When AI Writes Your Brief: Lessons from Fivehouse v. DoD
It was the kind of mistake you might expect from a first-year associate pulling an all-nighter, not from a 30-year veteran of the U.S. Attorney’s Office. In a federal courtroom in the Eastern District of North Carolina, Assistant U.S. Attorney Rudy Renfer filed a brief riddled with fabricated quotations, misattributed case holdings, and invented regulatory language — all generated by artificial intelligence and never verified against original sources.
The most stinging irony? The errors were caught not by opposing counsel at a white-shoe firm, not by a federal judge’s law clerk, but by the plaintiff himself — a pro se litigant representing himself without an attorney. Colonel Derence V. Fivehouse, a retired Air Force officer and experienced military lawyer, simply did what Renfer did not: he read the cited sources.
What followed was a cascade of professional consequences that ended a decades-long legal career and produced one of the most devastating judicial reprimands in the emerging canon of AI-related legal misconduct. The case of Fivehouse v. U.S. Department of Defense is now required reading for every attorney who has ever thought about pressing “generate” and then “file.”
The Case: Fivehouse v. U.S. Department of Defense
The underlying dispute was about healthcare, not technology. Colonel Derence V. Fivehouse, a retired U.S. Air Force staff judge advocate, brought suit against the Department of Defense over a policy change to TRICARE for Life — the government health insurance program that supplements Medicare for military retirees. The policy change at issue excluded coverage for GLP-1 receptor agonist medications, the class of drugs widely used for weight loss that has surged in popularity in recent years.
Fivehouse was no ordinary pro se litigant. As a former staff judge advocate — the senior legal officer advising military commanders — he had extensive legal training and decades of experience navigating complex regulatory frameworks. He chose to represent himself, and he brought to the case the meticulous attention to detail that his military career had demanded.
On the other side stood the full weight of the federal government. Rudy Renfer, the Assistant U.S. Attorney assigned to defend the Department of Defense, had spent 17 years in the U.S. Attorney’s Office for the Eastern District of North Carolina. He was, by all accounts, a seasoned litigator with a 30-year legal career behind him. The case should have been routine government defense work.
It became anything but.
The Brief That Unraveled
The problems began when Fivehouse sought to supplement the administrative record in the case. Renfer filed a response brief arguing that Fivehouse’s request was based “on speculation rather than evidence.” The brief cited Fourth Circuit precedent, quoted from federal regulations, and marshaled case law in support of the government’s position.
On the surface, it looked like a competent legal filing. Underneath, it was a minefield of fabrication.
The brief misquoted a Fourth Circuit case. It falsely attributed quotations to other cases that had never contained that language. It incorrectly described the holdings of several decisions it cited, transforming narrow, fact-specific rulings into sweeping endorsements of the government’s position. And — perhaps most alarmingly — it included at least two fabricated quotations attributed directly to the Code of Federal Regulations, the body of law that carries the force of statute.
These were not minor errors of citation format or page numbering. These were wholesale inventions: language that did not exist anywhere in the legal corpus, presented to a federal court as authoritative law.
| How Fivehouse Caught It
Colonel Fivehouse grew suspicious when the government’s brief “did not read like the sources it cited.” The phrasing felt wrong — too convenient, too perfectly tailored to the government’s argument. So he did what any diligent attorney would do: he pulled the cited authorities. What he found was that quoted language did not appear anywhere in the cited opinions or regulations. Narrow, fact-specific rulings had been reimagined as sweeping doctrinal pronouncements. Fivehouse methodically documented these discrepancies in his own reply submissions, laying out the evidence citation by citation. |
A pro se plaintiff — representing himself against the Department of Justice — had outperformed a 30-year government attorney on the most basic obligation of legal practice: making sure your citations are real.
Why AI Hallucinations Happen
To understand how an experienced attorney’s brief could contain wholesale fabrications, it helps to understand what generative AI actually does — and what it decidedly does not do.
Large language models (LLMs) like those powering ChatGPT, Copilot, and similar tools are, at their core, pattern-prediction engines. They generate text by predicting the most probable next word in a sequence based on statistical patterns learned from vast training data. They do not retrieve verified facts from a database. They do not look up cases on Westlaw. They do not consult the Federal Reporter or the Code of Federal Regulations.
