The Lawyer Who Used ChatGPT: What Happened to Steven Schwartz

The lawyer who used ChatGPT is Steven A. Schwartz, and the case he used it in is Mata v. Avianca. Schwartz, a New York attorney, was fined 5,000 dollars in June 2023 after a court filing he prepared cited six judicial opinions ChatGPT had invented. He used ChatGPT for legal research, asked the chatbot whether the cases were real, and filed what it told him. He was not disbarred, and he remains registered to practice in New York.

This article sets out what Schwartz used ChatGPT for, which six cases it invented, and how the court found out. It then covers why his colleague was sanctioned for a filing Schwartz researched, what the penalty actually required, and the verification step that transfers to lawyers in private practice. The account of the sanction below comes from the court's own 43-page opinion.

[cta-1]

Key Takeaways

The lessons for lawyers from the ChatGPT fake citations case are about who verifies a citation, at what moment, and against what source. Steven Schwartz asked ChatGPT for case law, asked ChatGPT whether that case law was real, and filed the answer. The court then sanctioned the lawyer who signed the filing as well as the lawyer who researched it.

1. The ChatGPT fake citations resolved. Reporter citations in the filing pointed at real published material, including the Miller cite, which lands inside Greenleaf v. Garlock, so a lawyer who checked only whether a citation resolves would have cleared them.
2. The duty attaches at signature. Peter LoDuca signed the filing and carried the Rule 11 certification for research he had not done, and Rule 11(c)(1) reached Schwartz as well.
3. Asking ChatGPT is not verification. Schwartz asked ChatGPT whether the cases were real, and the answer it gave him had the same provenance as the fake citations.
4. The order required a letter to the client. Judge Castel directed the respondents to write to the plaintiff, Roberto Mata, as well as to each judge a fabricated opinion had named.
5. Nobody was disbarred. The sanction for the ChatGPT fake citations was one 5,000 dollar penalty imposed jointly and severally on two attorneys and their firm, plus the letters.

Who Was the Lawyer Who Used ChatGPT?

The lawyer who used ChatGPT was Steven A. Schwartz, a New York attorney admitted in 1992 and in practice for 31 years by the time he used ChatGPT on the Mata filing. Schwartz practices workers' compensation and personal injury law at Levidow, Levidow and Oberman, a small Manhattan firm. The lawyer who used ChatGPT was not a novice, and he was not the attorney of record once the case reached federal court.

What Schwartz's Practice Looked Like

Four facts about the practice explain how a 31-year attorney ended up in this position, and none of them is inexperience.

1. He was admitted to the New York bar in 1992, after taking his law degree at New York Law School.
2. His practice is a plaintiff-side personal injury and workers' compensation practice, built on construction accident and malpractice matters.
3. Levidow, Levidow and Oberman is a small firm, without the research infrastructure or the supervising layers a large litigation department carries.
4. His habitual forum is New York state court, not the Southern District of New York.

Federal practice was not Schwartz's daily environment, so the certification rules he was found to have violated were not rules he worked under every week, a background Legal Dive set out in detail when the sanctions issued. The court did not treat unfamiliarity as an excuse, and the opinion nowhere suggests it should be.

What Did the Lawyer Use ChatGPT For?

The lawyer used ChatGPT for three things: to find case law supporting a tolling argument under the Montreal Convention, to produce the full text of the decisions it had named, and to confirm that those decisions were real. Schwartz used ChatGPT the way a lawyer uses a research associate, and he treated its answers as a research result rather than as generated text. He did not open a reporter, a docket or a research database to check any of them.

The Three Requests He Made

1. A research request: find authority that the Montreal Convention's two-year filing deadline can be extended.
2. A production request: supply the full text of the decisions just named.
3. A verification request: confirm that those decisions are real.

The Question He Was Trying to Answer

The substantive question was narrow. Mata's claim against Avianca had been filed outside the Montreal Convention's two-year window, and Schwartz needed authority that the window could be extended because of Avianca's bankruptcy. That is a question of federal appellate law, and whether the professional conduct rules permit a lawyer to use a consumer chatbot for it at all is a separate question from whether the output was accurate. ChatGPT supplied six airline cases that answered the question in Mata's favor, and none of them existed.

What Fake Cases Did ChatGPT Cite?

