Lawyer Disbarred for Using AI? What Actually Happened

It is 9:30 PM the night before a brief opposing a motion to dismiss is due. An attorney is missing one clean precedent on a procedural point, types a question into a general-purpose chatbot, and gets back judicial opinions complete with docket numbers and quotable paragraphs that fit the argument. The problem surfaces later when opposing counsel and the judge look up those citations and discover the cases never existed. That nightmare is why so many firm owners search lawyer disbarred for using ai before letting anyone in the office touch a new tool.

The short answer: If you searched lawyer disbarred for using ai, here is what the court records actually show: the attorneys in the landmark cases weren't disbarred just for using AI. Instead, they were fined under Rule 11, had pro hac vice admissions revoked, or had their law licenses suspended for filing fake, AI-invented citations without reading the cases—and in some instances, lying to the judge about it. The fix is architectural: never use an open chatbot for legal research, and require any AI working on your firm's files to cite the exact document and page number for every line.

Was a lawyer disbarred for using AI, or what did the courts actually do?

Most partners have heard a half-remembered headline about a lawyer disbarred for using ai. When you read the disciplinary orders and federal docket entries, a clearer pattern emerges. Courts are not punishing lawyers using ai; they are punishing lawyers who sign their names to unverified output and violate their duty of candor to the tribunal.

Here is what happened in two landmark court proceedings:

• Mata v. Avianca, Inc., 678 F. Supp. 3d 443, Case No. 1:22-cv-01461-PKC (S.D.N.Y. June 22, 2023): In a personal injury suit against Avianca, plaintiff's attorney Steven Schwartz used ChatGPT to research precedent opposing a motion to dismiss, and attorney Peter LoDuca filed the brief containing six nonexistent judicial opinions (including Varghese v. China Southern Airlines). Judge P. Kevin Castel did not disbar them; on June 22, 2023, he issued a Rule 11 order imposing a $5,000 joint and several fine on Schwartz, LoDuca, and their firm (Levidow, Levidow & Oberman, P.C.), plus ordering them to mail the sanction order to their client and each judge whose name was falsely attached to the fake opinions.
• People v. Zachariah C. Crabill, Case No. 23PDJ067 (Colo. O.P.D.J. Nov. 22, 2023): Colorado attorney Zachariah C. Crabill used ChatGPT in May 2023 to draft a motion citing fictitious cases, failed to alert the court when he realized they were fake before the hearing, and blamed an intern before confessing six days later via affidavit. On November 22, 2023, the Colorado Presiding Disciplinary Judge suspended Crabill for one year and one day (90 days active suspension, remainder stayed pending a two-year probation) for violating Colo. RPC 1.1, 1.3, 3.3(a)(1), and 8.4(c).

Under Rule 11, signing a filing certifies that an attorney read the authorities and verified the facts. Handing that duty to an unverified chatbot—and covering it up—triggers sanctions and suspensions.

Why do general chatbots invent fake citations and medical facts?

To avoid ai legal issues in your practice, it helps to understand why a consumer chatbot makes things up. Before I became a fractional CTO building AI systems for small businesses, I worked as an aerospace engineer, where you never deploy equipment without knowing its failure mode.

An open-ended language model predicts the next plausible word based on patterns in its training data. When you ask it a question out of the blue, it does not check Westlaw or read your client's PDF; it strings together words that sound like a judicial opinion.

That same failure mode—called a hallucination—hurts a personal injury or litigation firm outside of court briefs. If staff paste raw notes into a chatbot or run an ungrounded summarizer over medical records, the model can invent a Date of Service (DOS), merge two orthopedic visits, or skip an ER triage note. Putting an invented treatment date into a pre-suit demand package or discovery responses (like Interrogatories or Requests for Production) damages your credibility with the adjuster and opposing counsel.

How does a source-grounded AI system prevent hallucinated facts in your firm's files?

You do not fix hallucinations by telling staff to write better prompts in a consumer chatbot. You fix them by changing the architecture of the tool so it cannot invent facts from the open internet.

When I build document tools for a firm, we use RAG, or retrieval-augmented generation, which means the AI is locked to your uploaded documents and looks up the exact passage before answering. If the answer is not in the PDFs you provided, the system says the information is not found rather than guessing. Separate from my aerospace work, I've built RAG systems, including an AI document system that cut search time by 70%.

Every statement the tool outputs links directly to the source document and page number. In my live medical chronology demo, medical records go in and a dated medical chronology (med chron) comes out with every entry cited to its exact page in the PDF, so your paralegal clicks the page citation and verifies the Bates-stamped chart page in seconds.