Instead, they generate text that sounds like it belongs in a legal brief — because they have processed millions of examples of legal writing. This creates a particularly dangerous failure mode: the outputs are fluent, confident, and structurally convincing, even when the underlying content is entirely fabricated.
Legal citations are especially vulnerable to hallucination for several reasons:
- Structured format creates false confidence. A fabricated citation that follows the correct format — party names, volume number, reporter abbreviation, page number — looks authentic to a reader who does not verify it. The model has learned the pattern of a citation without understanding the substance.
- Case holdings are compressed and paraphrased. LLMs generate summaries of what a case probably held based on statistical association, not what it actually A narrow ruling about standing in a FOIA case can morph into a broad pronouncement about administrative law.
- Regulatory text is particularly susceptible. The Code of Federal Regulations is vast, technical, and frequently amended. LLMs can generate plausible-sounding regulatory language that blends real provisions with invented text, creating chimeric citations that are nearly impossible to detect without checking the original source.
- Confidence calibration is absent. Unlike a human researcher who might hedge — “I think the court said something like…” — an LLM presents its fabrications with the same authoritative tone as its accurate outputs. There is no internal signal to the user that the model is uncertain.
The result is a tool that can produce a convincing first draft of a legal brief in minutes — complete with citations that look right, read right, and are entirely wrong.
The Cover-Up That Made It Worse
If the fabricated brief was the first mistake, what came next was arguably worse. When Fivehouse flagged the errors in his reply submissions, Renfer had a choice: come clean immediately, or minimize and deflect. He chose the latter.
Renfer first asked the court for permission to replace the brief with a substitute document — a move that, by itself, might have been unremarkable. But his explanation was a masterclass in understatement. In a brief, two-page filing dated January 20, Renfer characterized the problem as a “clerical error,” stating that “the error was clerical in nature and resulted from the inadvertent filing of an unfinalized draft document.”
There was no mention of artificial intelligence. No acknowledgment that entire quotations had been fabricated. No candor about the nature or scope of the inaccuracies.
U.S. Magistrate Judge Robert T. Numbers II was not persuaded. On March 2, he issued a sharply worded order expressing “serious concerns” about the accuracy and credibility of the government’s filing. He ordered a show-cause hearing — the judicial equivalent of being summoned to the principal’s office, but with a court reporter and potential sanctions waiting.
The show-cause hearing took place on March 10. Renfer appeared alongside U.S. Attorney W. Ellis Boyle, a signal that the gravity of the situation had registered with the leadership of the office. Under oath, Renfer finally admitted the truth: he had used generative AI to draft the brief. He explained that he had accidentally overwritten a draft while under deadline pressure and had turned to AI to recreate it, believing he had corrected the AI’s output before filing.
| Judge Numbers repeatedly told Renfer that his accounting of events was “difficult to believe” and that his explanations “strain credulity.”
— From the March 10, 2026 show-cause hearing, Eastern District of North Carolina |
The progression — from fabricated brief, to “clerical error” excuse, to sworn admission under judicial pressure — transformed what might have been a correctable mistake into a crisis of candor. In the legal profession, the duty of honesty to the tribunal is not a suggestion. It is a foundational obligation. And Renfer had violated it not once, but twice: first by filing the brief, and then by obscuring how it was created.
The Professional Consequences
The fallout was swift and severe.
After the March 10 hearing, Renfer indicated a willingness to retire at the end of May. But the Department of Justice did not wait. The day after the hearing — March 11 — the involuntary termination process began in the morning and was completed by the afternoon. A 30-year legal career ended in a matter of hours.
On April 28, 2026, Judge Numbers issued an 18-page formal order of admonishment and reprimand. The document is remarkable for the precision and severity of its language. While the judge elected not to impose a financial fine — reasoning that Renfer had already lost his job and suffered severe financial and reputational harm — the written reprimand was devastating.
| “Renfer’s professional reputation, both locally and nationally, is in tatters. In this court, his name will be synonymous with a failure to uphold the basic duties of competence and candor expected of every attorney.”
