# Verify AI Case Citations Before You File

> A step-by-step way to confirm every AI-drafted citation exists, is quoted accurately, and is good law before you sign a filing. See the workflow here.

[Home](https://rankshieldlegal.com/) / [Blog](https://rankshieldlegal.com/blog/) / Citation Integrity Citation Integrity
# How to Verify AI-Generated Case Citations Before You File
To verify AI-generated case citations before you file, run three checks on every authority: confirm the case actually exists in a real reporter, confirm any quotation matches the published opinion word for word, and confirm the case is still good law. A citation only passes when all three hold. Skip any one of them and you can file a case that is fake, misquoted, or overruled, the three failure modes courts are now sanctioning.

By [Jamie Kloncz](https://rankshieldlegal.com/about/), Founder, RankShield ** 18 min read ** Published July 1, 2026

Generative AI does not know when it is wrong. It produces citations that look flawless, correct reporter, plausible court, clean Bluebook format, for cases that never existed. In a preregistered Stanford study of the leading legal research tools, Lexis+ AI answered incorrectly more than 17% of the time and Westlaw's AI-Assisted Research hallucinated more than 34% of the time, "one in six or more" queries, even on tools built specifically for law. [[1]](#ref-1) The risk is not that AI drafts your brief. The risk is signing a filing that no human verified. This guide is the verification workflow that goes between the draft and your signature.
The workflow below is deliberately mechanical. It does not ask you to judge whether a passage "sounds like" a real holding, because that judgment is exactly the instinct AI is engineered to defeat. Instead it converts a soft question, can I trust this citation, into three hard questions with binary answers: does the case exist, does the quotation match, and is the case still good law. Each question has an authoritative source that answers it, and each answer is something you can show to someone else. That is the difference between believing your citations are sound and being able to demonstrate it.
This article is informational and is not legal advice. It describes a general verification process and the professional-responsibility landscape as reported in public sources. It does not establish an attorney-client relationship, and it does not tell you how any specific rule applies to your matter, your jurisdiction, or your filing. RankShield Legal is a security and verification vendor, not a law firm, and nothing here substitutes for your own professional judgment.

## Why the signing attorney owns every AI citation
The duty to verify is non-delegable, and "the AI wrote it" is not a defense. Courts have been explicit: the lawyer who signs the filing is responsible for every authority in it, whether it came from an associate, a contract attorney, or a chatbot. The foundational case, *Mata v. Avianca*, ended in a Rule 11 sanction after ChatGPT invented six nonexistent decisions and the attorneys filed them without checking. [[1]](#ref-1)
Since then the pattern has industrialized. Damien Charlotin's public database tracking court decisions that flag [suspected AI hallucinations](https://rankshieldlegal.com/ai-hallucination-legal-filings/) now logs more than 1,300 proceedings and grows by several every day. [[1]](#ref-1) Sanctions have climbed from the $5,000 in *Mata* to six-figure orders. Verification is no longer diligence theater, it is the difference between a filed brief and a show-cause order with your name on it.
It helps to be precise about where the exposure actually sits. The risk is not that a lawyer used an AI tool. Many courts and clients now expect efficient drafting, and the tools are not going away. The risk is that a document carrying an attorney's signature made representations to a tribunal that turned out to be false, and that no one confirmed them before filing. A signature is a certification. When you sign, you are telling the court that you have read the filing and that its legal contentions are warranted. An unverified AI citation converts that certification into a gamble, and the house, in this case the court, always gets to check the cards.

## The three failure modes, at a glance
Before working through the checks one at a time, it helps to see why three separate checks are necessary rather than one. Each check catches a distinct failure mode, and a citation can pass one while failing another. A perfectly formatted citation to a real case can still carry an invented quotation. A real, accurately quoted case can still have been overruled last year. The checks do not overlap, so skipping any one of them leaves a live path to a sanctionable filing.
Failure mode What the citation looks like The check that catches it
Fabrication A real-looking citation to a case that does not exist in any reporter Existence: resolve the exact citation against an authoritative case-law source
Misquotation A real case cited around language it never contained, or a wrong pinpoint page Quotation accuracy: match the quoted text to the published opinion, verbatim
Bad law A real, accurately quoted case that has been overruled, reversed, or superseded Good-law standing: run the authority through a citator and read the negative-treatment flags

A citation passes only when all three checks hold. Failing any single check is disqualifying, regardless of how convincing the other two look.

