Imagine paying a lawyer to file a brief on your behalf, only to have a federal judge publicly note that your argument "collapses like an overmixed soufflé" the moment anyone checks the citations — because the cases don't exist. Not "hard to find." Not "outdated." Invented, apparently by an AI tool, and never caught before the brief hit the judge's desk.
That's exactly what happened in Clinco v. Commissioner, T.C. Memo. 2026-16, decided by U.S. Tax Court Judge Mark Holmes on February 9, 2026. It's the first time a Tax Court judge has explicitly flagged likely AI hallucinations in a filed brief — and it's a preview of a problem that's about to touch a lot more small businesses than you'd think, precisely because more of them are using AI tools to draft the paperwork behind their own finances.
What Actually Happened in the Case
Peter Clinco was a Los Angeles attorney who also ran MedCafe Westwood, a restaurant and bar near UCLA, spending 25–30 hours a week on the business alongside his legal practice. On his 2015 joint tax return — filed more than two years late, in September 2018 — Clinco reported gross receipts of roughly $1.6 million but claimed a net loss of about $400,000.
The IRS wasn't satisfied. Revenue Agent Yi Liu ran a bank deposits analysis across eight separate accounts and cross-referenced third-party information returns (Forms 1099-K and 1099-MISC). The reconstruction put MedCafe's actual gross receipts at approximately $2.29 million — about $2.2 million more than what Clinco had reported. Separately, Clinco had claimed $56,798 in depreciation deductions on two Pasadena rental properties without providing purchase invoices, closing statements, or any documentation establishing basis or placed-in-service dates.
On the merits, none of this went well for the taxpayer:
- Unreported income stood. The court found the IRS's bank-deposit reconstruction reasonable because it relied on third-party documentation plus Clinco's own admissions about the restaurant's cash percentage. Clinco argued some deposits were capital contributions rather than revenue, but the court rejected that theory — an email referencing a contribution, without more, "is not such a showing."
- The depreciation deduction was disallowed. Claiming the same depreciation again in a later year proves nothing about entitlement to it in the year at issue. Without basis, cost, and placed-in-service documentation, Section 167(a) offers no relief.
- A procedural challenge to the Notice of Deficiency failed too — Clinco's counsel argued the notice was invalid because it lacked a "wet" (physical) signature. The court dismissed this, citing established precedent (Tavano v. Commissioner) and IRS Manual provisions that permit delegated signatures.
The Citations That Didn't Exist
It's that last argument — the signature challenge — where things got strange. To support it, Clinco's attorney, Abraham Wagner, cited three cases:
- Cacchillo v. Commissioner, 130 T.C. 132 (2008)
- Miller v. Commissioner, 57 T.C. 440 (1971)
- Tefel v. Commissioner, 118 T.C. 324 (2002)
None of them exist. Judge Holmes wrote that Cacchillo v. Commissioner "does not, however, exist" and noted that the page citation for one of the others actually pointed to an unrelated case, Porter v. Commissioner. Taken together, the fabricated bundle of citations, in Holmes's words, "suggests something cooked up by AI."
Here's the part worth sitting with: no sanctions were imposed in this case. Judge Holmes stopped short of that step, but he didn't stop short of a warning. He wrote that "submitting a brief with fictitious caselaw is a recipe for sanctions" and specifically flagged Federal Rule of Civil Procedure 11(b) — the rule that requires attorneys to certify their filings are grounded in existing law — as the framework a future court could use to actually punish this. As one legal commentary summarizing the ruling put it, Holmes's closing note amounted to: the Tax Court hasn't sanctioned anyone for this yet. Yet.
This Is Bigger Than One Restaurant's Tax Return
Clinco isn't an isolated glitch. It's a data point in a fast-accelerating trend. A public tracker of AI hallucination cases in courts worldwide (maintained by legal researcher Damien Charlotin) has documented the pace accelerating sharply:
| Period | Documented cases with AI-fabricated citations |
|---|---|
| Mid-2025 | ~200 |
| January 2026 | 719 |
| April 2026 | 1,227 |
| June 2026 | 1,598 |
That's a rate of nearly 8 new cases per day by mid-2026, up from 5–6 a day just two months earlier. Hundreds of licensed attorneys have been named in these rulings, not just self-represented litigants unfamiliar with legal research. And the consequences have escalated well past a slap on the wrist:
- The landmark 2023 case, Mata v. Avianca, resulted in a $5,000 fine and public reprimand after two attorneys filed a brief citing six fake ChatGPT-generated cases.
- By late 2025, Couvrette v. Wisnovsky produced a record aggregate sanction of roughly $109,700.
- Q1 2026 alone saw more than $145,000 in AI-related sanctions across U.S. courts.
- A June 2026 ruling (Withers v. Aberdeen) went further still, imposing two-year suspensions and canceling a scheduled trial outright.
Courts have also started holding supervising attorneys personally liable when a junior associate's AI-assisted work goes unchecked — the fabrication doesn't have to be yours to become your problem if your name is on the filing.
Why a Small Business Owner Should Care
You're probably not filing Tax Court briefs. But the underlying failure mode in Clinco — trusting AI-generated output without verifying it against a primary source — shows up everywhere small business owners now use AI:
- Drafting demand letters, contracts, or responses to an IRS notice. If you or your bookkeeper use a general-purpose AI tool to draft correspondence citing tax code sections, regulations, or case law, verify every citation against the actual source (IRS.gov, the Cornell Legal Information Institute, or your CPA) before it goes out the door.
- Generating explanations for deductions. AI tools are good at producing plausible-sounding justifications for why an expense should be deductible. Plausible isn't the same as correct, and it's certainly not the same as documented. As Clinco shows, the IRS and the courts don't accept "it sounds right" — they want basis, invoices, placed-in-service dates, and a paper trail.
- Relying on AI-summarized guidance for a new law. Tax law changes constantly (OBBBA provisions, state-level regulatory shifts, and more all landed in 2026 alone). An AI model's training data has a cutoff, and even tools with live web access can misattribute or fabricate specifics. Cross-check anything with real financial consequences against a primary source.
- Bookkeeping automation. AI-assisted categorization and reconciliation tools are genuinely useful, but the same discipline applies: spot-check the output, especially anything that will end up supporting a deduction if you're ever audited.
The common thread in Clinco — and in the 1,598 other documented hallucination cases — isn't that AI is unusable. It's that nobody checked the work before it became part of the official record. That's a process failure, and process failures are exactly what good bookkeeping habits are built to catch.
Keep Your Financial Records Auditable, Not Just AI-Assisted
Clinco is ultimately a case about documentation: a bank-deposit reconstruction beat an undocumented capital-contribution theory, and undocumented depreciation didn't survive scrutiny even after two rounds of claiming it. Good records — not persuasive-sounding arguments — are what hold up when the IRS reconstructs your income from third-party data.
That's the same discipline plain-text accounting is built around. Beancount.io keeps every transaction in transparent, version-controlled plain-text ledgers — so if you ever need to reconstruct exactly how a deduction, contribution, or expense was recorded (and when), the history is right there, not buried in an AI-generated summary you have to hope was accurate. Get started for free and keep your books as auditable as your arguments should be.