Tax Court Sanctions Warning Over AI-Hallucinated Case Citations

Tax Court issues stark warning on AI verification duties

A Tax Court judge has delivered a sharp rebuke to a tax attorney who cited three nonexistent legal cases that appeared to be AI hallucinations, marking what may be the first Tax Court opinion to address fabricated citations generated by large language models.

In Clinco, T.C. Memo. 2026-16, Judge Mark V. Holmes warned that "submitting a brief with fictitious caselaw is a recipe for sanctions" after the taxpayer's attorney cited four cases supporting an argument about deficiency notice signatures—three of which did not exist. The IRS flagged the errors, but the attorney failed to correct the record and even repeated one fabricated citation in a subsequent filing.

The underlying dispute involved restaurant owner Peter L. Clinco's 2015 tax return, where the IRS determined gross receipts were underreported by more than $2.2 million. The court upheld the IRS determination, but Holmes devoted significant attention to the briefing problems that he said suggested "something cooked up by AI."

The fabricated authorities

The three cases Holmes identified as likely hallucinations illustrate how AI can generate plausible-sounding but entirely fictitious citations:

• Cacchillo, 130 T.C. 132 (2008): The cited page actually falls within an unrelated case about innocent-spouse relief
• Miller, 57 T.C. 440 (1971): The page reference lands in a case about renegotiation boards with no mention of deficiency notices
• Tefel, 118 T.C. 324 (2002): The citation points to a case concerning S corporation management fees

Holmes noted that while the Tax Court lacks a direct equivalent to Federal Rule of Civil Procedure 11(b), attorneys appearing before the court remain bound by the American Bar Association's Model Rules of Professional Conduct, which prohibit knowingly making false statements of law.

Why it matters

This decision arrives as tax professionals rapidly integrate generative AI into research and drafting workflows. Unlike traditional legal research platforms, large language models generate responses by predicting language patterns—they can produce authoritative-sounding citations to cases that never existed. The professional stakes extend beyond judicial sanctions: Treasury Circular 230 governs practice before the IRS, and the IRS Office of Professional Responsibility recently released "Introductory Guidelines for Responsible AI Use in Federal Tax Practice." For AICPA members, the Statements on Standards for Tax Services make clear that while AI may assist research, practitioners remain responsible for verifying all authorities, facts, and legal conclusions. Firms are now establishing written AI use policies requiring human verification of all cited authorities before filing.

Practical safeguards

Tax practitioners should implement specific controls when using AI tools. Written firm policies should define permitted uses, require human review of all AI-assisted work, and establish citation verification procedures. For litigation work, some firms now maintain verification logs or require second reviewers to cross-check authorities before filing.

The IRS Office of Professional Responsibility guidance emphasizes that responsible AI use requires firm-level governance including training, supervision, secure data protocols, and documentation. Practitioners must treat AI output as draft material requiring independent verification against authoritative sources.

Holmes emphasized that "such apparitions have made frequent appearances in legal briefing in recent years" and are "unacceptable." Several federal courts have now issued standing orders requiring attorneys to certify that AI-generated research has been independently verified.

The details were first reported by the Journal of Accountancy, which noted this appears to be the first Tax Court case addressing potential AI hallucinations in legal briefing.