Why do ChatGPT's references turn out to be fabricated?

Language models do not search for sources; they generate text that looks like a source. Author names, journal names and DOI patterns are learned shapes, so the entry looks convincing while no record exists. The same applies to Claude, Gemini, DeepSeek and other assistants.

Do assistants with web search (Perplexity, Copilot, ChatGPT search) fabricate too?

Less often, but yes: they can extract the wrong author or year from a page they found, or merge a summary with a citation. Checking is still needed; the tool corrects such 'real but wrong' entries with the record data.

Is checking an AI-written list a plagiarism risk?

No. The check compares your list with scholarly records; your text is not stored and is never sent to a similarity database. See the Privacy Policy.

What do I put in place of a fabricated reference?

The Find Evidence tool searches for a real, open-access source that supports your sentence and suggests wording the source safely backs.

For writers who use AI assistants

Are the references your AI gave you real? Check before you submit

ChatGPT, Claude, Gemini, DeepSeek and similar assistants produce fluent but non-existent references: real authors, plausible titles, DOIs that lead nowhere. Paste your list and every entry is compared with Crossref, OpenAlex, PubMed and other scholarly records.

ChatGPTClaudeGeminiDeepSeekCopilotPerplexityGrokLe Chat (Mistral)Qwen

Check AI-generated references Read the full guide

The three most common problems in AI references

Entirely fabricated entry

Author, journal and year look realistic; the article exists in no index. When no record is found and the DOI does not resolve, the entry is flagged red.

Real author, invented paper

A well-known author is credited with a paper they never wrote. The author's real publications are searched for a match; if none, the entry is marked suspicious.

Wrong DOI, year or journal

The paper is real, but the DOI points to another article or the year differs from the record. The correct record data is shown and a corrected entry is produced.

Six signs of a fabricated reference

The DOI does not resolve — If doi.org returns 'DOI not found', the reference is almost certainly fabricated — the single most reliable sign.
A title that fits too well — If the title reads as if written to support your exact sentence, be suspicious; real titles are rarely that tailored.
Author–journal mismatch — A prominent author shown in a journal they never published in — check their real publication list.
Inconsistent volume and pages — Vol. 12, no. 3, pp. 45–60 looks tidy, but if the journal had no such volume that year the entry is invented.
Not a single hit in Google Scholar — Search the title in quotation marks; zero results means the paper does not exist.
The assistant is 'sure' but cannot give a link — If asking for a DOI or link produces an apology and a new citation, the first one was invented too.

A safe workflow for AI-assisted writing

Check the reference list

Paste the assistant's citations; fabricated, faulty and real references are separated by colour.

Check claim–source agreement

Even a real source may not support your sentence; the claim is compared with the source content.

Find evidence for unsupported sentences

For a sentence with no or wrong source, real open-access evidence is searched.

Rewrite the bibliography in your style

Verified records are regenerated in APA 7, Vancouver, IEEE, Chicago and 16 more styles.

Before submission

Does every DOI actually resolve?
Did the author really write that paper?
Do journal, volume, issue and year match the record?
Does the source really support the sentence citing it?
Is any source retracted?

The same test, whichever assistant you used

Where a reference came from does not matter; every entry is compared with the same open scholarly records. No AI is involved in the reference verdict — it is a deterministic comparison of records. Not every unfound reference is fabricated: new, local or niche publications can be missing from indexes; verify yellow results by hand.