The six fake cases ChatGPT cited were Varghese v. China Southern Airlines, Shaboon v. Egyptair, Petersen v. Iran Air, Martinez v. Delta Airlines, Estate of Durden v. KLM Royal Dutch Airlines and Miller v. United Airlines. Every one of the six was fabricated, and every citation looked ordinary enough to pass a reading. The Federal Reporter citation on the fabricated Miller decision, 174 F.3d 366, falls inside a real and unrelated opinion, Greenleaf v. Garlock, which begins at 174 F.3d 352.

The Six Cases as They Were Filed

They were set out in the court's April 11, 2023 order, which directed LoDuca to annex copies of the decisions he had cited, with their fabricated reporter citations attached.

One real case sat in the same list. Ehrlich v. American Airlines, Inc., 360 N.J. Super. 360 (App. Div. 2003), is a genuine New Jersey appellate decision, which is part of why the filing did not announce itself as fabricated on sight. The sanctions opinion sets out each of the six and reproduces the fake Varghese decision in full as an appendix.

What the Fabricated Opinions Actually Looked Like

The fake Varghese opinion is the one the court examined closely, and the defects in it are the kind a practitioner can see.

1. The panel is wrong. The opinion seats Judge Higginbotham, a senior judge of the Fifth Circuit, on an Eleventh Circuit panel alongside Judges Jordan and Rosenbaum, who do sit on the Eleventh.
2. The reasoning does not hold together. The court's assessment was blunt: the legal analysis is gibberish.
3. The filing date is implausible. The Eleventh Circuit's electronic case filing system began operating in 2010, so a 2019 opinion absent from it is difficult to explain.
4. The internal citations are incoherent. The fake Petersen cites itself as precedent, and it confuses the District of Columbia with the state of Washington.

How the Court Found Out the Cases Were Fake

The court found out the cases were fake because nobody could locate them. Avianca told the court on March 15, 2023 that it could not find the authorities quoted in the opposition brief, and the court's own search could not find several of them either. Judge Castel then ordered the full texts produced. Schwartz asked ChatGPT for them, filed what it generated, and the fabricated Varghese opinion seated a Fifth Circuit judge on an Eleventh Circuit panel.

The Timeline From Filing to Sanction

The window to correct the record stayed open for three months.

What Turned an Error Into a Bad-Faith Finding

What moved the court from error to bad faith was the silence. Judge Castel found bad faith based on acts of conscious avoidance and false and misleading statements to the court, and recorded that Schwartz knew the full text of Varghese could not be found before the March 1 filing and said nothing. The opinion is explicit that the record would look quite different had the respondents come clean after Avianca's March 15 reply.

The Screenshots Schwartz Filed

The screenshots annexed to Schwartz's May 25 affidavit are reproduced in the opinion.

1. He asked the chatbot whether Varghese was a real case.
2. He asked whether the other cases it had provided were fake.
3. ChatGPT answered that it had supplied real authorities findable on Westlaw, LexisNexis and the Federal Reporter.

He treated that exchange as confirmation. The court did not. The confirmation and the fabrication came from the same place.

Why LoDuca Was Sanctioned When Schwartz Did the Work

Peter LoDuca was sanctioned because he signed the filing. Schwartz is not admitted to practice in the Southern District of New York, so when Avianca removed the case to federal court LoDuca filed the notice of appearance while Schwartz continued to perform all substantive legal work. A signature carries its own certification, and Rule 11(c)(1) reaches the attorney who did the work as well, which is why Schwartz was sanctioned alongside him. LoDuca testified that he did not read Avianca's brief questioning the cases before he responded to it.

LoDuca filed that notice of appearance on March 31, 2022, and the opinion records the arrangement plainly: he appeared because Schwartz is not admitted, while Schwartz continued to perform all substantive legal work.

Where the Exposure Actually Sits

For a practitioner the transferable point is about division of labor, not about chatbots. The signing attorney certifies research performed by someone else, which is the same exposure that arises in four familiar arrangements.

1. A partner signing a filing built on an associate's research.
2. A firm signing work produced by a contract attorney or a staffing agency.
3. In-house counsel signing a position worked up by outside counsel.
4. Any lawyer signing a document whose substance came out of software.

