The single most damaging way AI fails a lawyer is also the easiest to catch: it invents authorities that look completely real. Here is why it happens, what it has cost, and the workflow that stops it.
Educational guide · written for in-house counsel · current as of May 2026 · not legal advice
Why AI invents cases
A large language model does not look anything up unless it has been connected to a real database and instructed to. Left on its own, it generates text by predicting what is statistically likely to come next. A legal citation — a case name, a reporter, a volume and page, a year — is one of the most patterned strings of text in the English language. So when a model needs a citation, it produces something that matches the pattern: a plausible case name, a plausible reporter, a plausible pin cite.
That output is not a malfunction. It is the model doing exactly what it does. The problem is that a citation which merely looks right is worthless — and a model has no built-in sense of the difference. It will hand you a fabricated case with the same calm confidence it uses for a real one.
The mental model
Treat a general-purpose AI as a fluent, fast, and well-read colleague who never says "I don't know" and never checks. Useful for a first draft. Never a source.
What fabrication actually looks like
It is rarely as obvious as a gibberish citation. The dangerous failures are subtle:
Invented cases. A realistic case name and citation for a decision that does not exist.
Fabricated quotations. A real case, cited accurately — but with a quotation that appears nowhere in the opinion.
Misattributed propositions. A real case cited for a holding it does not contain, or that it actually rejects.
Stale law. A real case that has been reversed, overruled, or superseded — presented as good law.
Plausible-but-wrong specifics. A correct case with the wrong court, wrong year, or wrong pin cite.
Every one of these can survive a quick skim. None of them survives an actual check.
What it has cost lawyers
This is not a hypothetical risk. Courts across the United States have sanctioned lawyers for filing briefs built on AI-fabricated citations. A public database maintained by researcher Damien Charlotin tracks court decisions that have flagged AI hallucinations in filings; as of May 2026 it has logged more than 1,400 such matters, and the count climbs steadily.
Consequences that courts have imposed include monetary sanctions, fee awards to the opposing side, briefs struck in whole or in part, referrals to bar disciplinary authorities, and — perhaps most lasting — public opinions naming the lawyers involved. The reputational damage outlives the fine.
The professional-responsibility hooks are direct. ABA Formal Opinion 512 (July 2024) ties a lawyer's use of generative AI to the duties of competence (Model Rule 1.1) and candor toward the tribunal (Model Rules 3.1 and 3.3). Signing a filing is a representation that you stand behind what is in it. "The AI wrote it" has not worked as a defense, and there is no reason to expect it to.
It is not only briefs
Fabricated authority shows up in research memos, contract references to statutes and regulations, due-diligence summaries, and client emails. Anywhere AI cites something, the citation needs to be checked — not just in court filings.
The verification workflow
The fix is not complicated. It is a habit. Every authority an AI gives you is unverified until you have personally confirmed it in a real source.
Assume nothing is real. Treat every case, statute, quotation, and pin cite from the AI as a claim to be checked, not a fact.
Confirm it exists. Pull each authority in a primary source you trust — Westlaw, Lexis, Bloomberg Law, or the official court or government site. If you cannot find it, it is not real.
Read it. Open the actual opinion. Confirm it holds what the AI says it holds — not something adjacent, not the opposite.
Check every quotation verbatim. Find the quoted language in the source and confirm it word for word. Never trust quotation marks the AI produced.
Confirm it is still good law. Run it through a citator (KeyCite, Shepard's, or equivalent). A real case that has been overruled is still a serious error.
Verify the proposition, not just the cite. A genuine case attached to a holding it does not support is a fabrication in substance, even if the citation is real.
If you cannot verify it, cut it. An unverifiable authority does not go in the document. There is no middle ground.
Habits that reduce the risk
Verification is non-negotiable, but a few habits make fabrications less frequent and easier to catch:
Use retrieval-backed tools for research. AI connected to a real legal database, or used through a research platform, hallucinates less than a raw chatbot — but it still hallucinates. The workflow above does not become optional.
Ask for checkable output. Request pinpoint citations and direct quotations rather than paraphrase. Checkable claims are easier to verify and easier to catch when wrong.
Ask the model to flag uncertainty. Instruct it to mark anything it is not confident about and to tell you plainly when it is working from memory rather than a source.
Do not let confidence stand in for accuracy. A model's fluent, certain tone carries no information about whether it is right.
The bottom line
AI is a strong drafting and brainstorming partner and a genuine time-saver. It is not a source of legal authority. Keep those two roles separate and the citation risk becomes manageable.
A pre-filing checklist
Every case, statute, and regulation cited has been pulled in a primary source and confirmed to exist.
Every quotation has been matched word for word to the original.
Every authority has been run through a citator and is still good law.
Each authority actually supports the proposition it is cited for.
Anything that could not be verified has been removed.
A lawyer — not the AI — has read the final document and stands behind it.
Sources & further reading
ABA Formal Opinion 512, "Generative Artificial Intelligence Tools" (July 29, 2024) — ABA announcement