ChatGPT made up a quote
Open the source and search for a distinctive run of four or five words from the middle of the quotation, not the whole sentence. If those exact words are not in the source, the quote is wrong, and it is wrong in the way that costs you most: a quotation is the only sentence in your essay that a marker who knows the source can check in about ten seconds. This is measured, not folklore. Journalists at 22 news organisations reviewed 1,053 AI assistant answers containing a direct quote and found significant accuracy problems in 12% of them, in a test where every assistant had been told which sources to use. Below is why retrieval does not fix this, and the one-minute check that does.
First, the part that decides how worried to be: a tidied quote and an invented one are different problems, and the same one-minute check separates them.
A quote is generated text wearing a costume
Everything else a chatbot writes is understood to be its own words. A quotation is a factual claim about what somebody else wrote, in a format that says these are not my words, they are theirs, character for character. The model produces that claim the way it produces every other string, by predicting text. Nothing in the machinery treats the inside of a quotation mark differently from the outside.
So a quote is generated text wearing the costume of retrieved text, and the costume is the problem. Quotation marks promise an exactness the process behind them cannot keep by default.
This page is about the words between the quotation marks. Its neighbour, ChatGPT cited a paper that says something else, is about the sentence around them: a real source that does not support the claim you attached to it. Same family, different error. If your sentence reports what a study found and the study found otherwise, that page. If your sentence contains quotation marks and the source does not contain those words, this one.
The two also fail on different timescales. An unsupported claim takes a marker minutes to disprove, because they have to read enough of the paper to be sure it never says that. A misquote takes one keyboard shortcut: open the source, paste the quoted string, get nothing back. That asymmetry is why this error deserves an hour of your attention when a wrong page number does not.
Retrieval does not protect the words inside the marks
The reasonable instinct is that a tool which searches before it answers cannot misquote, because it has the source in hand. The largest study of the question says otherwise, and it was designed by the people whose words were being quoted.
In October 2025 the BBC and the European Broadcasting Union published News Integrity in AI Assistants. Twenty-two public service media organisations across 18 countries and 14 languages put 30 common news questions to the free consumer versions of ChatGPT, Copilot, Gemini and Perplexity in late May and early June 2025, then had their own journalists grade the answers. Every prompt asked the assistant to use that organisation's sources where possible, so each answer was a retrieval task with a named target. That left 2,709 graded answers, 1,053 of which contained at least one direct quote.
| Assistant | Answers containing a direct quote | Significant issues with quote accuracy |
|---|---|---|
| All four combined | 1,053 | 12% |
| Gemini | 290 | 20% |
| Copilot | 190 | 4% |
| ChatGPT | 262 | Not stated separately |
| Perplexity | 311 | Not stated separately |
The report names rates only for the combined figure, Gemini and Copilot, so the last two rows are blank on purpose rather than estimated off the chart. Copilot's 4% came with the fewest direct quotes of any assistant, which is its own lesson: the safest way to avoid misquoting is to quote less.
Two things make this unusually strong evidence. The graders were the newsrooms whose reporting was cited, so they knew the source text better than any outside reviewer could. And the assistants were not working from memory: they had been pointed at a publisher and were citing links. Grounding a model in a real document lowers the misquote rate. It does not remove it. Why AI makes up citations covers the underlying behaviour, and does Perplexity make up citations looks at the search-first tools.
Academic sources have been measured once, in the friendliest possible setting. CSIRO and Australian National University researchers ran four systematic literature reviews with the paper already loaded into the model's context and one instruction, extract exact quotes from it, then checked every quote against the document. McGinness and colleagues report error rates of 5% for GPT-3.5 Turbo and 2% for GPT-4 Turbo. Best case, single job, source in hand: still two to five quotes in a hundred that do not match.
What nobody has measured is the case you actually care about. There is no study of how often a chatbot misquotes a journal article inside a student essay. The two numbers above bracket it loosely, 12% in the wild against 2% under laboratory conditions, and anyone who quotes you a precise figure for your situation has invented it.
