A federal court in Oregon ordered a lawyer to pay $109,700 over AI-invented filings, and first-quarter sanctions alone topped $145,000. Courts have stopped teaching this lesson for free.
A federal court in Oregon ordered a lawyer to pay $109,700 in sanctions and costs for filing AI-generated errors, a figure the Washington State Bar's NWSidebar blog reported in March as a possible record. It did not stay remarkable for long. By ComplexDiscovery's tally, US courts imposed at least $145,000 in sanctions for AI-fabricated citations in the first quarter of 2026 alone.
Three years ago this was one famous cautionary tale. Now it is a line item.
| When | Where | The bill |
| 2023 | Mata v. Avianca (S.D.N.Y.), the first famous one | $5,000 |
| Reported Feb-Mar 2026 | Federal court, District of Oregon | $109,700 in sanctions and costs |
| 2026 | Whiting v. City of Athens (Sixth Circuit) | Full appellee fees, double costs, and $15,000 in punitive sanctions from each of two attorneys |
| June 2026 | New York appellate case | $10,500 combined: $8,000 from the attorney, $2,500 from the firm |
The scale behind the headlines: Damien Charlotin, a research fellow at HEC Paris, maintains the most widely cited database of AI-hallucinated content submitted to courts. It passed 1,200 cataloged cases worldwide by early 2026. NPR covered the stacking penalties in April. The American Bar Association's Litigation News summed up the era in four words: fake cases, real sanctions.
1. What changed between $5,000 and $109,700
In 2023, courts treated Mata v. Avianca as a teaching moment. The sanction was modest and the opinion read like a warning to the profession. The warning did not work, so the posture changed. Look at Whiting: the Sixth Circuit found more than two dozen fake citations and misstatements of fact, then reached for fee-shifting, doubled costs, and punitive sanctions on top. Courts are no longer pricing these as mistakes. They are pricing them as conduct.
And the failure keeps recurring for an uncomfortable reason: the tools got better. A hallucinated brief in 2023 often read slightly off. A hallucinated brief in 2026 reads clean, formats correctly, and cites cases whose names sound exactly right. Fluency is a formatting property. It carries no information about whether the case exists. Deadline pressure plus polished output is how experienced attorneys keep signing filings they never checked.
2. Every one of these is a workflow failure
Here is the take, and it is the whole article: none of the sanctions cases are technology failures. The model did what models do. It produced plausible text. The failure happened after generation, in the gap where verification was supposed to live. Which is good news, because your firm controls the workflow completely.
Source-locked drafting. The model may cite only from documents you loaded into it: the record, the authorities you pulled, your own research memos. Drafting from the model's memory of case law is how invented citations are born.
A cite-check gate with a name on it. Before anything is filed, a human pulls and reads every cited authority. Every one. The check gets logged with the checker's name and the date. If that sounds tedious, compare it to explaining $109,700 to your partners.
A tool policy by matter type. What is permitted on a research memo differs from what is permitted on a signed filing. Write the tiers down, and cover disclosure obligations where courts require them.
Train the actual failure mode. Teach associates that model confidence signals mean nothing. The sanctioned briefs did not look risky. They looked finished. That is precisely the property that makes unverified AI output dangerous in a profession built on citations.
The signature rule. Whoever signs owns every word and every cite. Rule 11 does not have an AI exception, and no judge in the database found "the tool wrote it" persuasive.
3. The mantra that survives contact with a courtroom
AI drafts. You verify. Firms that internalize those four words get the upside of drafting speed with none of the docket humiliation. Firms that skip the second half are funding the sanctions tracker one six-figure order at a time. The technology is allowed in nearly every courtroom in America. Unverified technology is what keeps ending up in the database.
Final thought
The remarkable thing about the 1,200-case database is how few of those matters involved bad lawyers. Mostly they involve normal lawyers, under normal deadline pressure, extending trust to output that had done nothing to earn it. Verification is not a tax on AI adoption. It is the price of the signature, same as it always was. The only new part is how the drafts get made.
AI, Properly is the Ayvede field report: AI news read plainly, with a takeaway you can use. Ayvede builds verification workflows and AI guardrails for law firms and other high-trust practices. Advise. Vet. Deliver. Ask a question in writing.