Sullivan & Cromwell Apologizes for AI-Generated Errors in Filing

Sullivan & Cromwell, one of Wall Street's premier law firms, notified U.S. Bankruptcy Judge Martin Glenn that an emergency filing in the Prince Global Holdings Chapter 15 case contained inaccurate citations and other errors described as AI "hallucinations." In an April 18 letter, partner Andrew Dietderich apologized, saying firm policies governing AI use were not followed and the secondary review process failed to catch the fabricated or misquoted authorities. Rival counsel at Boies Schiller Flexner identified the errors, prompting a corrected filing and direct apologies to opposing counsel.

The firm characterized the problem as instances where AI generated non-existent case citations, misquoted authorities, and produced incorrect article titles and numbers. The corrected submission contained a redline showing entire sentences and citation strings replaced or removed. Key operational failures Dietderich cited were human: deviation from documented AI-use policies and an ineffective secondary review step that did not detect the fabricated outputs. Courts and trackers now document more than 330 filings that include AI-related inaccuracies, which has led judges to impose sanctions in other matters.

This is a concrete example of downstream risk when generative models are used as research or drafting assistants without rigorous validation. Legal citations are high-integrity artifacts: a fabricated case citation is not a harmless stylistic error, it is a factual inaccuracy that can mislead courts, trigger sanctions, and damage firm reputation. The incident illustrates two failure modes relevant across industries: model hallucination producing plausible but false strings, and process failure where human reviewers rely on AI outputs without sufficient cross-checks.

Practitioners should treat model-generated factual assertions as untrusted inputs. Practical controls include mandatory citation verification workflows, automated citation-checking tools that cross-reference authoritative databases, stricter access controls to generative assistants for legal research, and auditable logs of which outputs informed a filing. Firms should also update training to require explicit attestation that citations were verified before filing.

This episode follows a growing set of cases where judges and opposing counsel have found AI-sourced errors in filings. The legal sector moved quickly to adopt generative tools for speed and efficiency, but the pattern of errors is prompting sharper scrutiny from courts and peers. For AI developers and platform teams, the legal domain highlights the limits of current factual grounding and retrieval approaches for LLMs. For model evaluation, the event argues for benchmark tasks that measure hallucination rates on structured, citation-heavy domains and for integrating retrieval-augmented systems with verifiable sources.

Whether the court imposes sanctions in this case, and whether Sullivan & Cromwell announces concrete process or tooling changes, will be important. Also monitor vendor responses: adoption of citation-validation APIs, improved provenance features, and tighter enterprise controls will follow as buyers demand safer workflows. Finally, expect regulators and bar associations to accelerate guidance on ethical AI use in legal practice.

The incident is neither a novel technical surprise nor an isolated PR mistake. It is a practical warning: generative AI can produce authoritative-looking but false artifacts that, when embedded in high-stakes outputs like court filings, create legal, ethical, and reputational risk. Teams must pair model use with robust verification and audit processes before deploying LLMs in regulated or evidentiary contexts.