Sullivan & Cromwell Admits AI Hallucinations in Filing

Sullivan & Cromwell, one of the United States' most prominent law firms, acknowledged that a bankruptcy court filing in the Prince Group case contained multiple AI-generated errors, including fabricated case citations and misquoted sections of the bankruptcy code. The firm said those mistakes, totaling around three dozen items, were identified by opposing counsel at Boies Schiller Flexner and that it filed a corrected motion and apologized to Judge Martin Glenn. "We deeply regret that this has occurred," said Andrew Dietderich, co-head of Sullivan & Cromwell's global restructuring group.

The exact AI system used was not disclosed in the firm's letter and public reports. Key failure modes reported were:

• •fabricated citations and non-existent legal authorities
• •inaccurate quotations and misstatements of statutory language
• •improperly summarized conclusions from real cases

Generative models are increasingly used to accelerate legal research and drafting, but these tools commonly produce plausible-sounding but incorrect outputs, known as hallucinations. The case exposes three operational failure points: improper adherence to usage policies, ineffective secondary review, and overreliance on tool outputs without source verification. Sullivan & Cromwell told the court it has "comprehensive policies and training requirements governing the use of AI tools," but the firm admitted those controls were bypassed in this instance.

The incident triggers immediate professional-risk questions. Lawyers have ethical obligations to ensure accuracy in court filings; automated workflows that reduce human verification can lead to malpractice exposure, sanctions, and reputational damage. For legal technologists and ML engineers building or integrating generative systems into legal workflows, the takeaways are clear: validate provenance, implement conservative retrieval-augmented pipelines, enforce mandatory human-in-the-loop checks, and log system outputs for audit.

This is not the first reported legal-sector hallucination, but it is notable because it involves an elite firm with significant influence over clients and legal tech adoption patterns. The misstep is likely to accelerate stricter internal governance at other firms, spur vendor differentiation on hallucination mitigation features, and encourage courts and bar associations to clarify guidance on AI-assisted drafting.

Expect law firms to tighten mandatory verification steps, increase provenance and citation tooling, and press vendors for stronger retrieval-grounded architectures and explainability features. Regulators and bar committees may issue sharper guidance on permissible AI use in pleadings and discovery.

If you integrate generative models in legal or regulated workflows, require: a defensible audit trail, verified-source retrieval (avoid purely generative-only citation), enforced secondary review with checklists, and conservative prompt designs that force source links rather than freeform assertions.