Aktiver — semantic AI platform Deterministic · traceable · behind your firewall
Insights / Enterprise AI

A federal appeals court sanctions lawyers over fake citations

On March 13, the U.S. Court of Appeals for the Sixth Circuit sanctioned two Tennessee lawyers, Van Irion and Russ Egli, for appellate briefs that cited non-existent cases, misquoted others and misrepresented the record. In a published opinion in Whiting v. City of Athens, written by Judge John K. Bush and joined by Judges Jane Stranch and Eric Murphy, the court ordered each lawyer to pay $15,000 to the court as a punitive sanction, to pay the other side’s full reasonable attorney fees on appeal across three consolidated appeals, and to pay double costs. It also referred both to the chief judge to consider disciplinary proceedings.

“All told, we found over two dozen fake citations and misrepresentations of fact in Whiting’s briefs,” the court wrote, adding in a footnote that this was a conservative estimate that excluded typos and sloppy citations.

/ 01What happened

The appeals arose from lawsuits over a 2022 fireworks show in Athens, Tennessee. When the court reviewed the briefs, it found that many cited cases did not exist, did not contain the language quoted, or did not discuss the point they were cited for. One example listed in the opinion’s appendix is a Sixth Circuit decision called “Berg v. Knox Cnty.” that, the court found, does not exist; the only source on Westlaw citing it was the appellant’s own briefing.

The court issued an order to show cause asking the lawyers, among other things, to provide copies of every authority they cited, to say who wrote the briefs, to say whether they used generative AI to write them, and to explain how they checked their citations. According to the opinion, they did not respond to those directives and instead argued that the order was void and motivated by harassment. The court rejected those arguments.

The court did not make a finding that AI produced the fake citations. It did not need to. Quoting a California appellate decision, it said no filing should contain citations, “whether provided by generative AI or any other source,” that a lawyer has not personally read and verified. And it quoted another court’s view that the existence of some genuine authority does not offset the harm of “false and hallucinated cases.” As the Volokh Conspiracy’s write-up highlights, the opinion states plainly: “Citing even a single fake case can be sanctionable.”

/ 02Why it matters beyond law

Courts have been sanctioning fabricated citations for some time. What stands out here is a federal appeals court explaining that it chose $15,000 each partly because “smaller fines have plainly been inadequate,” citing the continuing stream of such cases, and that the fake citations forced it to verify every single citation itself. It also sets out a simple rule: the professional who signs the document is responsible for every authority in it, however it was produced.

That rule generalises. Every regulated profession now has people using AI tools to draft work product: audit memos, clinical summaries, intelligence assessments, engineering reports. The legal system happens to have a quick, public way of catching a fabricated reference — the other side and the judge look it up. Most other fields do not. A made-up regulation in a compliance memo or a non-existent study in a clinical summary may sit undiscovered until it matters.

The failure mode is also specific. The opinion quotes the observation that hallucinations are more likely when a user asks a generative tool to supply a citation for a proposition that no real authority supports. In other words, the tool produces the reference the user wanted rather than reporting that none exists. That is the opposite of what a research tool should do.

/ 03Citations that cannot be invented

The structural fix is to make it impossible for a system to cite something that is not in its sources. If the documents an AI works from are first turned into a knowledge graph, with every statement linked to the passage that supports it, then any answer the system gives is assembled from those statements and carries their links. A claim with no supporting node simply cannot be produced, and the honest answer “no authority found” becomes a normal result rather than a failure.

That does not remove the professional’s duty to read what is cited. It does change the economics of checking. Verifying a citation that points to a specific passage in a document you hold takes seconds. Discovering that a citation points to nothing takes much longer, and in this case cost two lawyers $15,000 each plus fees and costs.

A system that can only cite what is in its sources cannot hand you a case that does not exist.

/ 04What to do

  • Set a verification rule. Every cited authority in AI-assisted work product is opened and read by a person before it leaves the building. The Sixth Circuit has now said this in plain terms.
  • Prefer grounded tools. For research, use systems that retrieve from a defined corpus and link every claim to a source, and that can say when nothing supports a proposition.
  • Keep records. Be able to say, if asked, which tools were used and how output was checked. The court asked exactly that.
Aktiver’s take — Enterprise data

The court’s rule is plain: whoever signs is answerable for every authority cited. Tools that will supply a case for any proposition make that rule expensive to keep.

Represent a brief’s support structure explicitly. Each proposition links to an authority, each authority to a document the firm holds, each quotation to a pinpoint passage. A citation with no target node cannot be generated, so “no supporting authority” becomes an ordinary query result. The same pattern serves a compliance memo citing a regulation or an audit file citing a control test. Ingestion Hub turns a research corpus, whether PDFs or exported databases, into an ontology that practitioners refine, and deterministic reasoning means a partner reviewing a draft sees the same evidence chain the associate saw. Verification shifts from re-searching every cite to opening the linked passage, which takes seconds and leaves a record of who checked what.

How we’d model it

Legal propositionAuthoritySource documentQuotationPassageCitationSigning attorneysupportedBycontainedInpinpointedToverifiedByLegalpropositionAuthoritySourcedocumentQuotationPassageCitationSigningattorneysupportedBycontainedInpinpointedToverifiedBy
  • Legal proposition —supportedBy→ Authority
  • Authority —containedIn→ Source document
  • Quotation —pinpointedTo→ Passage
  • Citation —verifiedBy→ Signing attorney

See where this applies →

Sources — read the originals
  1. Whiting v. City of Athens, Tennessee, Nos. 24-5918/5919/25-5424 (6th Cir. Mar. 13, 2026) — U.S. Court of Appeals for the Sixth Circuit opn.ca6.uscourts.gov ↗
  2. Lawyers Citing Nonexistent Cases Ordered to Pay Opponents' Attorney Fees, Double Costs, $15K Fine Each — The Volokh Conspiracy (Reason) reason.com ↗
Go deeper

See how Aktiver turns raw data into knowledge graphs your AI can reason over — deterministic, traceable, behind your firewall.

← All insights Subscribe via RSS → Book a demo →
Keep reading