How to Verify AI Legal Citations Before You File (ABA 512 Checklist)

To verify AI legal citations before you file, run every cite through three checks. First, confirm the case exists by pulling it from a primary source, not from the AI that produced it. Second, confirm the quoted language actually appears in that opinion. Third, confirm the case supports the proposition it is cited for.

ABA Formal Opinion 512 makes this verification a professional duty rather than an optional step, and the checklist below turns it into a Rule-by-Rule routine.

The four independent checks every citation passes before it reaches your draft.

The case that should change how you check cites

In early 2026, the Sixth Circuit did something that should make every litigator re-read their pre-filing routine. In Whiting v. City of Athens (Nos. 24-5918/5919, 25-5424, 2026 WL 710568), the court sanctioned two Tennessee lawyers $15,000 each in punitive sanctions, ordered full reimbursement of the appellees' fees, and tacked on double costs.

The brief contained more than two dozen problems. Some citations pointed to cases that did not exist. Others quoted language that was nowhere in the cited opinion. And some pointed to real cases that simply did not support the proposition they were cited for. The court called the conduct "well beyond sloppiness in drafting."

Worth noting, because it sharpens the duty: the Sixth Circuit did not expressly find the fakes came from generative AI. The rule it laid down is broader. No filing should contain citations, however generated, that the lawyer has not personally read and verified (National Law Review, March 2026). Whether a chatbot, a junior associate, or your own tired memory produced the cite, the signing lawyer owns it.

Read that list again, because it is the whole point of this guide. There were three distinct failure modes, and only one of them, the fabricated case, is the kind your gut already knows to look for.

If you want to know how to verify AI legal citations before you file, and this matters most for legal AI for in-house counsel teams running lean without a research librarian, you have to internalize that the easy failure (the fake case) is no longer where lawyers get caught. The dangerous ones are the real citation with the wrong quote, and the real citation that does not stand for what you claimed. Those survive the only check most people actually run.

This is a hands-on, printable checklist, mapped to ABA Formal Opinion 512 and to the three failure categories a federal appeals court just named out loud. For the full backstory on how we got here, the sanctions roundup from Mata v. Avianca onward lives in our AI hallucinations and sanctions piece. This post is the narrower thing: the task you run before the brief leaves your hands.

TL;DR

• A "does the case exist" check is no longer enough. The 2026 sanctions are landing on real cites with fabricated quotes and real cites that do not support the proposition.
• Whiting v. City of Athens (6th Cir. 2026) named three failure modes in one order: fake case, missing quote, and citation that does not support the claim. Build your checklist around all three.
• ABA Formal Opinion 512 (July 2024) ties this to Model Rule 1.1 (competence) and Rule 3.3 (candor). Independent verification is not optional, and a purpose-built legal tool does not transfer the duty.
• The Charlotin hallucination database now tracks well over a thousand cases globally, growing several a day. Reported US sanctions hit roughly $145,000 in Q1 2026 alone.
• The structural fix is not more willpower. It is researching in a tool where the source text sits beside the claim, so non-verification becomes hard rather than tempting.

Part of our legal AI verification and hallucination guide series.

For related verification / hallucination / vendor-trust coverage, see Inside a 4-Layer Legal AI Citation Verifier: A Technical Walkthrough and Damien Charlotin's Hallucination Tracker, Read Like a Risk Manager.

Why the old checklist fails in 2026

For two years, the standard advice was simple: pull every cite in Westlaw or Lexis and confirm it is real. That advice was built for the Mata v. Avianca era, when ChatGPT invented six cases out of thin air and a lawyer pasted them in. Those briefs are easy to catch. Run the cites, find the ones that return nothing, panic, fix.

The problem is that the threat moved. As models got better at sounding like law, they got better at producing citations that are half right. A real case name. A real reporter citation. A real court and year. And then a quoted holding the opinion never contained, or a pin cite to a page that says something adjacent but not what you needed.

