The Incident

On April 9, 2026, Sullivan & Cromwell — one of Wall Street’s most prestigious law firms, founded over 140 years ago and home to more than 900 attorneys — filed an emergency motion in a Chapter 15 bankruptcy proceeding before Chief Judge Martin Glenn of the U.S. Bankruptcy Court for the Southern District of New York.

The case involved Prince Global Holdings Limited, a group of British Virgin Islands entities whose founder, Chen Zhi, had been federally indicted in October 2025 on charges of orchestrating forced labor operations and large-scale investment fraud in Cambodia.

Ten days later, on April 18, S&C partner Andrew Dietderich — co-head of the firm’s Global Finance & Restructuring practice and a Chambers Band 1 practitioner of nearly three decades — sent a letter to the judge that will be studied in law schools and compliance departments for years.

The filing had contained AI “hallucinations.” Fabricated case citations. Misquoted legal authorities. Non-existent legal sources. Garbled references to the Bankruptcy Code. And the errors weren’t caught internally — they were flagged by opposing counsel at Boies Schiller Flexner.

Dietderich’s letter to Judge Glenn stated plainly: “The inaccuracies and errors in the Motion include artificial intelligence (‘AI’) hallucinations. ‘Hallucinations’ are instances in which artificial intelligence tools fabricate case citations, misquote authorities, or generate non-existent legal sources. We deeply regret that this has occurred.”

The Irony That Made This Global News

If this had happened at a small regional firm or a solo practitioner’s office, it would have been a footnote. It wasn’t.

Sullivan & Cromwell actively touts on its own website its role advising OpenAI — the creator of ChatGPT — on its partnerships with Microsoft, Oracle, CoreWeave, and AMD, and on the “safe and ethical deployment” of artificial intelligence. The firm has a dedicated AI practice that counsels organizations on AI governance, adoption, and compliance frameworks.

The firm that bills itself as a trusted guide for navigating AI risk had to apologize to a federal judge for not following its own internal AI policies.

That is not a punchline. That is a compliance failure case study.

What the Firm’s Own Policies Said

Dietderich’s letter didn’t hide behind ambiguity. He disclosed the details of S&C’s internal AI governance program, which included:

  • Two mandatory training modules on proper AI use in legal work
  • Tracked completion records to confirm attorney participation
  • Office Manual language explicitly instructing lawyers to “trust nothing and verify everything” when using AI tools
  • Comprehensive citation verification procedures before filing

The policies existed. The training existed. The safeguards existed. And yet, multiple layers of attorney review — at one of the most expensive law firms on earth — failed to catch fabricated citations before they were submitted to a federal court.

The errors were only caught after opposing counsel at Boies Schiller Flexner reviewed the filing. In a notable twist, BSF itself had previously been in the same position — having submitted AI-hallucinated citations in a high-profile brief.

The Remediation Response

To S&C’s credit, their response was substantive. The April 18 letter was accompanied by a detailed Schedule A cataloguing dozens of corrections across multiple documents, including:

  • Revised case citations
  • Corrected quotations from Chapter 15 precedents
  • Clarifications to supporting declarations
  • Replacement or removal of authorities that appear to have been entirely hallucinated

The firm undertook a full re-review of all filings in the Prince matter, confirmed no other AI-related issues existed, and filed a corrected version of the motion with a redline. Dietderich also personally called Boies Schiller Flexner to thank them for identifying the errors and apologize directly.

This Is Not an Isolated Incident

The S&C case is extraordinary because of the firm’s prestige and the irony of its AI advisory work. But the underlying problem is anything but unusual.

Researcher Damien Charlotin has catalogued over a thousand AI hallucination cases in U.S. courts. Bloomberg Law’s own database tracks 330+ similar filings. Courts have already begun sanctioning attorneys in multiple jurisdictions, and the precedents are solidifying across the country.

In Canada, the Alberta Court of Appeal’s decision in Reddy v. Saroya confirmed that attorneys bear ultimate responsibility for AI-hallucinated citations even when the work was produced by a third-party contractor. In Ontario, a Law Society discipline panel found that unverified reliance on AI tools like Grok could factor into costs decisions and suspension proceedings.

This is becoming a compliance category. Not a technology curiosity — a liability exposure.

For a full case-by-case breakdown of every major AI hallucination incident in U.S. courts this year, see our companion piece: The 2026 Legal AI Reckoning.

Why Policies Alone Are Not Enough: The Compliance Analysis

The S&C incident illustrates a failure mode that compliance professionals recognize well: the gap between written policy and operational behavior.

The firm had policies. The firm had training. The firm had procedures. None of it prevented the filing.

