legal studio

The fabricated ones always look right

In 2023 a lawyer filed a brief in the Southern District of New York citing six cases that did not exist. ChatGPT had invented them — complete with reporter citations, internal quotes, and plausible judicial reasoning — and when he asked the model whether the cases were real, it said yes. Mata v. Avianca cost him $5,000 and made him the most famous cautionary tale in the profession.

It did not stay rare. Damien Charlotin's AI Hallucination Cases database — a worldwide tracker of court decisions involving AI-fabricated material — stood at 1,598 cases as of June 9, 2026, and was growing at roughly eight new cases a day. The penalties escalated with the count. Couvrette v. Wisnovsky (D. Or.): 15 fake citations and 8 fabricated quotations across three briefs, $110,204.38 in total sanctions and fee-shifting, briefs struck, claims dismissed with prejudice. Withers v. City of Aberdeen (N.D. Miss., June 2026): both sides filed hallucinated material, and the judge removed all four attorneys from the case and barred them from appearing before that district for two years.

Here's what every one of those lawyers had in common: the fake authorities looked right. Fabricated cases come with clean Bluebook cites and quotes that sound exactly like the proposition you need. Looking right is what fabrication is. The only defense is mechanical — every authority gets opened at the source before it goes in the draft — and this chapter turns that defense into code you run, not a resolution you make.

Why it matters

• The verification gap is the profession's biggest skills gap, measured. Per a Bloomberg Law study, 76% of attorneys expect new lawyers to cite-check AI-generated materials — while only 20% of 3Ls report generative-AI proficiency. CLE programs increasingly warn about sanctions and even describe verification workflows; what stays rare anywhere on the market is repeated, graded practice of the procedure itself. That practice is this chapter's first lesson.
• Adoption is accelerating whether your firm has a policy or not. Thomson Reuters found 26% of legal organizations actively using generative AI in 2025, nearly double the year before, with 78% of law-firm respondents expecting it to be central within five years. Clio puts some level of AI use at 87% of large firms. The tools are already in the building; the discipline usually isn't.
• Accuracy fear is the blocker, and the fix is a workflow, not a better model. Thomson Reuters found 40% of reluctant legal professionals cite accuracy and reliability concerns — nearly double any other concern. You don't resolve that fear by trusting harder. You resolve it with a gate the draft cannot get past unverified.
• Privilege waiver by AI disclosure is now court-recognized risk, not a hypothetical. Three early-2026 federal decisions on overlapping facts split on it; in United States v. Heppner, disclosure to a third-party AI platform was held to destroy the confidentiality that work-product protection presupposes. Which matter content may enter which tool is a decision to make once, in writing — not per-prompt at 11pm before a filing deadline.

What AI specifically gets wrong

• It fabricates authority fluently, and confidence is not a signal. A model asked "are these cases real?" will cheerfully confirm its own inventions — that's the exact exchange from Mata. Verification happens in a citator, never in the chat window.
• It quotes real cases saying things they never said. A real case with a fabricated quote is still a fabrication — Couvrette's 8 invented quotations rode alongside citations to cases that exist. Existence, holding, and treatment are three separate checks; passing one proves nothing about the other two.
• It extracts clauses without telling you where they are. A contract-review summary that says "liability is capped at 24 months of fees" without a section reference isn't review support — it's a rumor about your own document. No pinpoint, no claim.
• It doesn't know what privilege is worth. The model will process a litigation strategy memo as happily as a lunch order. Whether that memo may leave the building — and through which tool, under what terms — is a question the model cannot answer and you must never leave to habit.

What you'll be able to do at the end

• Run every AI-cited authority through a three-verdict gate — exists, says what the draft claims, still good law — and produce a per-citation KEEP / FIX / STRIKE record with a file / do-not-file decision at the bottom.
• Hold every extracted clause to the pinpoint standard, triage it against a counsel-approved playbook (preferred, fallback, walk-away), and route anything outside the playbook to a named human — because negotiation positions are set by counsel, not assembled by a tool.
• Encode your team's privilege rules as a routing table — matter sensitivity crossed with tool clearance — so the answer to "can I paste this?" is a lookup, not a judgment call under deadline.
• Build the review log that records what was AI-assisted, what was verified, by whom, and who signed — the record you'd want to hand up when a judge with a standing order on AI use asks how this brief was prepared.

Every drill runs in plain Python on invented fixtures — no client data, no live tools, no research-platform login. The logic is the point: it's the same gate whether it lives in this script, your matter checklist, or the review workflow your firm adopts next quarter.