How to Spot AI-Fabricated References (ChatGPT, Claude, Gemini, DeepSeek…)

August 28, 2026 · 9 min read

ChatGPT, Claude, Gemini, DeepSeek, Copilot, Perplexity, Grok, Mistral's Le Chat, Qwen — whichever assistant you use, the shared weakness is the same: producing academic sources that do not exist, known as “hallucination”. This guide covers the practical signs of a fabricated reference, how assistants differ, and how to verify and repair your bibliography in minutes.

Why do AI assistants invent references?

Large language models do not retrieve text from a database; they generate text that looks probable given their training data. They know exactly what a reference looks like: surnames and initials, a year in parentheses, a sentence-case title, an italic journal name, volume(issue), a page range and a DOI beginning with 10.. So the invented entry is formally flawless — a real researcher's name, a real journal and a correct-looking DOI come together, but the paper was never written.

What changes from assistant to assistant?

• Chat without web search (the default modes of ChatGPT, Claude, Gemini, DeepSeek, Grok, Le Chat, Qwen).The riskiest case: the model writes “source-shaped” text from memory. Well-known authors and major journals are invented most often, because they are the most frequent patterns.
• Assistants with web search (Perplexity, Copilot, ChatGPT search, Gemini with Google grounding).They fabricate less but produce a different error: the wrong author or year extracted from a page, a news summary mistaken for the paper, parts of two sources merged into one entry. The result is a “real but wrong” reference, which is easier to miss.
• Deep-research / agent modes. The list is long and mostly real; the risk is that a source does not support your sentence. The source exists, the citation is wrong.
• Non-English topics. Journals in smaller languages are sparse in training data, so fabrication rates are markedly higher; the reference checker also searches regional indexes for entries without a DOI.

Six signs of a fabricated reference

1. The DOI does not resolve. Append the DOI to https://doi.org/and open it; “DOI not found” almost always means fabrication. Several DOIs can be checked at once with the DOI checker.
2. No hit in Google Scholar. Search the title in quotation marks. No results, or only unrelated works, is a red flag.
3. Author–topic mismatch. The author may be a real academic who never worked on that topic. Check the publication list (ORCID, Google Scholar); the author search tool pulls real publications straight from OpenAlex.
4. Journal–article mismatch.The journal is real but the stated volume/issue contains no such article; check the issue on the journal's archive page.
5. An over-tailored title. Titles that confirm your exact sentence, too specific to be true, are a classic hallucination sign.
6. A composite entry. The authors of one real paper + the title of another + the name of a third journal. Make sure every field belongs to the same record.

A wrong DOI is not the same as a fabricated DOI

The two look alike on screen but mean different things. A fabricated DOI exists in no record, and the rest of the entry is usually invented too. A wrong DOIbelongs to a real paper: a single shifted character, or the DOI of a different article. The checker flags both red — a DOI mismatch is grounds for suspicion — but when title and authors match strongly it shows the record's correct DOI and uses it in the corrected entry. If only the year differs (common for online-first papers) the source gets a yellow warning, not a suspicious flag.

Real author, invented paper

This is the most insidious form: a prominent name in the field is credited with a paper they never wrote but plausibly could have. The author checks out, the journal checks out, the year is reasonable — only the paper is missing. Automated checking searches the author's real publications for a similar title; if one exists it says “a real paper by the same author with a similar title was found — this may be what was meant”, otherwise the entry is marked suspicious. Opening the author's profile page takes under five minutes and spares you the committee's question.

How to verify a whole bibliography

Searching entry by entry takes hours for a long list. The faster route is to run the entire bibliography against open scholarly databases (Crossref, OpenAlex, PubMed, Europe PMC, Semantic Scholar, plus regional indexes) automatically:

1. Copy your reference list or prepare the Word/PDF file (the assistant's list can be pasted as is).
2. Paste it into the reference checker or upload the file.
3. Read the results: green = a matching record was found; yellow = the source was searched but there is no certain match, check by hand; red = dead or wrong DOI, or a strong conflict — most likely fabricated or misattributed.
4. Verified sources are rewritten in your chosen style (APA 7, Vancouver, IEEE, Chicago…); use “Copy corrected bibliography” to paste the whole list back into your manuscript.

What do I put in place of a fabricated source?

Deleting the red entry leaves the sentence unsupported. Find Evidence searches for a real, open-access source that supports your sentence as written; if it finds one it gives the reference in your style and suggests wording the source actually backs — you will usually need to soften the sentence, because no source is as certain as the assistant was. If nothing is found it says so, and the sentence is either rewritten as an unsourced observation or removed.

Even when the source is real: does it support the sentence?

Sources from search-enabled assistants and deep-research modes are mostly real; the problem is thecontent of the citation. The source may say the opposite, or merely touch the topic. Citation Check compares every claim in your text with the abstract or full text of the cited source and marks it full / partial / unsupported / could-not-verify. That is the second step after weeding out fabricated references.

FAQ: is a green result a guarantee?

No. Automated verification is a strong first filter; green means the entry matches a reliable scholarly record consistently. For critical work, opening the DOI link and the journal page remains a good habit. Nor is every unfound source fabricated: very new, local, book, thesis or niche publications can be missing from indexes — verify yellow results by hand. The real rule is never to submit with a red or yellow entry you have not looked at.

How to find real sources without the fabrication risk

The most robust approach is not to ask an assistant for sources but to find them directly in a scholarly database. The literature search tool scans real publications on a topic from OpenAlex; every listed item is a real record, so there is no hallucination risk. Use the assistant to edit and summarise, not to find sources — and if it did find some, check them before submission.