By Rick Yandle, Esq., CPA — federal criminal & civil tax defense
An IRS examiner helped write the law section of a civil fraud case with ChatGPT. We found it in the examiner's own activity record, obtained under FOIA, dated 05/12/2026.
We file a lot of Freedom of Information Act requests. Most of what comes back is routine. This one was not.
In a civil fraud case we are defending, we asked the IRS for the examiner's activity record ("time sheets"). That is the running log a revenue agent keeps on every case (IRS Form 9984). One entry, dated 05/12/2026, stopped me. The examiner wrote that she "researched Westlaw, Bloomberg, and received assistance from ChatGPT to update the law section of the fraud lead sheet."
Not a scheduling note. Not filing. The legal analysis behind a federal fraud penalty, prepared with help from ChatGPT, a tool that is known to invent case law.
What the record says
Here is the entry, word for word from the examiner's own log ("time sheet"):
"The Fraud Advisor gave the examiner feedback on the lead sheet. The examiner continued working on civil fraud lead sheet. The examiner researched Westlaw, Bloomberg, and received assistance from ChatGPT to update the law section of the fraud lead sheet."
Form 9984 is not written for the public. It is the agent's own log ("time sheet entries"), and we have it only because we asked for it under FOIA. Two things stand out. A Fraud Enforcement Advisor was already involved, so this was a developed fraud case, not a routine audit. And when it came time to write the law, the examiner reached for ChatGPT right next to the paid research services.
What a civil fraud lead sheet is
The fraud lead sheet is the workpaper where an examiner builds the fraud case. The agent lists the "badges of fraud," ties them to the taxpayer, and sets out the law.
The stakes are high. The civil fraud penalty runs 75 percent of the portion of the underpayment attributable to fraud (Section 6663), and once the IRS proves any part of the underpayment is fraudulent, the rest is treated as fraud too unless the taxpayer shows otherwise. To get there the government has to prove fraud by clear and convincing evidence, a much harder standard than the one that governs an ordinary audit. The law section of that sheet is where the government does that work.
It can also go criminal. If the examiner establishes affirmative acts of fraud and the criminal criteria are met, the examination is suspended and the Fraud Enforcement Advisor can recommend a referral to IRS Criminal Investigation, which then decides whether to take it. So the quality of that analysis is not a small thing.
The problem with ChatGPT here
Consumer AI makes up legal authority. This is documented, not a theory. A federal court in New York sanctioned lawyers who filed a brief full of cases ChatGPT had invented, quotes and all (Mata v. Avianca). Other courts have done the same since. The tool is fluent, confident, and wrong often enough that no careful lawyer trusts it on the law without checking every cite against the real book.
Put that tool on the government's side of a fraud case and the concerns line up. If any authority on that sheet came from ChatGPT, it has to be run down against real law before it supports a 75 percent penalty or a criminal referral; the record tells us the tool was used, not what was checked. The risk is broader than a fake citation: AI also writes confident analysis that misstates holdings and glosses over the distinctions that decide fraud cases. And a public chatbot is not a closed government system, so what taxpayer facts went into the prompts is a fair question.
What IRS policy required
IRS policy in effect at the time let employees use only Treasury- or IRS-approved generative AI. It barred entering taxpayer information or other nonpublic material into a public or unapproved system, and it required a human to verify AI-generated work. What the activity record does not tell us is which ChatGPT environment the examiner used, whether it was approved, what was entered, or how the output was checked. Those are the questions the entry raises.
What it means under audit
When the government claims fraud, we want to know how it built the case, not just the number at the bottom. The examiner's own record raises fair questions: what system was used, what was entered, what it produced, and how the resulting analysis was verified. Whether those facts end up affecting the administrative disposition, or become relevant in litigation, depends on what the answers show.
We are not saying a fraud case falls apart because an agent used ChatGPT. We are saying the government has to prove fraud by clear and convincing evidence, and our job is to find out exactly how it plans to do that. FOIA is one way we look.
What we are doing next
The examiner's ChatGPT use may have left prompt-and-response records, session history, or audit logs. None of that was in the FOIA response. So we asked whether it exists and was retained. That is a separate story, and it is not over.
A note for referring counsel
If you handle criminal defense, family law, or business matters and a client's tax problem is turning toward civil or criminal fraud, how the IRS built its case is worth scrutiny early, including what tools the examiner used to get there. I am always glad to consult with fellow attorneys whose clients have crossed into federal tax exposure.
Facing an IRS audit or criminal tax exposure? Schedule a case evaluation.