The firm's own AI training materials said it plainly: "trust nothing and verify everything." Someone didn't. On April 9, Sullivan & Cromwell, one of the most prestigious law firms in the United States, filed a motion in the Southern District of New York containing AI-generated legal citations that ranged from badly misquoted to entirely fictitious. At least 28 citations were erroneous, opposing counsel later found. Some cases referenced simply did not exist.
The filing came in a Chapter 15 bankruptcy case with unusually high stakes. Sullivan & Cromwell represents court-appointed liquidators from the British Virgin Islands pursuing claims against Prince Group and its owner, Chen Zhi. U.S. authorities have alleged that Chen directed scam compounds in Southeast Asia that victimised people worldwide and sought to recover billions in linked cryptocurrency. Chen was detained in Cambodia earlier this year and repatriated to China. This was not a routine commercial dispute.
Andrew Dietderich, the partner who heads the firm's restructuring practice, wrote directly to Judge Martin Glenn to take responsibility. The motion had misquoted language from the U.S. Bankruptcy Code and attributed to cited cases statements those cases never made. In one instance, a cited decision turned out to refer to an entirely different case in a different circuit. The firm withdrew the original motion and filed a corrected version with annotations marking where the AI had gone wrong.
Boies Schiller Flexner, acting for Prince Group and Chen, caught the errors and raised them with the court. They also argued the correction came too late: by the time Sullivan & Cromwell filed the revised motion, the defendants had already submitted their objections based on the flawed original. They asked the court to adjourn a scheduled hearing and hold a status conference. The downstream damage from bad citations, it turned out, didn't stop when the citations were withdrawn.
What makes the incident sting harder is the gap between stated policy and actual practice. Sullivan & Cromwell told the court it maintains rigorous AI protocols: mandatory training modules before any lawyer is granted access to generative tools, with completion tracked and verified. The training, the firm said, "repeatedly emphasises the risk of AI hallucinations, including the fabrication of case citations, misinterpretation of authorities, and inaccurate quotations." And yet the document that reached a federal judge contained exactly those failures. Dietderich acknowledged that the firm's oversight mechanisms had "malfunctioned," allowing defective content to advance unchecked.
The firm typically charges around $3,000 per hour. That figure has become a shorthand in coverage of this story, and it's not hard to see why. The hallucination problem in AI legal tools is not new, and its victims have generally been smaller practices under more financial pressure to cut corners. The suggestion that elite firms with rigorous policies and premium billing rates are equally exposed lands differently. It shifts the conversation from "reckless adoption" toward something more structural: that the verification burden AI places on legal professionals is real and persistent, regardless of how seriously a firm takes it in principle.
Courts in the U.K. have seen similar cases, with a senior High Court judge warning last year that lawyers submitting AI-generated fictions risk contempt proceedings. Deloitte, in an unrelated incident, issued a partial refund to the Australian federal government after an AI-assisted report contained factual errors. The pattern is consistent: AI tools accelerate drafting and research, and the productivity gains are real; the failures, when they occur, land in the most consequential documents at the most consequential moments. For legal work especially, that's the core tension the profession hasn't resolved.
Sullivan & Cromwell says it has launched an internal probe and implemented remedial measures. The corrected filing is now on the record. But the original one is too, and with it a clear illustration of what it looks like when a firm's protocols and its practice diverge, even briefly, under the pressure of a complex, fast-moving case.