Source: Stanford RegLab; Charlotin AI Hallucination Cases Download SVG

## Confirm the case actually exists
The first and most important check is existence: does the cited case appear in a real reporter, from the court and year claimed? Fabricated citations are the single largest AI failure mode, and they are invisible to the eye because the model copies real formatting, a realistic volume number, a plausible reporter, a court that hears that kind of case. [[1]](#ref-1) Formatting tells you nothing about whether the case is real.
Resolve every citation against an authoritative source, a case-law database such as Westlaw, Lexis, or CourtListener, and match the exact citation, not just the case name. A near-miss (right parties, wrong reporter or year) is a red flag, not a rounding error. If the citation does not resolve to a real opinion, treat it as fabricated and remove it before the filing moves any further.
The trap here is confirmation by resemblance. When you search a case name and see a real case with similar parties, the mind wants to close the loop and move on. That instinct is precisely what fabricated citations exploit. AI models frequently attach a real case name to an invented reporter citation, or pair a real citation with a subtly altered party name, so that a name-only search returns something that feels like a match. The discipline is to verify the specific string your brief relies on, volume, reporter, page, court, and year, and to confirm that the opinion those coordinates point to is the one you intend to cite. If any coordinate is off, you have not found your case; you have found a different case that happens to look similar, and relying on it is its own error.

## Verify the quotation matches the opinion
A real case can still be cited dishonestly. The second check is quotation accuracy: pull the actual opinion and confirm that any language your brief presents as a quote appears in the opinion, verbatim, and that the pinpoint page is correct. AI models routinely paraphrase a holding and present it inside quotation marks, or attach a real citation to a proposition the case never stated.
This check matters because opposing counsel and the court will do it for you. A quotation that does not appear in the cited opinion reads as either carelessness or misrepresentation, and under candor-to-the-tribunal obligations, neither is survivable. Match the quote to the published text before you rely on it, a real citation wrapped around an invented quote is still a fabrication.
There is a quieter version of this failure that deserves the same scrutiny: [the accurate quotation used to support an inaccurate proposition](https://rankshieldlegal.com/blog/fabricated-vs-misgrounded-legal-ai-citations/). A model may lift a genuine sentence from a real opinion and then place it under a heading or in a paragraph that stretches it well past what the court actually held. The words are real; the argument they are made to support is not. So the quotation check has two parts. First, confirm the words appear where the citation says they do. Second, read enough of the surrounding opinion to confirm the court was saying what your brief claims it was saying. A quotation ripped from its reasoning can mislead a court as effectively as one that was never written at all.

## Check the authority is still good law
The third check is the one most verification workflows skip: is the case still good law? A citation can exist and be quoted perfectly and still be worthless because the holding was overruled, reversed, superseded by statute, or limited by later decisions. Existence and accuracy do not tell you a case still controls.
Good-law standing comes from a citator, KeyCite in Westlaw or Shepard's in Lexis, that tracks negative treatment. Run every load-bearing authority through one and read the negative-treatment flags, not just the headline. A brief built on overruled law fails on the merits even when every citation is technically real, which is why a complete pre-filing check verifies all three dimensions, never just the first.
The reason this check gets skipped is that it feels like the least likely place to find a problem. A case you pulled and quoted correctly has already cleared two hurdles, and the temptation is to assume a real, accurately quoted opinion must still be authoritative. But a citator flag does not report on whether the case is real; it reports on whether later courts have undercut it. Those are independent facts. An opinion can be genuine, quoted flawlessly, and no longer good law because a higher court reversed it or a statute displaced it after it was decided. Reading the flag, and reading it carefully enough to distinguish "overruled on this point" from "questioned on an unrelated point," is what separates a citation that survives a merits fight from one that collapses the moment opposing counsel runs the same search.

## A repeatable pre-filing workflow
The three checks work best as a fixed sequence you run the same way every time, rather than a set of instincts you apply when something feels off. A fixed sequence removes the judgment call about which citations "look risky," because the honest answer is that fabricated citations are engineered to look safe. Run the same steps on every authority in the filing, load-bearing or not.