Stanford's HAI research, which we cover in the sanctions pillar, found that even purpose-built legal research tools hallucinated somewhere between 17 and 33 percent of the time in 2024 testing. Those are not all phantom cases. A large share are exactly this subtler species: right shell, wrong substance. If you want the structural side of this, our note on legal AI that avoids hallucinating cases covers how grounded retrieval cuts the rate at the source.

So the existence check is necessary and badly insufficient. If your verification protocol stops at "the case is real," you will pass a Whiting-style brief straight through.

The Sixth Circuit told you as much by enumerating the categories. Treat that enumeration as a gift. It is a free taxonomy of how you can still get sanctioned even after you "checked."

The pre-filing checklist, mapped to ABA 512

ABA Formal Opinion 512 frames the duty cleanly. Under Model Rule 1.1, competence now includes understanding the benefits and risks of the generative AI you use. Under Rule 3.3, you owe the tribunal candor, which means you cannot put a representation in front of a court that you have not independently verified.

The opinion is blunt that you "should not rely on" generative outputs without that independent check. The checklist below ties each step to the rule it satisfies and the failure mode it catches.

Run it on every citation, every quote, every proposition. Not the ones that feel risky. Every one.

1. Existence: catches the fabricated case (Rule 1.1)

Open the actual opinion in a real source. Not a summary, not the AI's restatement, the opinion itself. Confirm the case name, reporter citation, court, and year all resolve to one real document. If you cannot open it, it does not exist for filing purposes. This is the floor, not the ceiling.

One rule above all the others: never verify a citation with the same AI that produced it. Ask a chatbot "is this case real?" and it will often double down on its own fabrication. Take the cite to an independent primary source. If you do not have Westlaw or Lexis, free sources are enough to confirm existence: the Legal Information Institute at Cornell, govinfo.gov for federal materials, and the Supreme Court Case Finder for U.S. Supreme Court opinions. A cite that returns nothing in any of these is a red flag you cannot file past.

2. Quote fidelity: catches the real case with a fabricated quote (Rule 3.3)

This is the step almost everyone skips, and the one Whiting punished. For every quoted passage, find that exact language inside the opinion and confirm the words are there, in that order, on the page you pin-cited.

A quote that is paraphrased, stitched together from two paragraphs, or simply invented is a misrepresentation to the court even if the case is genuine. Search the opinion text for a distinctive phrase from your quote. If it does not appear verbatim, the quote is wrong.

3. Proposition support: catches the real case cited for the wrong proposition (Rule 3.3)

Read the surrounding holding, not just the quoted line. Does the case actually stand for the point you cited it for? A case can contain your exact sentence in dicta, in a dissent, or in a recitation of the losing argument, and still not support your proposition.

This is the hardest failure to catch because it requires reading, not searching. Ask: if opposing counsel reads the full opinion, can they argue your cite cuts the other way? If yes, you have a problem.

4. Good law: catches overruled or limited authority (Rule 1.1)

Confirm the case is still good law. The clearest 2026 example of why this matters is Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), which overruled Chevron v. NRDC, 467 U.S. 837 (1984).

A model trained on decades of pre-2024 administrative law will happily cite Chevron deference as live doctrine. Run the citator. An AI that does not know the law changed will not warn you that it did.

5. Pin cite accuracy: catches misplaced page references (Rule 3.3)

Confirm the pin cite points to the page where the proposition actually appears. Generated pin cites are frequently plausible and wrong. A court that flips to your cited page and finds nothing relevant reads that as carelessness at best.

6. Statutory and regulatory text: catches stale or misquoted statutes (Rule 1.1)

If you cite a statute or regulation, pull the current text. Codes get amended. A model's training cutoff is not the current U.S. Code. Confirm the section number, the operative language, and the effective version.