Here is why, and what compliance teams in any regulated industry need to internalize:

1. AI Makes Bad Work Look Finished

This is the core risk. AI-generated output doesn’t look like a rough draft. It doesn’t have obvious gaps or spelling errors that signal “check this carefully.” It looks complete, confident, and professionally formatted. Legal citations written by an AI look exactly like legal citations written by a human — until you verify them against the actual source.

Standard editorial review processes were not designed for this failure mode. Reviewers trained to catch typos and logical gaps are not automatically equipped to catch fabricated case law that reads correctly.

2. Automation Complacency Is Real and Predictable

As AI tools automate more steps in a workflow, human reviewers enter the process later — and with less context. The instinct to verify erodes gradually. This is not a moral failing; it is a documented cognitive pattern. When output appears complete and professional, verification behavior changes.

Compliance programs that install AI tools without adjusting review protocols for this specific risk are creating the conditions for exactly this outcome.

3. Emergency Conditions Compress Review Cycles

The S&C filing was an emergency motion. Time pressure is one of the most reliable predictors of compliance failures across every industry. When deadlines are compressed, review steps get shortened or skipped. AI tools that accelerate drafting do not automatically accelerate verification — and in emergency contexts, the verification step is the one most likely to be cut.

Any organization using AI in time-sensitive compliance or legal workflows must have explicit protocols for minimum verification steps that cannot be waived under deadline pressure.

4. Policy Awareness ≠ Policy Adherence

Two training modules and an office manual entry are a starting point, not a control. For AI governance specifically, compliance programs need to consider:

  • Workflow-embedded controls — verification checkpoints built into the document submission process, not just policy documentation
  • Spot audit programs for AI-assisted work product
  • AI citation verification tooling — tools like BriefCatch’s RealityCheck exist specifically because manual verification is insufficient at scale
  • Clear accountability chains — who is the named responsible reviewer for any AI-assisted filing?

The Broader Regulatory Trajectory

The S&C incident is occurring against a backdrop of rapidly tightening expectations. Courts are not sympathetic to “the AI did it” explanations. Judges have sanctioned attorneys in multiple jurisdictions. Several federal courts now require affirmative disclosure when AI tools were used in drafting court submissions.

For compliance professionals outside the legal industry, the trajectory is clear: regulators across sectors are moving toward explicit AI use accountability standards. The legal industry is simply the most visible early battleground because court filings are public record.

Financial services, healthcare, insurance, and government contractors should treat this incident as a forward-looking signal, not a legal-industry-specific curiosity.

What Compliance Programs Should Do Now

The Sullivan & Cromwell incident is a useful forcing function for compliance teams in any organization deploying AI in regulated work. The minimum viable response:

1. Audit your current AI use policies against actual workflows. Where is AI being used today? Does your policy actually cover those use cases? Are verification requirements operationally feasible in the contexts where AI is deployed?

2. Distinguish AI drafting controls from AI output verification controls. These are different problems requiring different solutions. Many organizations have policies about who can use AI tools; far fewer have controls specifically designed to detect hallucinated or fabricated content in AI output.

3. Treat emergency and compressed-timeline work as elevated AI risk. Build in explicit minimum verification requirements that apply regardless of deadline pressure.

4. Implement tooling, not just training. For document-heavy workflows, evaluate purpose-built AI verification tools. Manual verification of AI output at scale is not a reliable control.

5. Establish disclosure and escalation protocols. When AI-assisted work product contains errors that reach external parties, who is notified, how quickly, and through what process? S&C handled their disclosure well — most organizations do not have a practiced process for this.

6. Brief leadership on this incident. The “we advise OpenAI on ethical AI and still filed fake case law” narrative is a useful internal alignment tool. If it can happen there, it can happen anywhere. That is not hyperbole. It is the accurate lesson.

Final Thought

Sullivan & Cromwell did the right things after discovery: they corrected the record, apologized publicly and directly, remediated comprehensively, and Dietderich signed the mea culpa himself rather than pointing at a junior associate.

But the incident still happened. Policies didn’t prevent it. Training didn’t prevent it. Prestige didn’t prevent it. The AI governance program the firm markets to other organizations didn’t prevent it.

The lesson for every compliance function is not that AI cannot be used — it absolutely can, and the productivity benefits are real. The lesson is that AI changes the failure mode profile of knowledge work in ways that existing review processes were not designed to catch. Closing that gap requires more than a policy update and a training module.

It requires treating AI output verification as a first-class compliance control — with the same rigor, tooling, and accountability structure that organizations apply to any other high-stakes operational risk.

Sources: Bloomberg Law, Reuters, Above the Law, Canadian Lawyer, PYMNTS, Original Jurisdiction (David Lat), LawFuel, Sullivan & Cromwell AI Practice page. Case reference: Prince Global Holdings Limited and Paul Pretlove, Bankr. S.D.N.Y., 1:26-bk-10769.

This article is for informational purposes only and does not constitute legal advice.