When AI Gets Tax Law Wrong: Fake Citations, Hidden Prompts, and How to Check Your Work

AI tools now draft IRS response letters, summarize notices, and explain tax rules in seconds. Used well, they save time. Used carelessly, they produce something courts have started calling "AI slop": polished, confident text built on authority that does not exist. Tax cases are no exception, and the cost of a fabricated citation or a misquoted rule can land on the taxpayer as well as the advisor.

This post looks at two AI risks that matter for anyone dealing with the IRS: made-up legal authority and hidden instructions planted in documents. It ends with a practical checklist for using AI on tax matters without getting burned. Our earlier posts covered privacy when sharing returns with chatbots and how the IRS itself uses automated systems; this one is about the reliability of what AI hands back to you.

Why AI Invents Tax Law

Large language models predict likely words. They do not look up the Internal Revenue Code the way a research database does unless they are connected to one, and even then they can blend sources. Ask for a case supporting your position and a model may produce a name, a citation, and a quote that all look right. The format is convincing because the model has seen thousands of real citations. The substance may be invented.

Tax law makes this worse in a few ways:

• Numbers change every year. Inflation adjustments, phase-outs, and new legislation such as the One Big Beautiful Bill Act mean a model trained on older material may quote last year's figure, or a figure from a version of a bill that never passed.
• Similar rules sound alike. The time limit for filing a Tax Court petition, the deadline to request a Collection Due Process hearing, and the window to appeal a rejected offer in compromise are different, yet easy to confuse.
• Arithmetic is not a strength. Language models can make basic calculation and date-counting mistakes, which is dangerous when a deadline or a penalty depends on the count.

What Courts Are Doing About Fake Citations

Courts have now seen hundreds of filings containing AI-generated authority that does not exist, and the consequences are growing:

• The early warning. In Mata v. Avianca (S.D.N.Y. 2023), lawyers who filed a brief citing invented cases were fined $5,000 and publicly rebuked.
• A tax case. In Clinco v. Commissioner, T.C. Memo. 2026-16, the Tax Court found that cases cited in a brief appeared not to exist and called their presence unacceptable. It imposed no sanction that time, but noted that other courts are punishing this conduct more seriously.
• Steeper penalties. In March 2026, the Sixth Circuit in Whiting v. City of Athens fined each of two attorneys $15,000 and ordered them to cover the other side's appellate attorney costs and double costs after finding fabricated citations and misstatements in their briefs.

The lesson for taxpayers is direct. If you represent yourself in Tax Court, as many small-case petitioners do, the same honesty duties apply to you. Section 6673 lets the Tax Court impose a penalty of up to $25,000 when a taxpayer takes a frivolous or groundless position or uses the proceedings primarily for delay. Outside of court, § 6702 provides a $5,000 penalty for frivolous submissions to the IRS. An AI-drafted letter that relies on made-up law will not win you any goodwill with the examiner reading it.

A Subtler Trap: Real Citation, Wrong Content

Confirming that a citation is real does not finish the job. A model may pair a real volume and page with a case name that does not belong there, or attach a real case name to a holding it never made. The only reliable test is to open the actual source and read the passage you are relying on. If you cannot find the language the AI quoted, do not use it.

The same goes for IRS forms and publications. AI will sometimes describe a form line or an instruction that changed in a later revision. Always compare against the current version on IRS.gov, and check the revision date in the corner of the form.

Prompt Injection: When the Document Gives the Orders

The second risk runs the other direction. Prompt injection is the practice of burying commands in a file so an AI tool reading it follows those instructions instead of its user's. The text might be white on a white background, set in a tiny font, or tucked into metadata. A human reader never sees it; the AI does. Security groups such as OWASP list prompt injection as the top risk for applications built on large language models.

In August 2026, a Connecticut Superior Court sanctioned a self-represented plaintiff in Elliott v. New York Bariatric Group, LLC for embedding hidden text in his filings that told any AI reviewing them to side with him. He kept doing it after being warned, and the court revoked his electronic filing privileges, requiring paper filings in person. The court treated the attempt itself as the wrong, even though no court AI had read the filings.

Why should taxpayers and business owners care?

• Documents you receive may carry hidden text. If you paste a vendor contract, an invoice, or a counterparty's settlement proposal into an AI tool and ask for a summary, a hidden instruction could tilt the summary. Treat AI summaries of other people's documents as a starting point, not a conclusion.
• AI agents that act on email are exposed. Tools that read your inbox and can draft replies, pay bills, or move files can be steered by a cleverly crafted message. Limit what these tools are allowed to do without your approval.
• Never plant hidden text yourself. It may seem like a clever way to influence an automated reviewer at the IRS or a lender. Courts view it as deception, and in a tax setting it could undercut your credibility for the entire case.

A Practical Checklist for Using AI on Tax Matters

• Ground the tool in real sources. Give it the actual notice, the IRS page, or the code section, and ask it to answer only from those materials.
• Demand specific cites. Ask for the exact section, regulation, or page for every claim, then look each one up yourself on IRS.gov, the Tax Court site, or an official code source.
• Do the math separately. Use a calculator or spreadsheet for amounts and count deadlines on a calendar. Many IRS deadlines are counted from the notice date itself, not from the day it reached your mailbox.
• Check the year. Confirm that any dollar limit or rate is for the right tax year and reflects 2025 and 2026 law changes.
• Keep sensitive data out of public tools. Remove Social Security numbers and account details before using any consumer AI service.
• Keep a record. Save what you asked, what the tool said, and what you verified. If a question about your position comes up later, you can show your process.
• Have a professional review anything you file. A short review costs far less than a rejected argument, a penalty, or a missed deadline.

An Example

A hypothetical business owner receives a proposed penalty and asks a chatbot to write a reasonable cause letter. The draft cites two Tax Court cases and quotes a "first-year abatement" rule with a dollar limit. On review, one case cannot be found, the other is about a different penalty, and the dollar limit is invented. The real path is the IRS first-time abatement policy, which looks at the prior three years of compliance and has no dollar ceiling of that kind. With accurate facts and the correct policy, the corrected letter is shorter and far more persuasive.

Key Takeaways

• AI tools can produce convincing tax citations, quotes, and figures that are wrong or invented.
• Courts, including the Tax Court, are calling out fabricated authority, and sanctions are rising.
• Self-represented taxpayers face the same honesty duties, plus possible § 6673 and § 6702 penalties.
• Prompt injection can hide instructions in documents; treat AI summaries of others' documents with caution and never hide text yourself.
• Verify every source in the original, do the math yourself, and have filings reviewed before they go out.

Want a second set of eyes on an IRS letter, notice response, or Tax Court filing before it goes out? Schedule a consultation with SMB CPA Group, contact us with your questions, explore our solutions, or review our pricing.

Disclaimer: The information presented on this website or any SMB CPA Group, PC online platform is for general information and illustrative purposes only. It should not be considered tax, legal, financial, or other professional advice. The content is not intended to provide definitive answers or solutions to specific business situations and is not a substitute for careful research or the advice of a licensed professional that knows your unique circumstances.

Readers are advised not to rely solely on SMB CPA Group materials or use them as the basis for any business, legal, tax, or accounting decision without first seeking independent subject matter expertise and counsel. All case studies shown are hypothetical and intended for demonstration purposes only; results shown are not guarantees of performance or outcomes.

Please contact an SMB CPA Group professional directly to discuss your specific questions or business situation. Our team would be happy to speak with you about a tailored consultation to your needs. We also encourage all readers to seek counsel from licensed attorneys, financial advisors, CPAs, Enrolled Agents, or other qualified professionals prior to making decisions related to their finances or enterprises.