The four ways a quotation goes wrong
The BBC and EBU evaluators documented each of these with named examples. They are worth separating because the fix differs.
Invented outright. Asked whether Viktor Orban is a dictator, Gemini reported that he is described as Putin's bridgehead in the EU. The phrase, quoted in German in the original answer, is not in the source given for it and appears to be fabricated. This is the rarest of the four and the easiest to catch, because no search anywhere finds it.
Smoothed. This is the common one and the one people miss. Asked why Birmingham bin collectors went on strike, Perplexity produced three problem quotes in a single answer. Two, attributed to the Unite union and to Birmingham City Council, are not in the cited sources. The third was altered from the BBC article it cited:
| What the BBC source reported | What the assistant printed inside quotation marks |
|---|---|
| The bottom line is that our members can't afford to have savage pay cuts of up to £8,000 with no mitigation. | Our members cannot withstand drastic pay reductions of up to £8,000 without any compensation. |
Read those twice. The meaning survives, the number survives, and almost every content word has been swapped for a synonym: afford became withstand, savage became drastic, no mitigation became without any compensation. That is not a malfunction. It is paraphrasing, the most ordinary thing a language model does, applied to text that must not be paraphrased and then wrapped in marks saying it was not.
Misattributed. Answering a question about the Ukraine minerals deal, ChatGPT took a correctly worded description of the deal as balanced and fair and gave it to Volodymyr Zelensky. It was said by the then economic affairs minister, Yulia Svyrydenko, and the cited source carried the correct attribution. The words were right there. The name was reassigned anyway.
Quotation marks around a paraphrase. An NPR evaluator noted that Perplexity failed to use quote marks to distinguish between direct quotes and paraphrases, calling it misleading and something that can be considered plagiarism. This version is the likeliest to reach your draft intact, because nothing about it looks broken. How to avoid plagiarism covers the boundary.
What the style guides let you change
Students tend to assume a lightly reworded quote is a small sin. Every major style guide disagrees, and the size of what they do permit is the surprise.
| Style | What you may change without telling the reader |
|---|---|
| APA | The case of the first letter, some end punctuation where meaning is unaffected, single to double quotation marks, and footnote callouts. Any other change, including omitting words, must be explicitly indicated. |
| MLA | Note numbers and symbols, in-text citations, typographic flags and most formatting features. Otherwise transcribe the source exactly. |
| Chicago | Obvious typographic errors, corrected silently, unless the source is an older, manuscript or unpublished work. Section 13.7 otherwise requires reproducing exactly not only the wording but the spelling, capitalization, and internal punctuation of the original. |
The permitted set is entirely typographic. Capitalisation at the join, the shape of a quotation mark, a footnote marker, the odd stray typo. Not one of the three lets you substitute a word. APA puts the rule in a single sentence: the quotation must match the wording, spelling, and interior punctuation of the original source, even if the source is incorrect. You are obliged to carry the source's mistakes across unchanged, marking them if they confuse.
Which inverts how the smoothed quote reads. The model did not damage the sentence, it improved it, and a tidier quotation breaks the rule as squarely as a clumsy one, because the rule is about identity rather than quality. APA's writing advice is the instruction this whole page is built on: when quoting, always check your paper against the source to ensure that there are no discrepancies.
Checking a quotation in under a minute
Open the source. Then do not search the sentence.
Searching the full quoted sentence is the mistake almost everyone makes, and it fails so often that people decide they have the wrong version of the source, or give up and trust the quote. We measured why: 373 sentences from four real PDFs, three machine learning papers and the BBC and EBU report itself, scored for anything a person cannot reproduce by typing.
| Feature that breaks an exact-string search | Share of the 373 sentences |
|---|---|
| A word hyphenated across a line break | 18% |
| Curly quotation marks | 15% |
| A curly apostrophe | 8% |
| An en dash or em dash | 7% |
| At least one of the above | 42% |
| Contained a clean five-word run with none of them | 99.7% |
Four in ten sentences carry a trap, and almost every one of them still contains a short window that does not. Your keyboard's apostrophe and the PDF's apostrophe are different characters, and a search box compares characters, not intentions. A word split across a line as hallucin- and ate is stored as two pieces. Five words from the middle of a line dodge all of it.
