What are AI Hallucinations and Fact Check?

GPTZero's hallucination check reads a document and asks two questions: does each citation point to a real source, and do the sources actually support the claims? The check is available on the dashboard and in the browser extension.

It does not decide whether a sentence is true or false. It shows whether a citation can be found, and whether the sources it retrieves support the claim, contradict it, or only partly address it. Read the source before treating a result as final.

The exact steps to run AI Hallucinations and Fact Check are at the end of this article.

What an AI hallucination is

An AI hallucination is a fluent statement that is false or unsupported. A model predicts likely wording. It doesn't look up a source before it answers, so it can fill in missing or ambiguous details with something that sounds right.

In a paper or essay, that usually shows up in one of these ways:

• A fake citation. The title, authors, journal, DOI, or URL look academic, but the source does not exist. One real detail is not enough: a real author can be paired with a fake paper, or a real journal with a DOI that goes nowhere.

• A misleading citation. The source exists, but it does not say what the document claims it says.

• A fake quote or statistic.

• An overconfident summary. A mixed finding is rewritten as a simple, certain conclusion.

How the check works

GPTZero looks for checkable, objective claims and for the citations attached to them. It then searches online, academic, and public material, including scholarly articles, preprints, and news, and pairs each claim with sources that might support or contradict it. Suggested sources can be copied in MLA, APA, Chicago, IEEE, or BibTeX.

A citation is compared part by part, not as one blob of text. The parts are the title, authors, publication date, URL, DOI, arXiv id, and publisher. Each part can match, match only partly, match weakly, fail to match, or be missing. A title that matches and a DOI that does not is a warning, not a pass.

Claims are judged by how the retrieved sources relate to the sentence:

• Support means the sources agree with the claim, or the cited source backs it.

• A conflict means the sources say something else.

• No clear answer means the sources only partly overlap, mention the topic without backing the claim, or do not settle it.

A suggested source can also be flagged when it looks AI-written. That flag is about who appears to have written the source, not about whether the claim is true.

Where you'll see it

On the dashboard

The tile is labeled AI Hallucinations and says Check claims and cite sources. Enter your text, select that tile before you scan, and click Scan. You can run it alone or with other scans. Open Hallucination in the results. On a phone, that result is labeled Hallucinations.

If the document has no bibliography, the result says “Paste a bibliography to check hallucinated claims and sources.” Cited-source checks need a bibliography and in-text citations. Claims with no citation can still be checked against the web.

Three sections follow. Each shows All clear, or how many items need attention.

• Citations is the existence check. Open a citation and choose Similarity breakdown to see Title, Author, Date, URL, DOI, arXiv ID, and Publisher, each marked Match, Partial match, Weak match, No match, or Missing. The summary counts sources that contain errors or were not found online, may contain errors, were verified online, or are not verifiable.

• Cited claims asks whether the cited source supports the sentence. It counts claims that conflict with the source, have no clear answer in the source, or are supported by the source.

• Uncited claims finds web sources for sentences that are not cited. It counts claims that conflict with online sources, lack clear online support, or are supported by online sources. The note says Highlight text to find citations. When sources are found, Generate Bibliography turns them into a bibliography. A source can be marked Likely AI, Likely human, Likely partly AI-written, or Uncertain if source is AI-generated.

Inside Citations and Cited claims, Sort by severity puts the items that need attention first.

On a free account, the result asks you to upgrade. The title is Upgrade for Hallucination Detector. If you are not logged in, the title is Get a free Hallucination Detector report.

In the extension and Google Docs

On a web page, open the GPTZero extension and choose Fact Check. While that check is open, the widget footer reads Fact Check. In Google Docs, the same item is in the GPTZero sidebar. If text is selected, the menu says Fact Check Selection and checks that selection. Otherwise it says Fact Check and checks the document.

This is the claim-and-source half of the same check. The header is Select the claim to dig deeper. Each claim shows how many sources Support it, Disagree with it, or are Nuanced. A claim with nothing retrieved says No sources found. Open a claim for Sources for this claim. A source can explain itself with Why Support, Why Disagree, or Why Nuanced, and you can copy it in APA, BibTeX, Chicago, IEEE, or MLA.

If no claims are found, the result says “We didn't find any claims in this text” and asks you to try a different text.