them: plausible authors, journal names, years, volumes and issues, except the paper does not exist. This is one of the most typical hallucination scenarios, because the model's job is to generate text that looks like a citation, not to retrieve real papers from a database. You do not need intuition to judge; the four steps below verify one reference in about ten minutes.
Why the fakes look so convincing
The model has seen enormous numbers of citation formats in training: which journals a field publishes in, how titles are phrased, how author names are ordered — it has absorbed the statistical patterns. Ask it for sources and it assembles one on the spot following those patterns. The more standard the format, the less it proves — formatting correctly is precisely what it is best at.
The four-step check
Step one: search the exact title. Put the full title into Google Scholar, CNKI, or PubMed, character for character. A real paper almost always surfaces its original page; if only a few scattered words match and the full title finds nothing, treat it as fake for now.
Step two: verify the DOI. Most legitimate papers carry a DOI shaped like 10.xxxx/xxxxx; look it up on doi.org or Crossref. If the DOI resolves to nothing, or resolves to a different paper, discard the citation outright.
Step three: check the journal's volume and issue. On the journal's official site, open the archive for that year: did that issue exist, and do the page numbers fit? Fabricated references often break here — a journal that never had a volume 40 in 2019, for instance.
Step four: check the authors. Search each name plus the research area: do these people exist, and have they published in this field? Fake citations often pair real scholars with the wrong field, or invent people entirely.
Signals that settle it immediately
An unfindable title plus an invalid DOI — with those two combined, no further checking is needed. A citation giving only journal and year, with no volume, issue, or pages, counts as unverified by default. There is also a subtler case: the paper is real, but it does not say what the AI claims — read at least the abstract before citing, and confirm it actually supports your sentence.
The right way to use AI with references
For papers, reports, and grant materials, never let a general chat model produce references from thin air. The safer workflow: find the real papers in a database yourself first, then hand the AI the abstracts or full texts to summarize, compare, and polish; or use a tool with live retrieval that returns clickable sources, and open each one to verify. Anything submitted under your name must pass the four checks above for every single reference — in academic settings, responsibility for a fabricated citation belongs to the submitter, not the tool.
