Park v. Kim and the Second Circuit's AI warning shot

On January 30, 2024 — seven months after Judge Castel sanctioned the lawyers in Mata v. Avianca — the United States Court of Appeals for the Second Circuit handed down Park v. Kim, 91 F.4th 610.

The case is short. The opinion is dense. The conduct described is, on first read, hard to believe — because by the time Attorney Jae S. Lee filed her reply brief citing a ChatGPT-fabricated decision, the entire legal profession had been talking about Mata for half a year.

She did it anyway.

What Park v. Kim added to the AI sanctions landscape isn't a new factual pattern. It's a new procedural posture, a new enforcement vehicle, and a new floor for what federal courts will consider "the basic obligations of counsel." Every litigator with an appellate practice needs to understand what changed.

The underlying case

Before getting to the brief, the underlying case matters — because the AI sanctions in Park arrived on top of an already-broken representation.

Minhye Park, a South Korean resident, filed a medical malpractice action against Dr. David Dennis Kim in the United States District Court for the Eastern District of New York in June 2020. The case proceeded under federal diversity jurisdiction. Discovery was contentious — much of the medical record was located in South Korea, and the defendant pressed for disclosure that the plaintiff resisted.

Over the course of two years, Judge Pamela Chen entered multiple discovery orders. The plaintiff repeatedly failed to comply. The court warned, on more than one occasion, that continued noncompliance would result in dismissal. The plaintiff continued to resist.

In August 2022, Judge Chen dismissed the case under Federal Rules of Civil Procedure 37 (discovery sanctions) and 41(b) (failure to prosecute). The Second Circuit later characterized the noncompliance as "sustained and willful intransigence." The plaintiff appealed.

That is the case that landed on the Second Circuit's docket. Before any AI question was raised, the underlying representation had already failed at the district court level. The appellate posture was, in candid terms, weak.

The reply brief

On appeal, Park's counsel — Attorney Jae S. Lee of JSL Law Offices in Uniondale, New York — was charged with defending the position that the district court had abused its discretion in dismissing the case.

Lee's opening brief was filed late after two extensions. The reply brief was filed July 25, 2023 — more than a week past the extended deadline. It included this citation:

Matter of Bourguignon v. Coordinated Behavioral Health Servs., Inc., 114 A.D.3d 947 (3d Dep't 2014)

The case does not exist. The Appellate Division reporter volume and page are real. The Third Department is real. The date is plausible. The case is fabricated.

When the Second Circuit could not locate the decision, the court ordered Lee to produce a copy. She could not. In her response, Lee acknowledged that she had used ChatGPT to find precedent that might support her arguments. She had not, in her own words, "read or otherwise confirm the validity of the (non-existent) decision she cited."

She had not read it because there was nothing to read.

The Second Circuit's response

The court's analysis of Lee's conduct occupies a small portion of the opinion, but its language is precise enough to be quoted in essentially every state bar's AI guidance issued since.

The panel — Judges Barrington Parker, Allison Nathan, and Sarah Merriam — wrote that "citation in a brief to a non-existent case suggests conduct that falls below the basic obligations of counsel." The court then took two actions: it referred Lee to the Court's Grievance Panel, and it ordered her to furnish a copy of the decision to her client — translated, since the client was Korean-speaking.

Three things make this disposition distinct from Mata.

First, it was a federal Circuit court ruling, not a district court order. Mata set a sanctions precedent for one district. Park established the same proposition at the appellate level, binding within the Second Circuit and persuasive everywhere else. When a federal appeals court says that filing a fabricated citation falls "below the basic obligations of counsel," that language travels.

Second, the enforcement mechanism was different. Rule 11 of the Federal Rules of Civil Procedure — which Judge Castel relied on to sanction the lawyers in Mata — does not apply in the courts of appeals. The appellate analog is Federal Rule of Appellate Procedure 38, which the Second Circuit did not invoke. Instead, the court used its inherent authority to refer Lee to its Grievance Panel. The Grievance Panel — not the three-judge merits panel — would decide whether to impose discipline, what discipline, and on what timetable. The available range includes formal reprimand, suspension, and disbarment.

Third, the sanction has no fixed financial ceiling. Schwartz and LoDuca paid $5,000 each. Lee's exposure is open-ended; the Grievance Panel can impose any discipline its rules allow.

Why the conduct mattered more, not less, the second time

The legal profession had been talking about Mata v. Avianca for seven months when Lee filed her reply brief. Every state bar had been moving toward AI ethics guidance. ABA Formal Opinion 512 was already in the drafting pipeline. Mata had been covered in Bloomberg Law, the ABA Journal, every legal trade publication, and several mainstream outlets.

Lee's defense — that she had used ChatGPT and not verified the result — was already a defense the legal profession had rejected. By July 2023, the "I didn't know the AI could fabricate" position was no longer a colorable explanation. It was an admission.

This is the part of Park that the bar guidance picked up on most aggressively. The court did not need to lecture Lee on why her conduct was problematic. It just had to describe what she did. The legal community had already done the work of establishing that this conduct was unacceptable.

