Most coverage of AI hallucination in legal filings focuses on fake case names — citations to cases that simply do not exist. A quieter, arguably more dangerous failure mode gets less attention: a real, existing case, correctly cited, carrying a quotation the case never actually contains.
Why fabricated quotations are harder to catch
An existence check passes immediately, since the case is real. A skimming reviewer who recognizes the case name may not think to verify the exact quoted language, especially if it reads plausibly in context. This combination — a real citation with fabricated content — is specifically why an existence-only verification process is insufficient, and why the Charlotin AI Hallucination Cases database documents this as a distinct, recurring category of error.
Why AI tools generate fabricated quotations
A language model generating a quotation is producing statistically plausible text in the style of judicial writing, attributed to a real case it may have genuine (if imperfect) training exposure to. The model can correctly recall that a case exists and addresses a certain topic while still generating quoted language that was never actually written by that court, because the underlying generation process is not a database lookup unless the tool has retrieval-augmented access to the actual opinion text.
How to catch it
Take any quoted language in your draft and search for a short, distinctive fragment of it (five to eight consecutive words) using a full-text search tool like CourtListener's. If the fragment does not appear in that specific opinion, the quotation is fabricated or misattributed, regardless of how accurate the surrounding paraphrase feels.
Why courts treat this as seriously as fake case names
A fabricated quotation attributed to a real judge is arguably more misleading than an invented case name, because it lends false authority to language the court never actually wrote, potentially shaping how a reader (including the presiding judge) understands the cited case's holding. Courts applying Rule 11 or equivalent sanctions frameworks have not distinguished between the two failure modes in terms of severity.
Expert perspective
Researchers behind the Stanford RegLab/HAI hallucination studies have specifically flagged fabricated quotations as a category that is likely under-detected relative to fabricated case names, precisely because existence checks (the most common and fastest verification step people actually run) do not catch it.
A quotation-checking checklist
- Full-text search a distinctive fragment of every quotation, not just the citation it's attributed to.
- Confirm the surrounding context matches how you're using the quote, not just that the words appear somewhere in the opinion.
- Treat block quotes with the same scrutiny as short inline quotes — length does not correlate with accuracy.
A common question
Can a paraphrase, not in quotation marks, also be a problem?
Yes, if it misrepresents the case's actual holding, even without direct quotation marks. This overlaps with the proposition-support check: confirm the case actually says what you're characterizing it as saying, quoted or not.
Related reading
- Existence, Quote, and Proposition Checks: What Each One Actually Catches
- How CourtListener Works for Lawyers
- AI-Hallucinated Case Law: Real Examples From the Public Record
- How to Verify a Legal Citation in Under 2 Minutes
- Can ChatGPT Cite Fake Cases?
Check a brief before you file it →
Why fabricated quotations survive review more often than fake case names
A fake case name is a binary failure: the case either exists in a reporter or it does not, and a single database search resolves the question definitively. A fabricated quotation attributed to a real case is a more subtle failure, because the surrounding sentence often reads coherently, the citation format is correct, and a reviewer skimming for obviously wrong case names has no reason to slow down and check the quoted language word for word. This asymmetry in detection difficulty is likely part of why fabricated quotations appear to be underrepresented relative to fake case names in documented sanctions cases: they are not necessarily less common, they may simply be caught less often.
A specific technique for catching this reliably
Rather than reading a quotation and asking "does this sound right," extract a short, distinctive phrase from the middle of the quotation, five to eight words that would be unlikely to appear verbatim by coincidence, and run it through a full-text search tool like CourtListener against the specific case. If the exact phrase does not surface in that opinion, treat the entire quotation as unverified until you can locate the actual source language, or conclude it does not exist as quoted.
Block quotes deserve more scrutiny, not less
There is a natural tendency to assume a long, detailed block quote is more likely to be accurate than a short inline quotation, on the theory that a fabrication would be harder to sustain at length. The documented cases do not support this assumption. Longer fabricated quotations appear in the record alongside short ones, and length alone provides no meaningful signal about accuracy.
Final takeaway
Treat every quotation, regardless of length or how confidently it reads, as unverified until you have confirmed the exact language against the actual opinion text. This single habit closes one of the more commonly underestimated gaps in AI-era citation verification.
A common question
Is this specifically an AI problem, or did fabricated quotations happen before generative AI tools existed?
Misquotation and misattribution predate AI drafting tools, arising occasionally from sloppy human research or over-aggressive paraphrasing presented as direct quotation. What has changed is scale and confidence: AI tools can generate large volumes of plausible-sounding fabricated quotations quickly, with a uniformly confident tone that gives no outward signal distinguishing a fabrication from an accurate quotation, whereas a rushed human researcher's error often carries some tell (an approximate rather than exact quote, an unusual paraphrase) that a careful reader might notice.
Related reading
- Existence, Quote, and Proposition Checks Explained
- How to Check for AI Hallucinations in Legal Briefs
This is precisely the failure mode the quote-match layer of a three-layer verification tool is built to catch automatically, running the same fragment-search logic described above across every quotation in a document at once.