- **Extract every citation** Pull each cited authority out of the draft into a checklist, including string cites and footnotes. A citation that never makes it onto the list never gets verified, and the ones buried in footnotes are the ones most likely to slip through.
- **Resolve the exact citation** Match the full citation string against an authoritative case-law source. Confirm the reporter, volume, page, court, and year all point to the same real opinion. Treat any near-miss as a failure, not a typo to wave through.
- **Match every quotation** For each passage in quotation marks, open the opinion and confirm the language appears verbatim at the pinpoint page cited, and that the surrounding reasoning supports the proposition your brief attaches to it.
- **Run the citator** Send every authority you are relying on through KeyCite or Shepard's and read the negative-treatment flags. Confirm nothing load-bearing has been overruled, reversed, or superseded.
- **Record what you checked** Note which citations you verified, against what source, and when. A manual pass with no record protects the brief but proves nothing later. The record is what you hand a judge, a client, or a carrier if the work is ever questioned.
The order matters. Existence comes first because there is no point matching a quotation in an opinion that does not exist, and good-law standing comes last because it only matters for authorities that are real and accurately quoted.

## Manual verification versus automated certification
Doing all three checks by hand, on every citation, on every filing, does not scale, and it leaves no record that you did it. A manual pass protects one brief; it does not produce evidence you can hand a judge, a client, or your malpractice carrier. As courts add standing orders requiring an AI certification on filings, the gap between "we checked" and "here is proof we checked" becomes the exposure.
This is the problem RankShield Legal was built to close. Rather than flagging suspect citations for a human to chase, it [resolves each cited authority against live case-law](https://rankshieldlegal.com/ai-legal-citation-checker/) for all three checks, existence, quotation accuracy, and good-law standing, and issues a verifiable certificate before the filing is signed, sealed to a tamper-evident transparency log anyone can independently check. To be clear about what that means: RankShield does not promise an AI that never hallucinates, and its check is only as complete as the case-law corpus it can reach, which it records with every result. It certifies which of your citations are real, accurately quoted, and good law, so a fabricated case never reaches the court, and you can prove it.
The distinction worth holding onto is between catching an error and being able to show your work. A careful associate can catch a fabricated citation. What a manual process cannot easily produce is a durable, independently checkable record that the verification happened, covered every authority, and returned a specific result for each one. That record is what changes as courts move from expecting diligence to expecting proof of diligence. When a standing order asks you to certify that your citations were verified, the difference between an attestation you assert and one that is sealed to a log a third party can inspect is the difference between your word and evidence.

## What "hallucination-free" actually means
"Hallucination-free" is not a property any current legal AI tool has demonstrated. The Stanford benchmark measured error rates well above zero on tools built for law. [[1]](#ref-1) Treat verification as something your process does, not something your software guarantees.
It is worth being blunt about a phrase that shows up in marketing for legal AI tools: "hallucination-free." No responsible reading of the current evidence supports it as a claim. The Stanford study that measured the leading tools found error rates that are the opposite of zero, more than 17% for one tool and more than 34% for another, even though both were built specifically for legal research. [[1]](#ref-1) A tool that answers incorrectly one time in six is not hallucination-free, and treating any tool as if it were is how unverified citations reach filings in the first place.
The safer mental model is that verification is a property of your process, not a feature of your software. No model, and no verifier, removes your obligation to confirm what you file. What a verification layer can do is narrow the surface you have to check by hand, apply the same three checks to every citation without fatigue, and produce a record of the result. That is genuinely useful, and it is a different claim from promising that errors never happen. RankShield is a security vendor, and the honest version of its value is not that it makes hallucinations impossible but that it makes them detectable and the detection provable, within the limits of the case-law corpus it can reach, which it discloses with each result.