When the Check Has to Run

The certification attaches at the moment of signature, which means the verification check belongs in the signing workflow rather than in the drafting workflow. A review step that runs before a document reaches the signatory is the step that discharges the duty, and a step that runs after it does not.

What Were the Consequences of Using ChatGPT in These Legal Proceedings?

The consequences of using ChatGPT in these legal proceedings were a 5,000 dollar penalty and a set of mandatory letters, not a suspension and not disbarment. Judge Castel imposed the penalty jointly and severally on Peter LoDuca, Steven Schwartz and Levidow, Levidow and Oberman P.C., payable into the Registry of the Court. The order also required each respondent to write, within 14 days, to the plaintiff Roberto Mata and to every judge whose name a fabricated opinion carried, attaching the opinion, the hearing transcript and the April 25 affirmation.

What the Penalty Required

1. It is one penalty, not two. The 5,000 dollars was imposed jointly and severally on three respondents, and it was paid into the court rather than to Avianca.
2. The firm was a respondent. Levidow, Levidow and Oberman P.C. carried liability alongside both attorneys, so the exposure was not confined to the individuals who touched the filing.
3. The client had to be told. The letter to Roberto Mata put the error in front of the person paying for the work.

Why the Research Failure Alone Was Not Enough

The court was explicit that the research failure by itself would have been a lesser matter, and the distinction is the difference between a mistake and a sanctionable one.

1. Sloppy research alone would have been objectively unreasonable under Rule 11, which is a lower finding than bad faith.
2. What justified the bad-faith finding was the conscious avoidance that followed Avianca's March 15 reply.
3. The false statements made to the court while the filing stood uncorrected compounded the original error.

[cta-2]

Was the Lawyer Who Used ChatGPT Disbarred?

The lawyer who used ChatGPT was not disbarred. Steven Schwartz was not suspended either. The sanction was a 5,000 dollar penalty imposed jointly and severally on Schwartz, Peter LoDuca and the Levidow firm, plus an order to write to the client and to each judge whose name a fabricated opinion carried. Schwartz remains registered to practice law in New York, and no public disciplinary determination against him has issued from the Appellate Division.

What the Registration Record Shows

The New York Unified Court System's attorney registration record lists Steven Allan Schwartz at Levidow Levidow and Oberman, admitted in 1992, with a status of currently registered. The court system's own search pages are not reachable to automated clients, so that record is also published in an accessible copy. The status is a registration status rather than a disciplinary clearance, and it is worth reading as the narrow fact it is.

Federal Sanction and State Discipline Are Separate Tracks

A Rule 11 sanction and a state disciplinary proceeding run on separate tracks, and the outcome of one does not resolve the other.

1. A federal court sanctions conduct in a case before it, under the rules of that court and its inherent power.
2. Attorney discipline in New York sits with the Appellate Division's grievance committees, which act on their own record and on their own timetable.

A monetary sanction can therefore stand without any disciplinary determination following it, which appears to be what happened here.

What Happened to Mata's Case Against Avianca

Mata lost. Judge Castel granted Avianca's motion to dismiss on June 22, 2023, the same day he issued the sanctions order, on the ground that the claim was time-barred under the Montreal Convention's two-year deadline. Roberto Mata was injured in August 2019 and the operative complaint against Avianca was filed in February 2022.

What Mata Alleged

Mata's own claim is on the record.

1. Mata alleged that a metal service cart struck his knee on an Avianca flight in August 2019.
2. The claim was first brought in New York state court.
3. It was affected by Avianca's bankruptcy, and was then filed as the present action in February 2022.
4. Avianca removed the action to federal court, where the docket terminated on July 7, 2023.

Why the Dismissal Matters to the Story

The fabricated cases were invented to win the exact point the court decided against.

1. Schwartz went looking for authority that the Montreal Convention's two-year window could be extended.
2. ChatGPT produced six airline decisions holding that it could.
3. The court held the claim was out of time, so the proposition would not have carried the case even if the authority behind it had been real.

This filing carried a verification failure and a merits failure at once.

