Yesterday, Adam Feldman of Legalytics noted this observation—no great surprise, yet it is helpful to have a tangible demonstration:
This piece investigates how courts have responded once lawyers discovered, or should have realized, that AI-related inaccuracies had crept into the record. The analysis began with Damien Charlotin’s publicly accessible AI Hallucination Cases Database and independently evaluated 100 of the most recent cases within the dataset that involved U.S. matters and were resolved. A tighter analytical subset consisted of cases with a meaningful AI connection, a final judicial outcome, and enough information to assess counsel’s conduct after discovery.
The findings point to a clear, practical distinction. The severity of the initial error remains central: nearly every high-materiality incident led to a serious consequence. Yet post-discovery behavior also mattered. Serious consequences occurred substantially more often when lawyers offered misleading explanations, deflected blame, repeated the misconduct, or introduced additional inaccuracies after being notified. Prompt admission, by contrast, correlated with a markedly lower incidence of serious consequences.
Candor cannot erase significant harm to a client or to the judicial process. It can, however, prevent an already grave mistake from becoming a separate issue of professional responsibility….
And some more (read Feldman’s full post for even more):
The data suggests that the initial response should be investigative rather than defensive. Counsel should preserve the prompts, outputs, research histories, drafts, and communications tied to the filing; determine who drafted, reviewed, and signed it; and examine every authority, quotation, and factual assertion produced through the same workflow.
The correction ought to address the entire problem rather than only the error identified by opposing counsel. That may require withdrawing or replacing the filing, notifying the court and opposing counsel, and accurately explaining how the material entered the record. Describing fabricated authorities as typographical errors, or assigning responsibility primarily to a subordinate, client, vendor, or software product, increases risk.
Remedial measures are most persuasive when they are concrete and voluntary. Courts have credited steps such as mandatory review of primary sources, supervisory approval, AI-use policies, training, audits of related filings, and reimbursement of unnecessary expenses. Once counsel receives credible notice, every subsequent filing, explanation, and omission becomes part of the court’s assessment of professional judgment and candor….
