AI Hallucination Cases — the running record of court decisions

The Institute's reading

This is the closest thing the profession has to a live tracker of AI failure in court, and it is worth consulting rather than reading — a maintained record of decisions in which judges addressed filings containing fabricated citations, misquoted authority or invented holdings. It spans jurisdictions and court levels, and it includes lawyers at firms of every size, self-represented litigants, and on occasion filings prepared by counsel who then disclaimed responsibility.

Two things become apparent quickly. The first is the monotony of the fact pattern: a lawyer used a tool, did not verify what it produced, filed it, and was found out when opposing counsel or the court could not locate the authority. There is very little variation. Almost none of these involve an exotic failure or a subtle error — they involve nobody opening the case. The second is that courts have distinguished sharply between the AI use and the conduct after discovery. Firms that acknowledged the error promptly have generally fared far better than those that were evasive or attempted to defend the fabricated authority, which several attempted. The sanction that follows a cover-up is consistently worse than the one that follows the mistake.

For a managing partner the practical use of this record is internal. An abstract warning about verification is easy to nod at; a list of named lawyers in published decisions is not. It is also the clearest possible answer to a partner who believes their own seniority makes the checking step unnecessary, since the record contains a number of very senior people.

Key propositions

• The fact pattern is consistently the same: a tool was used, the output was not verified, the filing was made.
• The failures are not subtle — in most cases nobody opened the cited authority at all.
• Consequences have included monetary sanctions, referrals, stricken filings and findings recited in published opinions.
• Courts have treated the response after discovery as more serious than the original error.
• The record spans firm sizes and seniority levels, including experienced litigators.

In practice

• Verification of filed authority is the single highest-value control a litigation practice can implement.
• A firm needs a decided response to an incident before one occurs, because the record shows the response drives the outcome.
• This record is more effective internal training material than any policy document.

Where authorities disagree

Some argue the record overstates the problem — that a few hundred failures against millions of filings is a low rate, and that the cases cluster among the least sophisticated users. Others note it almost certainly undercounts, since it captures only failures caught and addressed in a written decision, missing everything found in review, settled quietly, or never noticed. Both are probably right. Neither changes what a firm should do, because the control that prevents the failure costs very little.

Related entries