— U.S. Magistrate Judge Robert T. Numbers II, April 28, 2026 |
Judge Numbers found that Renfer “intentionally submitted a brief containing false materials to the court” and made clear that fabricated authority “will not be tolerated.” The order referred Renfer to the Department of Justice’s Office of Professional Responsibility for further investigation — a process that could result in additional disciplinary consequences, including potential disbarment referrals.
U.S. Attorney Boyle, for his part, called the filing “unacceptable.” He issued a warning letter to all staff in the office and scheduled mandatory training on AI use in legal practice to prevent similar incidents.
| Consequence | Detail |
| Involuntary Termination | Initiated and completed the day after the show-cause hearing (March 11, 2026) |
| Judicial Reprimand | 18-page formal admonishment issued April 28, 2026 by Judge Numbers |
| OPR Referral | Referred to DOJ’s Office of Professional Responsibility for investigation |
| Office-Wide Response | Warning letter to all staff; mandatory AI training scheduled by U.S. Attorney Boyle |
| Reputational Damage | Judge Numbers wrote that Renfer’s name would be “synonymous with a failure to uphold the basic duties of competence and candor” |
How Renfer Could Have Prevented This: A Practical Guide for Attorneys
The Fivehouse case is not a cautionary tale about the dangers of AI. It is a cautionary tale about the dangers of complacency, shortcuts, and the failure to apply the same professional rigor to AI-assisted work that attorneys are trained to apply to every other aspect of their practice. Here are ten concrete steps that would have prevented this outcome.
1. Verify Every Citation Against Original Sources
This is the non-negotiable rule. Every case cited in a brief must be pulled from an authoritative legal database — Westlaw, Lexis+, official court reporters, or government regulatory databases — and read in the original. No exceptions. If an AI generates a citation, treat it as an unverified lead, not an established fact. Check the case name, the reporter citation, the page number, and — critically — the quoted language and the holding.
2. Read Every Cited Case in Full
It is not enough to confirm that a case exists. You must confirm that it says what you claim it says. AI models routinely transform narrow holdings into broad propositions. A case about standing in a specific regulatory context becomes a sweeping rule about administrative deference. The only way to catch this is to read the case — not a summary, not a headnote, but the opinion itself.
3. Use AI as a Drafting Assistant, Not a Research Replacement
Generative AI can be genuinely useful for structuring arguments, identifying potential lines of reasoning, drafting routine language, and overcoming writer’s block. What it cannot do is legal research. It does not access legal databases. It does not Shepardize cases. It does not check whether a regulation has been amended. Use it for the writing. Do the research yourself.
4. Cross-Reference AI Output with Established Legal Databases
Build a workflow where every AI-generated draft goes through a systematic cross-referencing process. Run every citation through Westlaw or Lexis. Check every regulatory reference against the current CFR on eCFR.gov. Verify that quoted language appears verbatim in the source. This should be a required step, not an optional one.
5. Implement a Personal Verification Checklist
Before filing any document, walk through a checklist:
- Have I independently verified every case citation?
- Have I confirmed that every quotation appears verbatim in the cited source?
- Have I checked that my characterization of each holding is accurate?
- Have I verified all regulatory citations against current, official text?
- Have I Shepardized or KeyCited every case to confirm it remains good law?
- Has a second person reviewed the brief? If not, why not?
6. Request Extensions Rather Than Cut Corners
Renfer stated that he turned to AI because he accidentally overwrote a draft while under deadline pressure. Courts grant extensions. They do it routinely. A motion for a brief extension of time, honestly explained, will never end a career. A fabricated brief will. When the pressure mounts, the professional response is to ask for more time — not to outsource accuracy to a machine.
7. Be Transparent with the Court About AI Use
If AI contributed to the drafting of a document and errors are discovered, immediate transparency is essential. Renfer’s initial “clerical error” explanation — with no mention of AI — compounded the original problem and destroyed his credibility. Courts are far more likely to treat an honest mistake with leniency than a discovered cover-up. Candor is not just an ethical obligation; it is the best strategic choice.