Four documents we already had is not a random draw of the literature, so treat those percentages as ours rather than as a constant of nature. The direction is not in doubt, and it is not only us. The CSIRO team above hit the same wall: yes or no checking of the quotes, they report, failed because of trivial errors related to unicode characters, so they moved to fuzzy matching, robust to unexpected unicode characters and hyphenations. Professional researchers could not make exact matching work either. You are not doing it wrong.
So, the routine:
- Pick four or five distinctive words from the middle of the quote. Content words, not of the. Avoid anything with an apostrophe, a dash or a hyphen in it. Search that.
- If it fails, pick a different window. A first failure is usually typography. A second, elsewhere in the sentence, is meaningful.
- Drop to three words if you have to. Short enough survives a hyphenated break, and you can read outward from it by eye.
- For a book, search Google Books. It indexes the interior of print books and takes a phrase in quotation marks. A snippet is enough to confirm wording. If the book has no preview, the quote cannot be settled that way and you need the page itself.
- Read what you find, do not just note that you found it. Half the failures on this page are near misses: the passage exists and the words differ.
Three outcomes. The words match and you are done. The words are close but not identical, so you are holding a paraphrase: quote the source correctly or drop the marks and cite it as a paraphrase. Or nothing resembling it is there, in which case stop treating it as a quote and check whether the reference itself is real. How to check if a citation is real is the full routine, and our citation checker takes a whole reference list at once.
When the words are real but the speaker is not
Sometimes the fragment search succeeds and the quote is still wrong, because the words belong to somebody else. The Ukraine minerals case is the blunt version: a quote handed to the wrong person while the cited source named the right one.
The version you meet in an essay is quieter. You search the paper you were told to cite, find the words, and they sit inside that paper's quotation marks. Your source was quoting somebody else and the model collapsed the two, leaving what looks like a direct quotation from a work you have never opened.
That is a secondary source, and it has a rule. APA asks you to find the primary source, read it, and cite it directly, using secondary citation sparingly, for instance when the original is out of print, unavailable, or in a language you do not read. When you cannot reach the original, name the primary source in your sentence, add as cited in the work you actually used, and list only that work in your references.
The tell is simple. When you find the quote, look at whether it sits in the author's own prose or inside their quotation marks. If it is nested, you have a secondary source, and citing it as though you read the original is the error, even though every word is right.
If you have already handed it in
Check the others first. A misquote rarely travels alone, because whatever produced one produced the rest of that draft's quotations the same way. Ten minutes of fragment searches across every quoted sentence tells you the size of the problem, which is what you need before saying anything to anyone.
Then tell your marker before they find it, in one plain sentence: the quotation in paragraph four is not accurate, here is the source wording, here is the correction. No paragraph of apology. A student who reports their own citation error reads as someone who checks their work. The same error found by a marker who then opens two more quotes is the version where the whole reference list gets examined. If the sources themselves do not hold up, my professor said my sources do not exist covers that conversation.
Be accurate about which failure you had. A tidied quotation is a citation accuracy problem. One invented and pinned on a named person is worse in kind, and pretending otherwise will not survive the follow-up question. If you fix the draft by turning quotes into paraphrases your length changes, which is its own small headache: do quotes count in the word count has the rules.
The habit that prevents all of it costs ten seconds a quote. Copy the words out of the source rather than out of the answer, and keep the passage beside the sentence while you write. A quotation pasted from the document you were reading cannot be smoothed, because nothing rewrote it in between.
That is the arrangement CiteOwl is built around. It searches OpenAlex, CrossRef, Unpaywall and the web, reads what it retrieves, and writes each claim from a passage it keeps beside the sentence, so the words attributed to a source are the words that came out of it, and every edit arrives as a diff you accept or reject before it enters your draft.