This is the one place a statutes tool earns its keep: a public statutes and regulations API exists precisely to fetch current USC, CFR, and 50-state code text on demand, which is exactly the kind of authority you do not want to trust to a model's memory. (Note the scope: that API is statutes and legislation only. There is no public case-law search endpoint, so case verification still happens against a real case-law source like Westlaw or Lexis.)

Try Vaquill AI on your matterBook a demo

7. Confidentiality of the inputs: catches a different kind of exposure (Rule 1.6)

This one is not about the citation, it is about how you ran the check. Rule 1.6 covers client confidentiality. If your verification workflow means pasting privileged facts into a consumer chatbot that trains on your inputs, you have traded a citation risk for a confidentiality breach. Before you do, read whether ChatGPT is confidential for legal work and what its terms actually say about training on your prompts.

Know where your inputs go before you paste them. Our note on where your legal AI data actually goes walks through what to ask a vendor.

Run every cite top to bottom. A no at any gate stops the cite from reaching the brief.

What a half-right citation actually looks like

Here is the failure mode the existence check misses. Imagine an AI hands you this:

Smith v. Jones, 543 U.S. 112, 119 (2004): "a contractual limitation of liability is unenforceable where it shields a party from its own gross negligence."

The case name, reporter, volume, page, and year all look like a real U.S. Reports cite. Run step 1 and the volume resolves. But open volume 543 of U.S. Reports to page 112 and the case is about something else entirely, or the volume's pages do not run to a Smith v. Jones at all. The quoted sentence about gross negligence never appears in it. That is a real-shell, wrong-substance hallucination, and it sails through "is the case real" because the shell is plausible.

The fix is mechanical. Pull the opinion, use find-in-page for a distinctive seven-word string from the quote ("shields a party from its own gross"), and if it returns zero hits, the quote is invented no matter how real the citation looks. This is step 2, and it takes about thirty seconds per quote once you build the habit.

How to verify AI legal citations without losing your morning

A seven-step check on every cite sounds brutal, and run manually it is. The honest reality from talking to litigators is that the steps that get skipped under deadline pressure are exactly steps 2 and 3, the quote and the proposition.

Those are also the expensive ones to skip. So the question is not "do I have the discipline," it is "is my workflow built so the lazy path is also the safe path."

This is where the structure of the tool matters more than the model behind it. Most general chatbots, and even some legal products, hand you an answer with citations and ask you to go elsewhere to confirm them. That gap, between the claim and the source, is where verification quietly dies.

The better pattern is grounded research: the system retrieves real opinions first, generates the answer from those retrieved documents, and shows you the source text next to the claim so you can read the holding without leaving the page. We explain the mechanics in how AI legal research works with RAG, and you can see the accuracy testing behind it on our benchmarks page.

The reframe is the whole thesis. "Verify before you file" is a habit, and habits break under stress. "Research in a tool where the source sits beside the claim" is structural, and structure holds when discipline does not.

The goal is to make non-verification awkward rather than convenient.

To be clear about the limit, because ABA 512 is clear about it: a purpose-built tool does not transfer the duty. Stanford's numbers prove that even the named legal incumbents miss.

Grounded research lowers the rate and shortens the time to confirm a cite. It does not let you stop reading the opinion. Anyone selling you "verification-free" legal AI is selling you a future sanction.

A note on scale, so you take this seriously

It is tempting to treat the sanctions cases as cautionary outliers, the kind of thing that happens to other lawyers. The data says otherwise.

Damien Charlotin's hallucination tracker, widely cited including by Bloomberg Law, now logs well over a thousand cases globally as tracked, with the US share in the high hundreds and the count climbing several entries a day. Reported US sanctions reached roughly $145,000 in the first quarter of 2026 alone, including a reported Oregon matter near $110,000. This is not a rare event anymore. It is a recurring line item in court dockets.

What changed is not that lawyers got lazier. It is that the tools got persuasive enough to pass a casual look, and the casual look is what most verification still amounts to.

