What a fabricated citation costs
Not a projection or a survey of concerns. These are decided outcomes: money paid, findings entered, contracts partly refunded.
The case list
A public database maintained by Damien Charlotin, a research fellow at HEC Paris, logs decisions in which a court or tribunal found, or clearly implied, that a party relied on AI-fabricated material. The inclusion standard is strict: a judicial finding, not a news report.1
It passed 2,000 entries worldwide by September 2026, with more than thirteen hundred in the United States. Over one stretch that year it was adding close to eight entries a day, faster than the five or six a day recorded a few months earlier. Self- represented litigants make up the larger share, but practicing lawyers account for hundreds of the entries.2
The financial outcomes climbed as the novelty defense stopped working. In March 2026 a federal court of appeals fined two lawyers $15,000 each, with fees and double costs on top, for briefs containing more than two dozen fabricated citations and misstatements of the record.3 The largest publicly reported sanction is the $110,000 in fees and penalties a federal district court in Oregon imposed in December 2025, after lawyers filed fifteen fabricated citations and eight invented quotations across three briefs.4 Bar suspensions connected to AI filings have followed.5
The direct cost is the smaller half. A sanctions order is a public document with names in it, and it is discoverable by every future opposing counsel and every client running a conflicts check.
A government asked for its money back
In October 2025 Deloitte agreed to refund part of a A$439,000 engagement with Australia's Department of Employment and Workplace Relations. The report, which reviewed the IT system used to automate penalties in the welfare compliance regime, contained a fabricated quotation from a federal court judgment and references to academic papers that do not exist.1
The refund was over A$97,000. A revised version was published, this time disclosing that a generative AI system had been used in its preparation.2
Two details make this the more instructive case. The errors were found by an outside academic reading the report, not by the firm's review process. And the subject matter was a government's automated penalty system, which is to say a report about the risks of automated decision-making failed through automated decision-making.
You are bound by what your system says
In February 2024 the British Columbia Civil Resolution Tribunal found Air Canada liable for negligent misrepresentation. Its website chatbot had told a passenger he could apply retroactively for a bereavement fare. The airline's actual published policy said otherwise. The chatbot had invented a policy that suited the question.1
The airline argued that the chatbot was a separate entity responsible for its own statements. The tribunal rejected that, holding the airline responsible for all information on its website regardless of which component produced it, and awarded damages, interest and fees.2
The award was small. The principle is not, and it generalises past airlines: an organization is answerable for what its systems tell people, and telling someone a policy exists tends to make it exist.
What these three have in common
In each case a competent organization had a review process, and in each case the review did not catch it. That is the pattern worth taking seriously. A fabrication produced by a capable model is fluent, specific and shaped exactly like the thing the reader was hoping to find. It survives review precisely because it is good.
Which is why a checking step is a weak place to put the whole defense. That argument is the subject of the next piece, and the obligations arriving to meet it are in the one after.