8. Understand AI’s Fundamental Limitations
Attorneys using AI tools must understand what these tools are. They are pattern-prediction engines trained on text. They generate plausible language, not verified facts. They do not “know” the law. They do not access databases in real time (unless specifically connected to retrieval tools). Outputs can be confidently wrong. Understanding this is not optional — it is a prerequisite for competent use.
9. Maintain a Human-in-the-Loop at Every Stage
AI should never be the final author of any legal filing. A human attorney must review, verify, and take personal responsibility for every word in every document filed with a court. This is not a bureaucratic formality. It is the mechanism that catches the errors AI cannot see in itself.
10. Stay Current on Court-Specific AI Disclosure Requirements
Courts across the country are adopting rules that require attorneys to disclose AI use in filings or to submit sworn statements that all citations have been verified by legal staff. Ignorance of these local requirements is not a defense. Check each court’s standing orders, local rules, and any applicable state bar guidance before filing.
| The Golden Rule of AI in Legal Practice
If you would not stake your law license on the accuracy of a sentence, do not file it. AI-generated text demands the same — or greater — scrutiny as text drafted by the most junior associate in your office. The signature on the filing is yours. The responsibility is yours. No algorithm changes that. |
The Bigger Picture: Courts Respond
Fivehouse v. DoD is not an isolated incident. It joins a growing list of AI-related legal missteps that have prompted sanctions, reprimands, and professional humiliation across the country. Attorneys in New Jersey and Alabama have faced sanctions for filings containing fabricated citations generated by AI tools. The pattern is consistent: an attorney uses AI to draft or supplement a brief, fails to verify the output, and files a document containing invented authorities.
The judiciary is responding. Courts like the Western District of North Carolina now require sworn statements confirming either that no AI was used in the preparation of a filing or that all AI-assisted citations have been independently verified by a licensed attorney or legal staff. Other jurisdictions are developing similar requirements.
Bar associations are issuing formal guidance emphasizing three pillars: supervision (AI output must be reviewed by a competent attorney), transparency (courts and clients should know when AI tools are used), and verification (every citation and factual assertion must be confirmed against authoritative sources).
These developments signal a profession that is not rejecting AI — but demanding that its use meet the same standards of diligence and honesty that govern every other aspect of legal practice.
Conclusion: The Duty That Cannot Be Outsourced
Artificial intelligence is a powerful tool. It can accelerate drafting, illuminate arguments, and help attorneys manage crushing workloads. Used responsibly, it has the potential to expand access to justice and improve the quality of legal work.
But the story of Rudy Renfer and Fivehouse v. U.S. Department of Defense makes one thing unmistakably clear: the duties of competence and candor cannot be delegated to an algorithm. An AI model does not have a law license. It does not take an oath to the court. It does not face disbarment. The attorney does.
Colonel Fivehouse, representing himself against the Department of Justice, did what the government’s own attorney failed to do. He read the sources. He checked the citations. He held the opposing party to the standards the profession demands. That a pro se plaintiff became the quality-control backstop for a DOJ filing is not just embarrassing — it is a warning.
| “In this court, his name will be synonymous with a failure to uphold the basic duties of competence and candor expected of every attorney.”
— U.S. Magistrate Judge Robert T. Numbers II |
The lesson is not that attorneys should avoid AI. The lesson is that AI makes the fundamentals more important, not less. Verify your sources. Read your citations. Be honest with the court. And never, ever let a machine file something you have not personally confirmed to be true.
The technology is new. The obligations are not.
| Key Takeaways
• AI-generated legal citations are prone to hallucination — plausible-sounding fabrications that do not correspond to real authorities. • Every citation in an AI-assisted brief must be independently verified against original sources before filing. • When errors are discovered, immediate transparency with the court is both an ethical obligation and the best strategic response. • Attempting to characterize AI-generated fabrications as “clerical errors” compounds the misconduct and destroys credibility. • Courts are rapidly adopting disclosure requirements and verification mandates for AI-assisted filings. • The professional consequences of unverified AI filings include termination, judicial reprimand, bar referral, and permanent reputational damage. |
This article is intended for educational and informational purposes and does not constitute legal advice. The facts described are drawn from publicly available court records and news reporting on Fivehouse v. U.S. Department of Defense, Eastern District of North Carolina.