What the ChatGPT Lawyer Case Changed for Lawyers

What the ChatGPT lawyer case changed for lawyers is that fabricated citations stopped being a hypothetical risk and became a docketed one. Mata v. Avianca was the first widely reported sanction of its kind, and the penalties imposed on lawyers since have gone well beyond a single 5,000 dollar fine. Courts and tribunals now log these filings in volume: the AI Hallucination Cases database maintained by Damien Charlotin records 2,149 matters globally, 1,473 of them in the United States.

The Layers That Moved After It

1. The ethics layer. The American Bar Association addressed generative AI directly in ABA Formal Opinion 512, which treats the existing duties of competence, confidentiality and reasonable fees as the governing framework rather than creating new ones.
2. The court layer. A growing number of federal judges now require a certification about AI use in filings, usually in the form of a standing order rather than a local rule.
3. The penalty layer. Monetary sanctions, referrals to disciplinary bodies and public admonitions have all followed since.

The specific penalties, the orders and the running count are set out case by case in the record of lawyers who have been fined for AI-generated citations, which is the fuller landscape rather than the single incident described here.

How to Verify AI-Generated Legal Citations Before Filing

Verifying AI-generated legal citations before filing means pulling and reading the opinion text, not confirming that the citation resolves to something. The verification step Schwartz skipped was that second read, and it matters because the fabricated citations in the Mata filing did point at real material: inside the invented Varghese decision, 92 F.3d 1074 holds two brief Eleventh Circuit orders in unrelated cases, 772 F.3d 1278 is Witt v. Metropolitan Life, and the Federal Reporter cite given for Zicherman, a decision the invented Varghese opinion cited in turn, is Miccosukee Tribe v. United States. A citation that resolves is not evidence that the case it names exists.

The Checks That Would Have Caught This Filing

Each check below is written as a trigger rather than a principle, because a principle does not say when to act.

1. Pull the opinion text every time a citation comes from a generative tool. The trigger is the tool, not the suspicion, and reading the opinion is also what catches a real case cited for a proposition it does not hold.
2. Check the panel against the court before relying on an unfamiliar appellate decision. A Fifth Circuit judge sitting on an Eleventh Circuit panel is an error a lawyer who knows the bench sees in one read.
3. Never accept a model's answer about its own output. Schwartz asked whether the cases were real and the answer came from the same process that produced them.
4. Run the check at signature, not at drafting. The signing attorney carries the certification, so the step belongs in the workflow that precedes signature, and it covers work from associates, contract attorneys and vendors as well as from software.
5. Record who verified what, and against which source. The respondents' problem after 15 March was that no one could say, which is what turned a research failure into a bad-faith finding.
6. Decide the client-disclosure and billing position before it is needed. The order here required a letter to the client, and a firm that has not settled in advance who writes that letter settles it under pressure.

Where the Discoverability Risk Sits

The verification record cuts both ways. A log of prompts and outputs is what shows that a check was performed, and it is also material that may be discoverable or produced to a court if a filing is later questioned. Schwartz's own screenshots became an exhibit to a published federal opinion. A policy that mandates a record should therefore say where the record lives and who can reach it.

What Goes Into the Tool

A separate check applies to inputs. A consumer chatbot's default terms may permit the provider to retain what is typed in and use it to improve the model, so the question of what a general chatbot does with client information is a confidentiality question distinct from accuracy. A firm policy that addresses only output verification leaves the input side unanswered, and the confidentiality duty does not wait for an inaccurate citation to be triggered.

Drafting Tools Carry the Same Obligation

Nothing in the above is specific to case law. A tool that produces a brief, a memorandum or a motion raises the same traceability question as a tool that produces a citation, and the choice of software for drafting briefs and memoranda is partly a choice about whether the output can be traced back to something readable. Where it cannot, the verification burden falls entirely on the reader, which is the position Schwartz was in without recognizing it.

Spellbook and ChatGPT for Legal Work: Where Spellbook Fits, and Where It Does Not

Spellbook and ChatGPT are built for different legal work, and neither one would have done the job Schwartz needed done. Spellbook is a contract tool: it works on the agreement in front of the lawyer, and it does not perform litigation case-law research, so it would not have found or checked an Eleventh Circuit tolling precedent. ChatGPT is a general-purpose chatbot that will answer a case-law question whether or not it holds an answer.

Where the Same Risk Shows Up in Transactional Work

That limitation is the honest frame for a page about this case, and it also locates where a transactional lawyer meets the identical risk. The risk is not fabricated case law. It is an assertion about a document that a lawyer cannot trace back to the document, and it appears in three ordinary places.

