The Tax Court spotted three citations an AI invented. Nobody else did.

On February 9, 2026, the U.S. Tax Court decided an unremarkable case about a Los Angeles restaurant's underreported cash receipts — and used four pages of the opinion to do something it hadn't done before: rule on what to do when a lawyer's own brief cites cases that were invented by a large language model.

Three of the four cases the taxpayer's attorney cited to challenge an IRS notice of deficiency don't exist. None of the three were subtle errors — a wrong year, a transposed digit. They were confident and specific: real-looking case names, real volume and page numbers, and a holding that fit the argument exactly.

Three cases that were never decided

The attorney, representing the late Peter Clinco's estate, argued that the notice of deficiency was invalid because no IRS employee had signed it by hand. To support that theory he cited Cacchillo v. Commissioner, 130 T.C. 132 (2008) — a case holding, he wrote, that an unsigned notice of deficiency strips the Tax Court of jurisdiction. "Cacchillo v. Commissioner" does not exist. Page 132 of volume 130 of the Tax Court Reports falls inside Porter v. Commissioner, a case about an unrelated innocent-spouse standard of review.

• Miller v. Commissioner, cited as 57 T.C. 440 (1971) — page 440 of that volume falls inside Winfield Manufacturing Co. v. Renegotiation Board, which never mentions a notice of deficiency.
• Tefel v. Commissioner, cited as 118 T.C. 324 (2002) — no case by that name exists; the page cited is a paragraph of Hillman v. Commissioner, about S corporation management fees.

The Commissioner's lawyers flagged all three in their answering brief. Clinco's attorney didn't withdraw them — he listed Cacchillo again, in his table of authorities, in his reply.

The persuasiveness of Clinco's argument collapses like an overmixed soufflé when one looks at the citations used to prop it up.

Judge Mark Holmes, Clinco v. Commissioner, T.C. Memo. 2026-16

A warning, not a sanction

Submitting a brief with invented caselaw, Judge Holmes wrote, "is a recipe for sanctions and a clear violation of Rule 11(b)" — and he quoted Chief Justice John Roberts's own verdict on the practice for good measure: "Always a bad idea." Then the court didn't impose one.

That restraint is doing real work. Sanctioning a lawyer for a research shortcut nobody can prove he took would punish the wrong thing — the citation-checking failure is the actual defect, whatever produced the citations, and it doesn't require a confession to matter.

The check nobody ran

Circular 230 doesn't distinguish between a lawyer's case citation and a preparer's authority for a return position. Section 10.22 requires the same independent verification either way — a duty tax return review already exists to satisfy for one of them, and Clinco shows exactly what its absence looks like for the other.

Where this stops being a court's problem

A restaurant's cash receipts made this a Tax Court case, not a tax return review. But the failure sitting inside it shows up the moment a firm asks an AI tool to support a position with authority instead of prose — a code section, a revenue ruling, a case on point. The tool will produce one. Whether it exists is exactly as checkable, and exactly as unchecked, as it was in Clinco's brief — unless the check is attached to a specific reviewer and a specific line, not left to whoever happens to reread the file before it goes out.