Modern practice management platforms are moving the same way. Clio's Manage AI extracts deadlines from court documents with side-by-side source verification and a 90-day AI Event Log, while Smokeball's Archie and Filevine's LOIS answer questions from matter documents. Whether you use your practice management software or a custom assistant like my law firm AI assistant demo working alongside tools like Clio or Filevine, the rule stays the same: traceable or it doesn't count.

What rules should a small firm set for lawyers using AI safely?

A small firm owner can protect the practice with four house rules for lawyers using ai and the paralegals supporting them:

• Never use an open chatbot for case law research: Use legal research databases and citators to find precedent, and require the signing attorney to read the opinion before any citation goes into a brief.
• Keep client files out of free consumer chat tools: Never paste client narratives, medical records, or depositions into public web chatbots that train on user inputs.
• Require page-level citations on all internal document work: For medical chronologies, deposition summaries, discovery responses (ROGs and RFPs), and internal procedure lookups, only use source-grounded tools that cite the exact document and page number.
• Keep human verification mandatory before signing: Make it a firm rule that a paralegal or attorney checks the cited source pages before a demand package, discovery response, or filing leaves the office.

What this looks like in practice

Picture an 8-person litigation and personal injury firm preparing a pre-suit policy limits demand and potential complaint after a commercial truck collision:

1. The paralegal uploads the records into a source-grounded workflow: Instead of pasting clinical text into a chatbot, the paralegal runs the Bates-stamped PDFs through a page-cited medical chronology workflow (/medical-chronology).
2. The system builds a dated, page-cited table: The tool extracts each visit by Date of Service (DOS)—listing provider, chief complaint, ICD-10 and CPT details, imaging, and treatment plan—with the exact page citation attached to every entry.
3. The paralegal clicks and verifies the source pages: Before drafting the demand package or discovery responses, the paralegal clicks the page links to verify the lumbar MRI date and check an ER nurse's triage note on prior back stiffness against the chart page.
4. The attorney signs with confidence: When the attorney reviews the demand package or signs motions, every factual contention is backed by a verified page number.

Where AI falls short

Even when you lock an AI tool to your own documents and require page citations, you still need to know its limits:

• Checking whether case law is still good law: A language model is not a citator like KeyCite or Shepard's. It cannot verify whether an appellate opinion was overruled, distinguished in your circuit, or superseded by statute, and under Rule 11, a vendor cannot stand before a grievance panel for you.
• Silent omissions in messy or split files: Restricting an AI system to your firm's documents stops it from inventing fake cases, but it can still miss a handwritten margin note, a faint faxed clinic sheet, or an exhibit split across files. Page citations make review faster, but never replace checking the record.

What I'd do first

Here are four concrete steps you can take this month:

1. Write a one-page verification policy this week: Put in writing that no case citation or medical fact goes into a brief, discovery response, or demand letter unless a human has opened and read the opinion or PDF page.
2. Draw a hard wall between legal research and document summarization: Ban general-purpose chatbots for finding case law, and keep precedent checks inside your verified legal database.
3. Make page numbers mandatory on internal summaries: Require every entry in a medical chronology, deposition summary, or Interrogatory draft to include the exact page or Bates number.
4. Test a page-cited workflow on a real document: Open my medical chronology demo or law firm AI assistant demo to see how a source-grounded system shows the exact passage and page number for every answer.

Start with the problem, build what fixes it, prove it works. To give your team traceable AI tools that cite every page, see how I help law firms or book a call to look at your workflow.

Sources

• Mata v. Avianca, Inc., 678 F. Supp. 3d 443, Case No. 1:22-cv-01461-PKC (S.D.N.Y. June 22, 2023) via CourtListener — https://www.courtlistener.com/docket/63107798/mata-v-avianca-inc/ — Supports the facts that plaintiff's attorneys submitted a brief citing six nonexistent judicial opinions generated by ChatGPT (including Varghese v. China Southern Airlines), and that Judge P. Kevin Castel issued a 34-page Rule 11 sanction order on June 22, 2023 imposing a $5,000 joint and several fine on Steven Schwartz, Peter LoDuca, and Levidow, Levidow & Oberman, P.C., plus mandatory notification letters to their client and the judges whose names were falsely used.
• People v. Zachariah C. Crabill, Case No. 23PDJ067 (Colo. O.P.D.J. Nov. 22, 2023) — https://www.coloradosupremecourt.us/pdf/PDJ/2023/23PDJ067_Stip.pdf — Supports the facts that Colorado attorney Zachariah C. Crabill used ChatGPT in May 2023 to draft a motion citing fictitious cases, failed to alert the court upon discovering they were fake before the hearing, falsely blamed a legal intern before confessing six days later via affidavit, and was suspended on November 22, 2023 for one year and one day (90 days active suspension, remainder stayed pending a two-year probation) for violating Colo. RPC 1.1, 1.3, 3.3(a)(1), and 8.4(c).