1. A redline that recommends a position without showing which clause it responds to.
2. A summary of a counterparty's paper that paraphrases a term rather than quoting it.
3. An answer about market practice with no benchmark behind it that can be opened.

How Spellbook Is Built to Be Checked

The product is built so the underlying material can be opened.

1. The Review feature proposes redlines as native tracked changes in Word, with citations that link back to the source language in the contract, so a lawyer reading a suggestion can see what it was drawn from.
2. Spellbook Ask answers legal questions grounded in 130+ legal research sources and attaches the citation to the answer.
3. The Compare to Market feature measures a clause against 2,300+ benchmark standards drawn from 20M+ real contracts, so a market-practice claim has a stated denominator.

The difference from a general chatbot is where the answer comes from, not a claim that a wrong answer is impossible. A lawyer still reads the clause.

A Side-by-Side on Four Dimensions

The scope row is the one that matters on this page. Spellbook is a commercial contract specialist used by 5,000+ in-house teams and law firms across 80+ countries, and none of that reaches a question about Eleventh Circuit precedent. A firm that needs case-law research needs a research platform, and the discipline of reviewing a document against a defined standard is a different discipline from checking whether a cited authority exists. Pricing is quoted per seat and scales with team size, with a free seven-day trial available without a sales call.

Frequently Asked Questions

Should Lawyers Stop Using ChatGPT or AI Tools?

No. The sanction in Mata v. Avianca did not turn on the use of a generative tool. It turned on filing unverified output and then failing to correct the record once the problem was raised. The practical response is to match the tool to the task, verify output against a primary source before it reaches a filing, and decide in advance who performs that check and records it.

Can ChatGPT Be Used to Represent Clients in Court?

No. Representing a client in court requires admission to the bar of that court, and the rules on the unauthorized practice of law do not admit software. ChatGPT cannot appear, cannot sign a filing, and cannot carry the Rule 11 certification, which is why the Mata sanction fell on two admitted attorneys and their firm rather than on the tool that produced the citations.

Does Using ChatGPT Affect the Lawyer-Client Relationship?

It can, in two directions. Time saved on mechanical work can be redirected to the work that requires a lawyer's judgment. Entering client information into a consumer tool may also expose confidential material, and a disclosure obligation can follow an error, as it did here: the court required a letter to the client. Whether use needs to be disclosed, and whether it affects a fee, are questions better settled in a firm policy than case by case.

Is a Legal-Specific AI Tool Safer Than ChatGPT for Legal Work?

A legal-specific AI tool is safer than ChatGPT for legal work on two measurable dimensions and on no others. The first is data handling: legal tools are commonly sold with zero data retention terms and SOC 2 Type II reports, while a consumer chatbot's default terms may permit the provider to train on what is typed into it. The second is grounding: a tool that answers from a defined source set and attaches the citation can be checked, whereas a general chatbot generates the citation along with the answer. Neither dimension removes the lawyer's own verification duty, and no tool makes a filing accurate.

Are There Other Lawyers Who Have Been Fined for Using AI?

Yes, and the penalties have grown larger than the 5,000 dollars imposed here. Courts in several jurisdictions have sanctioned attorneys for fabricated citations since 2023, with monetary penalties, referrals to disciplinary bodies and public admonitions among the recorded outcomes. The running record of those matters is logged case by case, and the pattern across them is more instructive than any single figure.

[cta-3]

Where This Leaves a Transactional Practice

Schwartz's error was not the use of a chatbot. It was that nobody opened the thing the chatbot produced. The transferable obligation is that any output reaching a signature has to be traceable to a source the signing lawyer can read, whether the work came from software, an associate or a vendor, and that the check runs before the signature rather than after it.

What That Changes in the Workflow

1. Name the person who verifies, for each document type, before the document exists.
2. Put the check in front of the signature rather than after the draft.
3. Require the source, not the summary, for anything a generative tool produced.
4. Keep the verification record somewhere the firm can find it under pressure.

The Review feature is built for that read on the contract side: it shows what a suggestion was drawn from, in Word, before anyone signs. Try Spellbook free to see how the Review feature handles a contract already on the desk.