## What a verifiable certificate proves, and what it does not
Because RankShield Legal [issues a certificate and seals it to a transparency log](https://rankshieldlegal.com/citation-certification/), it is important to be exact about the claim that certificate makes. It attests to the architecture of the check and to the result it produced: that the three checks ran, against a recorded corpus, at a recorded time, and returned a specific verdict for each citation. Sealing that result to a tamper-evident log means a third party can confirm the record has not been altered after the fact. That is a claim about process integrity and evidence, not about the underlying law.
The certificate does not, and cannot, decide your case. It does not tell you that a real, accurately quoted, good-law authority supports your argument, that judgment remains yours. It does not extend past the case-law corpus the check could reach, which is why that scope is recorded with every result rather than assumed. And it is not a privilege determination. Where RankShield's broader platform describes privilege handling, what the architecture attests to is isolation and the consent under which processing occurred, not a legal conclusion that privilege was preserved in your matter. Keeping those boundaries visible is part of the honesty the tool is built around: it proves what it checked, and it stays quiet about what it did not.
On the security posture underneath all of this, RankShield describes its cryptography as quantum-safe rather than quantum-proof. The distinction is deliberate. Quantum-safe means the design uses algorithms chosen to resist the attacks a future quantum computer is expected to enable, so that records sealed today remain verifiable later. It does not mean any system is immune to every future advance, and no honest vendor would claim otherwise. The point of the transparency log is precisely that its integrity can be checked over time rather than taken on faith.

## Building citation checks into your practice
The workflow only protects you if it runs every time, which means it has to be part of how filings get produced rather than a step someone remembers under deadline pressure. The most reliable way to guarantee that is to make verification a gate the filing cannot pass without clearing, the same way a signature block or a certificate of service is not optional. When the check is a required step rather than a good intention, the failure modes stop depending on whether anyone felt suspicious that day.

- Treat every AI-assisted citation as unverified until it has cleared all three checks, existence, quotation accuracy, and good-law standing, and apply the same standard to citations you believe you recognize.
- Run the checks in order and on every authority, not only the ones that look load-bearing, because fabricated citations are designed to look ordinary.
- Keep a record of what was verified, against what source, and when, so that "we checked" can become "here is the proof," which is the standard courts are moving toward.
- Do not rely on formatting, plausibility, or a tool's marketing language as a substitute for resolution against an authoritative source.
- Remember that the signing attorney owns the result. Software can narrow the work and document it, but it does not transfer the professional responsibility that attaches to your signature.
This section describes a general approach to process design and is not legal advice. Your jurisdiction, your court's standing orders, and your professional-responsibility rules govern what you must actually do before filing.

Test yourself
## Citation verification self-check
A few questions on confirming AI-drafted citations before you sign a filing.

- 1 What are the three checks every AI-generated citation must pass? Format, length, and tone Existence, quotation accuracy, and good-law standing Author, court, and page count **Answer:** Existence, quotation accuracy, and good-law standing Confirm the case exists in a real reporter, that any quotation matches the opinion verbatim, and that the case is still good law. A citation passes only when all three hold.
- 2 Does correct Bluebook formatting tell you a citation is real? Yes, format confirms the case exists No, models reproduce format perfectly for cases that never existed Only for federal cases **Answer:** No, models reproduce format perfectly for cases that never existed Fabricated citations copy realistic reporters, courts, and years, so format carries no information about whether the case is real. Only resolving the exact citation does.
- 3 Are AI legal research tools "hallucination-free"? Yes, tools built for law never hallucinate No, a Stanford study measured error rates well above zero on tools built for law Yes, once they are certified **Answer:** No, a Stanford study measured error rates well above zero on tools built for law The Stanford study found Lexis+ AI incorrect more than 17% of the time and Westlaw hallucinating more than 34% of the time. Verification is a property of your process, not a software guarantee.
- 4 Who is responsible for an AI-generated citation that turns out to be fake? The AI vendor The signing attorney, because the duty to verify is non-delegable No one, if the tool made the error **Answer:** The signing attorney, because the duty to verify is non-delegable Courts have been explicit that the lawyer who signs the filing owns every authority in it. Mata v. Avianca ended in a Rule 11 sanction after unchecked ChatGPT citations were filed.
Honest self-check. There is no sign-up, and nothing is stored.

Questions answered
## Straight answers to the common questions
The questions readers ask about this topic, answered directly. **No forms, no sales pitch.**

JAMIE KLONCZ · SEO AGENCY NAPLES ************** ONLINE
Pick a question on the left, or search above. You will get the direct answer, the way an answer engine would give it.