Whiting is the bellwether because an appeals court stopped treating "the case is fake" as the only sin and started counting the subtle ones. Other courts will follow that framing. Build your checklist for the court you will face in 2027, not the one Mata faced in 2023.

The shorter version you can tape to your monitor

If you remember nothing else, remember the three questions that map to the three Whiting failure modes:

1. Does this case exist, in a real source I opened myself?
2. Is this quote actually in that opinion, word for word, on this page?
3. Does this case actually hold what I said it holds?

Then add the three that catch the rest: is it still good law, is the statute text current, and did I run this check somewhere that protects my client's confidences. Six questions. Every cite. No exceptions under deadline, because the deadline is precisely when the subtle errors slip through.

The lawyers in Whiting almost certainly believed they had checked. They had not checked the right things. The difference between a clean filing and a $15,000 sanction in 2026 is not whether you verified. It is whether your verification asked question two and question three at all.

FAQ

How do I verify an AI legal citation before filing?

Run three core checks on every cite. Confirm the case exists by opening the opinion in a primary source you trust, not in the AI that produced it. Confirm the quoted language appears verbatim on the pin-cited page. Confirm the holding actually supports the proposition you cited it for. Then add a citator check for negative treatment. A cite that fails any of these does not go in the brief.

How can I check if an AI-generated case citation is real for free?

You do not need a paid database to confirm existence. The Legal Information Institute at Cornell, govinfo.gov, and the Supreme Court Case Finder all let you pull federal and many state opinions at no cost. Search the case name and reporter cite. If it returns nothing across these sources, treat the citation as fabricated.

Can ChatGPT verify its own citations?

No. The tool that invented a citation will often defend it when asked "is this real," because it is predicting plausible text, not checking a database. Always verify against an independent primary source. Using the same model to confirm its own output is the most common way fabricated cites survive to filing.

What is ABA Formal Opinion 512?

ABA Formal Opinion 512 (July 2024) is the American Bar Association's guidance on generative AI in legal practice. It ties AI use to existing duties: competence under Model Rule 1.1, candor to the tribunal under Rule 3.3, and confidentiality under Rule 1.6. The core takeaway is that lawyers must independently verify AI outputs and cannot delegate that responsibility to the tool.

Does a purpose-built legal AI tool remove the need to verify citations?

No. Stanford HAI research found that even legal-specific research tools produced incorrect or hallucinated answers in a meaningful share of queries, roughly 17 to 33 percent in 2024 testing. A grounded tool lowers the error rate and shortens verification, but the duty to read the opinion stays with you. ABA 512 is explicit that the tool does not transfer the duty.

How do I check whether a case is still good law?

Run a citator. Westlaw's KeyCite and LexisNexis Shepard's flag negative treatment, including reversals and overrulings. The clearest recent reason this matters is Loper Bright Enterprises v. Raimondo, 603 U.S. 369 (2024), which overruled Chevron. A model trained on older law will cite the dead doctrine as live, so a citator pass is non-negotiable before filing.

What happens if I file a brief with a fake AI citation?

Courts are imposing real sanctions. In Whiting v. City of Athens (6th Cir. 2026), two lawyers were fined $15,000 each plus the opponents' fees and double costs. Damien Charlotin's hallucination tracker logs well over a thousand cases globally, and reported US sanctions reached roughly $145,000 in the first quarter of 2026. See our AI hallucinations and sanctions roundup and the Oregon $110K sanction breakdown for the pattern.

Are ABA Formal Opinions binding on lawyers?

ABA Formal Opinions are persuasive, not binding, but state bars and courts routinely adopt their reasoning. We break down exactly how much weight Opinion 512 carries in are ABA Formal Opinions binding.

For more on a research workflow that grounds answers in real opinions and surfaces sources you can verify, see /features/legal-research.

Arshita Anand

Co-Founder & CEO · Attorney

Arshita leads product and strategy at Vaquill, building the legal AI suite that solo, small-firm, and in-house US lawyers use to run a matter end to end.