In the months following Park, federal and state courts began citing both Mata and Park together in AI-related sanctions orders. Park did not displace Mata. It compounded it. Where Mata answered the question "is this conduct sanctionable," Park answered "is this conduct still sanctionable now that everybody knows."

The answer was yes.

The discovery noncompliance link

There is one analytical thread in Park v. Kim that often gets lost in the AI-sanctions framing. The Second Circuit affirmed the underlying dismissal because of discovery noncompliance — months and months of court orders ignored, of warnings given, of dismissal threatened. The AI fabrication came on appeal, defending the resulting dismissal.

In other words: the lawyer whose representation produced sustained discovery violations at the trial level was also the lawyer who used ChatGPT to write the brief defending that representation, without verifying the citations.

The two failures are not unrelated. Both reflect a particular kind of practice failure — one where procedural compliance is treated as an optional layer rather than the foundation of professional representation. In the trial court, that meant ignoring discovery orders. In the appeals court, it meant not reading the cases cited in one's own brief.

For appellate practitioners, this is the deeper lesson. Appellate practice has always demanded a higher standard of preparation than trial practice — there are fewer opportunities to recover from a mistake, the panel reads every brief carefully, and the case authority cited is the entire weight of the legal argument. Using generative AI to draft an appellate brief without verifying every citation is a category mismatch. The tool's failure mode is precisely the work the appellate brief is supposed to do well.

What appellate practitioners need to understand

For lawyers who practice in any federal or state appellate court, Park v. Kim establishes four operating principles.

One — the duty applies to every paper filed. This is obvious but worth stating. Reply briefs are still briefs. Procedural motions are still court submissions. Letter briefs filed after argument are still subject to candor and verification duties. The Grievance Panel referral in Park came on a reply brief — a paper that some lawyers, inappropriately, treat as a lower-stakes document. The court treated it the same as any other filing.

Two — appellate sanctions move through a different procedural vehicle, and the timeline is longer. Unlike Rule 11, where the trial court can resolve sanctions quickly, Grievance Panel referrals move on their own schedule. The lawyer's exposure persists for as long as the Panel's process lasts. The reputational damage is immediate; the disciplinary outcome can be years away.

Three — appellate courts have inherent authority to act independent of Rule 11. This matters because some practitioners assumed that the appellate courts were procedurally hamstrung in AI-citation cases. Park foreclosed that assumption. The court did not need a specific rule. Its inherent authority to police the conduct of counsel before it was sufficient.

Four — every state bar's AI guidance now cites Park alongside Mata. The case is no longer a Second Circuit precedent in any practical sense. It is national authority on what AI conduct in appellate practice violates the rules. Practitioners in jurisdictions outside the Second Circuit who think the case doesn't apply to them are mistaken.

What changed between Mata and Park, and what didn't

Mata v. Avianca established that generative AI could fabricate plausible-looking authority and that lawyers using it without verification could be sanctioned. Park v. Kim established that the lesson of Mata was binding on every lawyer practicing in federal court regardless of whether they had personally read the Mata opinion.

What didn't change between the two cases: the underlying failure mode. In both, the lawyer used a consumer generative AI tool to find authority. In both, the tool produced fabricated citations. In both, the lawyer did not independently verify against an authoritative source. In both, the lawyer's defense rested on the proposition that she had trusted the AI to be accurate.

The proposition was unavailable. Both courts said so. The accumulated weight of bar guidance, ethics opinions, and case law that has emerged since says so even more clearly.

If you are a lawyer using generative AI for legal research, the questions Park v. Kim puts to you are simple. Did the case you are about to cite actually pass through an independent validation step against an authoritative legal database? Can you produce, on demand, the contemporaneous record of that validation? Is that record tamper-evident and time-stamped?

If the answer to any of those questions is no, your workflow is not the workflow that Mata and Park require. It is the workflow that produced Mata and Park.

The infrastructure question

We built Discover Docket — and the DDEAS framework inside it — to make the Park v. Kim failure mode structurally impossible.

JILL cannot return a citation that hasn't been validated against current authoritative case law for the jurisdiction your matter sits in. The validation step is not a checkbox the lawyer remembers to tick. It is the gate between the AI's output and the lawyer's screen. If a case doesn't exist, the lawyer never sees the citation, because the workflow doesn't allow it.

Beyond validation, every output JILL produces carries a confidence score, every action is recorded in a cryptographic audit log, and the log is chained into a tamper-evident sequence. The lawyer who needs to demonstrate — to a court, a grievance panel, or a malpractice carrier — that her work was rigorously verified has the record. The lawyer who needs to do that work without infrastructure is in the position Attorney Lee was in: explaining herself to a court that has already decided what her conduct represents.

We built DDEAS so that no lawyer using Discover Docket is ever in that position.

Read the DDEAS framework →

If you want the longer analysis of how the architecture closes the Mata v. Avianca failure modes, the companion article walks through it.