← PREV NEXT → [REQUEST ACCESS →](https://rankshieldlegal.com/contact/)

- **How do you verify AI-generated legal citations?** Run three checks on every citation. First, confirm the case exists by resolving the exact citation against an authoritative case-law database, matching the citation, not just the case name. Second, confirm quotation accuracy by pulling the opinion and matching any quoted language and pinpoint page verbatim. Third, confirm the case is still good law using a citator such as KeyCite or Shepard's to catch overruled, reversed, or superseded authority. A citation passes only when all three hold; failing any one means removing or replacing it before filing.
- **What happens if you file a brief with a fake citation?** You can be sanctioned under Rule 11 and equivalent state rules, referred for discipline, and named publicly in the opinion. Consequences have escalated sharply, from the $5,000 sanction in *Mata v. Avianca* to six-figure orders in 2026, and a database now tracks more than 1,300 court decisions flagging suspected AI hallucinations. "The AI generated it" is not a defense, because the duty to verify cited authority is non-delegable and attaches to the attorney who signs.
- **Can correct formatting tell me a citation is real?** No. Fabricated citations copy correct Bluebook format, realistic reporters, and plausible courts and years, which is exactly why they are dangerous. Format is a presentation choice the model reproduces perfectly; it carries no information about whether the underlying case exists. The only reliable existence signal is resolving the exact citation against an authoritative source and confirming it returns the real opinion. A citation that looks perfect and a citation that is real are not the same thing, and only verification tells them apart.
- **Are AI legal research tools "hallucination-free"?** No current tool has demonstrated that. In a preregistered Stanford study of leading legal research tools, Lexis+ AI answered incorrectly more than 17% of the time and Westlaw's AI-Assisted Research hallucinated more than 34% of the time, roughly one in six queries or more, even though both were built for law. "Hallucination-free" is a marketing phrase, not a measured property. Treat verification as something your process does on every citation, not a guarantee your software provides, and confirm each authority against an authoritative source regardless of which tool produced it.
- **Why check whether a case is still good law if the citation is real?** Because existence and accuracy do not tell you the holding still controls. A genuine opinion you have quoted correctly can still have been overruled, reversed, superseded by statute, or narrowed by later decisions, and a brief built on it fails on the merits even though every citation is technically real. A citator such as KeyCite or Shepard's reports negative treatment that a case-law search alone does not surface. Run every load-bearing authority through one and read the flags carefully enough to tell a fatal reversal from an unrelated critique.
- **Does an automated verification certificate replace my professional judgment?** No. A verifiable certificate attests to the process and its result, that the three checks ran against a recorded corpus at a recorded time and returned a specific verdict for each citation, sealed to a tamper-evident log a third party can inspect. That is evidence the verification happened; it is not a decision about whether an authority actually supports your argument. The signing attorney still owns that judgment. RankShield Legal is a security and verification vendor, not a law firm, and the certificate documents diligence rather than substituting for it.

## References

- Stanford RegLab (Magesh, Surani, Dahl, Suzgun, Manning, Ho). Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools. Journal of Empirical Legal Studies, 2025 (preprint May 2024). https://hai. stanford. edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries Charlotin, D. AI Hallucination Cases database. 2026 (1,300+ proceedings; updated daily). https://www. damiencharlotin. com/hallucinations/ Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S. D. N. Y. June 22, 2023). https://law. justia. com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/.

Written by
## [Jamie Kloncz](https://rankshieldlegal.com/about/)
Founder, RankShield
Jamie Kloncz is the founder of RankShield, the verifiable AI and quantum security platform behind RankShield Legal. An engineer by training, he built RankShield after his own devices and business were attacked, including an AI voice-cloning scam that targeted his family, on one conviction: unverifiable security is the real danger, so every consequential action should leave a receipt anyone can independently check.
[More about Jamie →](https://rankshieldlegal.com/about/)

Try it · Free
## Check a citation against live case-law
Paste a citation from an AI-drafted brief and see whether the case actually exists, resolved against live case-law. Free, no sign-up. Then request early access to certify a full filing.
[Try the citation checker](https://rankshieldlegal.com/ai-legal-citation-checker/)

Keep reading
## Related guides
[Citation Integrity Fabricated vs. Misgrounded: The Two Ways Legal AI Gets Citations Wrong Read guide →](https://rankshieldlegal.com/blog/fabricated-vs-misgrounded-legal-ai-citations/)[Citation Integrity What Happens When You File a Brief With a Fake AI Citation Read guide →](https://rankshieldlegal.com/blog/fake-ai-citation-sanctions/)[Citation Integrity Legal AI Citation Checkers Compared: What Each Method Actually Catches Read guide →](https://rankshieldlegal.com/blog/legal-ai-citation-checkers-compared/)