Sullivan & Cromwell, one of the most prestigious law firms in the country, filed a bankruptcy motion in federal court in Manhattan that contained citations to cases that do not exist, misquoted passages from the US bankruptcy code, and inaccurately summarized real cases. Judge Martin Glenn received an apology letter from Andrew Dietderich, the cohead of the firm's restructuring group. The errors were caught by opposing counsel at Boies Schiller Flexner during routine review of the filing. Sources told the Financial Times that S&C has an enterprise license for OpenAI's ChatGPT, though the firm has not officially confirmed which model produced the bad citations.
The pattern is by now depressingly familiar. Morgan & Morgan, Levidow Levidow & Oberman, and a string of smaller firms have already been sanctioned for filing AI-hallucinated citations. What makes this one notable is the firm: S&C is not a solo practitioner Googling for shortcuts. It is a top-five Wall Street firm with billable rates north of two thousand dollars an hour, an enterprise AI deployment, and the resources to build whatever verification workflow it wants. The hallucinated citations made it through anyway. The failure mode is the same as every prior incident: a lawyer asked the model to find supporting cases, the model produced something that looked like supporting cases, and the lawyer signed the brief without independently verifying that the cases existed.
The technical reality, for anyone still confused: ChatGPT does not have a verified database of court rulings. It generates text that pattern-matches what legal citations usually look like. When the right citation exists in the training data, it can reproduce it accurately. When the right citation does not exist or the model is uncertain, it will produce a plausible-looking citation anyway, with a real-sounding case name, a real-sounding reporter volume, and a real-sounding page number. There is no internal flag that says "I made this up." The model does not know it does not know. Enterprise licenses do not change this; they change billing and data retention, not the fundamental tendency of language models to confabulate when asked to retrieve facts.
The fix is operational, not technological. Any firm using AI for legal research needs a hard rule that every citation gets independently verified against Westlaw, LexisNexis, or PACER before the brief is signed, and the verification step has to be auditable. Tools that ground retrieval in actual case databases — there are several aimed at this market now — make the workflow cheaper, but they do not absolve the lawyer of the duty to check. The lesson from S&C is not that ChatGPT is uniquely bad at law; it is that even sophisticated organizations with enterprise tooling will skip verification if the workflow does not force them to. Judges are running out of patience. Bar referrals and monetary sanctions are arriving more often. The firms that will survive the transition are the ones that treat the model output as a draft to be checked, not as a finished work product to be signed.