Every reported sanction has the same shape: a tool produced something plausible, and nobody checked it before it was filed. The failure is procedural, not technological.
Start a conversation with the AI Adoption Concierge, already scoped to verification & quality control. Choose the question closest to yours, or describe your situation directly.
There is now a substantial public record of what happens when AI output reaches a court unchecked. A database maintained by legal researcher Damien Charlotin tracked roughly 1,490 decisions worldwide — more than a thousand of them in the United States — by May 2026, in which a party relied on AI-hallucinated material and a court responded. The penalties have escalated from four-figure fines to fifteen thousand dollars per attorney in a federal court of appeals, and to the first suspensions. What is striking about the record is how consistent it is: the tool behaved exactly as generative models do, and the firm had no step at which someone confirmed the citations existed. Verification is a workflow problem with a workflow answer.
Three separate problems. Checking citations is the narrow one; deciding who reviews what, and under which duty, is the durable one.
The narrow, bounded check that separates every sanctioned firm from every unsanctioned one.
investigateMatching depth of review to consequence, so checking does not consume the time the tool saved.
investigateExisting supervision duties turn out to be the most useful frame a firm already has.
investigateHow the Institute approaches verification — orientation and process design, not advice on your jurisdiction.
Larger than most firms assume and growing steadily. Damien Charlotin's public database of AI-hallucination decisions tracked roughly 1,490 worldwide, over a thousand of them in the United States, as of May 2026. The consequences have escalated over that period: from modest fines, to $15,000 per attorney in a federal appellate matter, to bar discipline. In April 2026 the Nebraska Supreme Court suspended an attorney whose appellate brief contained 57 defective citations out of 63, including twenty cases that did not exist. Because the record grows monthly, any figure needs a date attached.
Very few firms conclude that, and the ones that try tend to discover the tools are in use anyway. A prohibition that is not enforced produces the worst configuration: people using consumer tools, with client material, outside any workflow the firm can see or check. The more workable posture is a defined set of approved tools, a mandatory verification step for anything leaving the firm, and enough training that people understand what they are checking for. Prohibition also forfeits the productivity, which competitors are not forfeiting.
ABA Formal Opinion 512, issued in July 2024 as the ABA's first ethics guidance on generative AI, addresses competence, confidentiality, communication, candor toward the tribunal, supervisory responsibility, and fees. Supervision is the most practically useful of those for verification: firms already have a framework for reviewing work produced by someone who does not carry the ultimate responsibility, and AI output fits that frame closely. States have issued their own guidance and it is not uniform, so a firm should read its own jurisdiction rather than assume the ABA opinion governs.
Not if it is scoped to consequence. Confirming that cited authorities exist and stand for what a draft claims is a bounded task, and it is dramatically faster than the original research would have been. What does erase the saving is undifferentiated review — treating an internal research memo and a brief for filing identically. Firms that tier their review by where the work is going generally keep most of the gain; firms that either skip checking or check everything to the same depth tend not to.
Describe how AI is being used and by whom. The Institute will help you scope a workflow that fits.