Citation roulette: the verification habit that keeps you out of the AI-hallucination court database

Overview

This is a verification habit: how to check every quote, case, statistic and link a model hands you before it lands in something with your name on it.

Why now: as of June 2026, Damien Charlotin’s public AI Hallucination Cases database has logged around 1,600 court decisions worldwide in which a party was caught relying on hallucinated material — fabricated citations, fake quotes, misstated authority — and it is growing by five or six new cases a day. On one day alone, 31 March 2026, US courts noted suspected AI hallucinations in seventeen separate decisions.

By the end of this you’ll have a four-pass checking routine you can run on any AI-assisted deliverable in a few minutes.

The content

The obvious read is that hallucinated citations are a lawyer problem — a niche failure for people who quote case law for a living. The overturn: lawyers are simply the ones who get caught in public, because a court reporter writes the failure down. Your slide deck, board paper or client memo has no judge checking the footnotes. There is no reason to think the fabrication rate is any lower in your documents; only the audit is missing.

What makes fabricated sources dangerous is that they are plausible by construction. A model generates the citation that should exist given everything around it — the right-looking journal, a real author, a believable year. It is not lying so much as filling a shape. That is exactly why skim-reading fails: the fake reads better than a real source, because nothing about it is surprising. The single highest-profile example, Mata v. Avianca, drew only a 5,000-dollar sanction in 2023 — the money was nothing; the named-in-every-CLE reputational hit was the actual penalty.

So treat the model as a drafting tool and yourself as the verification layer. The habit is one sticky idea — trace-to-source — applied in four passes. Quotes: find the exact words in the original document, not a summary of it. Cases and studies: confirm the thing exists and says what you claimed, not merely that a similar-sounding title exists. Statistics: open the underlying table, because models drift figures and decimal places. Links: click every one; a dead or redirected URL is the cheapest tell there is.

Where this breaks is worth saying plainly. Trace-to-source catches fabrication, not framing — a real study can be cited accurately and still be cherry-picked or out of context. And web-connected models that retrieve live sources reduce invention but do not remove it; a retrieved snippet can still be mis-attributed to the wrong document. Verification is a floor, not a ceiling.

Try it

Run this on your own current AI-assisted deliverable — the one closest to going out — not a practice document.

You are my citation auditor, not my drafter. Below is a draft I produced with AI help.

List every checkable claim in a table: every direct quote, named case or
study, statistic, and URL.

For each row give me:
- the exact claim as written
- what an independent person would need to open to confirm it (primary
  source, not a summary)
- a CONFIDENCE flag: VERIFIABLE-IN-PRINCIPLE vs CANNOT-LOCATE-WITHOUT-ACCESS
- any internal tell (oddly round number, author/year mismatch, a quote
  too neat for the source)

Do NOT reassure me anything is correct — you cannot confirm it. Only tell
me where I must go to check it myself.

[paste draft]

The trap to avoid: the model cannot verify its own output. Asking “are these real?” invites it to confirm its own fabrications. This prompt makes it build your checklist, then you do the opening, clicking and reading. If a row is CANNOT-LOCATE, the default is to cut the claim — not to ship it and hope.

Additional reading

• AI Hallucination Cases Database — Damien Charlotin — the running, daily-updated scoreboard of court decisions worldwide where hallucinated material was caught.
• In One Day (Mar. 31), 17 U.S. Court Decisions Noting Suspected AI Hallucinations — Volokh Conspiracy / Reason — on why most cases go unrecorded.
• Mata v. Avianca, Inc. — Wikipedia — the original 5,000-dollar sanction that became the canonical example.
• AI Hallucination Cases — Simon Willison — an early read on the database and why lawyers keep returning to the tool anyway.

Editor’s note

Lawyers get caught because someone with subpoena power reads their footnotes; the rest of the population ships unaudited citations into decks and memos every week and never finds out. Do the verification pass. It’s easy, and it’s worth it.