How to Spot a Fake Quote or an Invented Source, Including Ones Made Up by AI

Two kinds of fake travel well online. One is the misattributed quote: a real sentence, credited to someone who never said it. The other is the invented source: a book, study or court case that does not exist at all. AI chatbots can produce the second kind by accident.

Both can be caught with the same habit: go back to the original, and check that it exists and says what it is claimed to say.

01Why quotes end up with the wrong person

Garson O’Toole, the researcher behind the website Quote Investigator, traces where quotations come from. In an interview with NPR, he explained why so many witty lines get credited to Mark Twain: “If you preface a quotation by saying it’s from Twain, then people are prepared to laugh at it.”

The writer Ralph Keyes, author of The Quote Verifier, has a name for the people words attach themselves to. A newspaper review of his work reports that “Keyes calls some of these people ‘flypaper figures’ because words tend to stick to them.” A second review describes another pattern he found: quotations that start out long and clumsy tend to end up “shorter, more graceful and more melodious in the retelling.”

Put those together and you get the typical fake quote: a polished line, attached to a famous name, with no date and no source. Our quote checks show the pattern many times over.

02How to find where a quote first appeared

The aim is the earliest appearance, with a date — not the most popular one.

• Internet Archive offers full-text search inside digitised books. Its help page explains that “This allows you to search inside books and other text items. It uses the OCRed text file derived from the processing of uploaded text formatted files”. Because it relies on scanned text, it is worth trying spelling variations.
• HathiTrust searches the text of its digitised collection; a University of California help page says “The HathiTrust corpus consists of 17.5 million volumes comprising 6.2 billion pages.” Some are marked “Limited (search-only)”: the same page notes that “No one may access ‘Limited (search-only)’ volumes for reading purposes.” Our reading: even a search-only hit can show that the words appear in a book with a known date, which may be all you need.
• Google Books also searches the full text of books. Its About page says “Book Search works just like web search”, and where rights allow, “you’ll be able to see a preview of the book, and in some cases the entire text.”

What you are looking for: the oldest dated source that contains the words, and whether it credits the person the quote is now pinned on. If the earliest version is decades after the person died, or credits someone else, you have your answer.

03Invented sources: what the evidence shows about AI

The researchers Kalai, Nachum, Vempala and Zhang describe the underlying behaviour: “Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty.” A guessed reference can look exactly like a real one — author, title, journal, year.

The research.

• In a 2023 study in Scientific Reports, Walters and Wilder had two versions of ChatGPT produce 84 short papers on 42 topics and checked all 636 citations. They found that “55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated.” Even among the real citations, “43% of the real (non-fabricated) GPT-3.5 citations but just 24% of the real GPT-4 citations include substantive citation errors.”
• A 2023 study in Cureus by Bhattacharyya and colleagues checked references in medical articles generated by ChatGPT and found that “47% were fabricated, 46% were authentic but inaccurate, and only 7% were authentic and accurate.”

Both studies tested 2023 versions of these tools. Newer models may do better or worse; the point is that a citation from a chatbot is a lead to check, not a source.

Why it happens. Kalai and colleagues’ 2025 paper argues that “language models hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty.” That is their argument rather than a settled consensus.

55%	GPT-3.5 citations that were fabricated (Walters and Wilder)
18%	GPT-4 citations that were fabricated (Walters and Wilder)
47%	References in ChatGPT-generated medical articles that were fabricated (Cureus)
7%	Of those references, authentic and accurate (Cureus)

Both studies tested 2023 versions of these tools.

04When invented sources reach a courtroom

United States, 2023. In Mata v. Avianca, a federal judge in New York found that lawyers had submitted “non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT” and “then continued to stand by the fake opinions after judicial orders called their existence into question.” The court was careful about what it was punishing: “Poor and sloppy research would merely have been objectively unreasonable.” It found bad faith “based upon acts of conscious avoidance and false and misleading statements to the Court” and imposed “A penalty of $5,000”. It also made a point worth keeping: “there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.”

England, 2025. In a judgment of 6 June 2025 covering two cases, the High Court described what these tools can do: “The responses may make confident assertions that are simply untrue. They may cite sources that do not exist.” In one of the two cases, a schedule prepared for the court listed forty-five citations that had been put before it, and “In eighteen instances, the case cited does not exist.” The court referred lawyers in both cases to their professional regulators.

These are not isolated. A database maintained by the researcher Damien Charlotin, which records court decisions where a court found or implied reliance on hallucinated material (and some where AI use was alleged but not confirmed), listed more than 2,000 cases when we checked it in September 2026.

United States, 2023	Mata v. Avianca: fake opinions created by ChatGPT; bad faith found, and a penalty of $5,000
England, 2025	Two cases: in one, eighteen of forty-five citations put before the court did not exist; lawyers referred to their regulators

Existing rules impose a gatekeeping role on attorneys, the US court said.

05Five quick checks for any source

1. Does the DOI work? Many journal articles have a DOI, a permanent identifier. Paste it into doi.org. The DOI system’s own documentation says it shows a “DOI Name Not Found” error page when it “cannot find” the identifier. Our advice: a DOI that works but leads to a different paper is just as much a warning sign.
2. Can Crossref find it? Crossref’s Simple Text Query “allows you to retrieve DOI names for journal articles, books, and chapters by cutting and pasting a reference or reference list into the query box”. Our reading: no match does not prove a source is fake — not everything has a Crossref DOI — but a confident-looking journal citation with no match deserves a closer look.
3. Does the journal exist? Periodicals are given an ISSN, which you can look up in the ISSN Portal. Be careful what that proves: the ISSN International Centre notes that an ISSN “does not guarantee the quality or validity of the contents”.
4. Do the details fit? In the Cureus study, fabricated and inaccurate references often had the wrong volume, page numbers or year. Check that the volume and pages exist in that journal, in that year.
5. Can you find the author? Researchers can register for an ORCID, which the organisation describes as “a unique, persistent identifier free of charge to researchers”. Our reading: not every researcher has one, so a missing ORCID proves nothing on its own.

06Quotes on pictures

For a quote shared as an image, two tools help.

• Google Lens lets you “Use your camera or an image to search”, and can copy text out of an image so you can search for the words themselves.
• TinEye is a reverse image search engine that “uses image recognition technology rather than keywords, metadata or watermarks”; you submit an image “to find out where it came from, how it is being used, if modified versions of the image exist”. One caution from its own help pages: “there might be a long gap between when the image was posted and when our crawlers first found it” — its “first found” date is not necessarily when the image was first posted.

07The short routine

1. Copy the exact words and search for the earliest dated version in book and newspaper archives.
2. Check that the named person could have said it — and that the earliest source credits them.
3. For any study or book, check the DOI, the journal and the details.
4. Treat any citation from an AI chatbot as unverified until you have opened the source yourself.
5. If you cannot find the original, don’t pass it on with a name attached.

How we checked this · Last checked September 2026

• Each point was checked against the sources listed below, including the two court judgments, the studies of AI-generated citations and the search tools' own help pages.

Checked September 2026.

How we check every article →

08Sources