What happened in the LeDoux v. Outliers AI citation case?

On July 24, 2026, Judge Tiffany M. Cartwright of the Western District of Washington fined attorney Jocelyn Stewart $3,000 under Rule 11. Her filings had dozens of inaccurate citations, including a Washington Supreme Court case that does not exist, and the court found her conduct tantamount to bad faith.

How much have courts fined lawyers for AI hallucinations?

In the orders we track, from $1,500 to $15,500 per lawyer, plus fee awards, double costs, State Bar reports and referrals. ComplexDiscovery counted at least $145,000 in sanctions in the first quarter of 2026. No court publishes an official total.

Does AI hallucination liability only apply to lawyers?

Court sanctions fall on lawyers, but clients pay too. In Douglas v. Deutsche Bank the bank's appellate brief was stricken, and in Couvrette v. Wisnovsky the clients' claims were dismissed. Any team that sends unchecked AI citations to clients or regulators carries a similar risk without a judge to catch it.

What should a small team's AI verification policy include?

A named person opens the primary source for every citation, statistic and quote an AI tool produced, logs who checked what and when, and applies the same check to corrections. Nothing AI-drafted leaves the team without that log. The copy-paste policy is in this article.

AI Hallucination Sanctions Tracker 2026: Fines by Case

Image: Unsplash, used under the Unsplash License.

Seven orders, one pattern. A lawyer filed citations that a generative AI tool made up, nobody opened the sources, and a judge did. The table lists each sanction with its docket number, checked against the order itself on October 20, 2026. Two of the orders are from late 2025; we keep them because later courts measure their fines against them.

The tracker

Case	Court	Sanction	Order date	What the court found
Douglas v. Deutsche Bank National Trust Co.	D.C. Court of Appeals (No. 24-CV-1099)	The bank's appellate brief stricken; no fine	Sep 3, 2026	Four cited authorities did not exist. The attorney said she used Google's generative AI search tool to find cases and did not verify them
Southland Homes v. Lam	Cal. Superior Court, Orange County, Appellate Division (JAD26-04)	$2,500 and $1,500 on two attorneys; order sent to the State Bar	Aug 4, 2026	A writ petition rested on fabricated citations and legal propositions
LeDoux v. Outliers, Inc.	W.D. Wash. (3:24-cv-05808)	$3,000; every later brief must certify that citations were verified	Jul 24, 2026	Dozens of inaccurate citations across at least five filings, including a Washington Supreme Court case that does not exist. The "corrected" motion swapped in real quotations that changed the meaning
Doiban v. OLCC	Or. Ct. App. (A181982, 347 Or App 742)	$10,000	Mar 18, 2026	At least 15 fabricated case citations and at least nine quotations that appear nowhere in Oregon case law. Counsel said AI only drafted an outline and staff searched Westlaw and Lexis
Whiting v. City of Athens	6th Cir. (Nos. 24-5918/5919, 25-5424)	$15,000 each for two attorneys, plus the other side's appeal fees and double costs	Mar 13, 2026	Over two dozen fake citations and misrepresentations of fact. The court asked whether generative AI wrote the briefs; the lawyers did not answer
Couvrette v. Wisnovsky	D. Or. (1:21-cv-00157)	$15,500 on lead counsel; briefs stricken; the clients' claims dismissed with prejudice; co-counsel later ordered to pay $14,205.66 of the defense fees	Dec 12, 2025 (fee order Mar 23, 2026)	Fifteen nonexistent cases and eight fabricated quotations across three summary judgment briefs in a family winery dispute
United States v. McGee	S.D. Ala. (1:24-cr-112)	$5,000 to the CJA fund; public reprimand; referral on removal from the court-appointed counsel panel	Oct 10, 2025	A motion drafted with an AI program contained fabricated citations the lawyer did not check

The amounts are what each court ordered, not what has been paid. ComplexDiscovery's April 2026 tally counted at least $145,000 in sanctions for fabricated citations in the first quarter of 2026; it counts Couvrette in that quarter, although the sanctions order is dated December 12, 2025. No court keeps an official total, so read any running sum, ours included, as a floor.

Douglas is the row to show a client. The fine there was zero, and the cost landed on the bank's case: its brief was stricken.

Brass scales and a gavel on a light wooden desk, representing a verification checklist before AI output goes external

Image: Pexels, used under the Pexels License.

What changed since the last review

Previous update: September 2026. This review: October 20, 2026, with every row re-read on CourtListener or govinfo.gov.

Added Southland Homes v. Lam (Orange County appellate division, August 2026) and Douglas v. Deutsche Bank (D.C. Court of Appeals, September 3, 2026).
LeDoux: on August 18, 2026 the court granted summary judgment and dismissed the remaining claims with prejudice (Dkt. 269). The sanction order stands on its own.
Couvrette: the fee order of March 23, 2026 put part of the defense fees on co-counsel; on June 12, 2026 the court ordered the lead lawyer to show cause why he should not be held in contempt for not complying with its order; on July 22, 2026 the plaintiffs filed a Ninth Circuit appeal (No. 26-4712).
Fixed four errors from the earlier version: the Couvrette and McGee dates, the Whiting docket numbers, and the claim that all five original cases were 2026 sanctions for proven AI use. Details in the correction note below.

The case behind the July update

Joann LeDoux, an Army nurse, sued the maker of nootropic supplement kits she said were adulterated with amphetamines. Her lawyer, Jocelyn Stewart, filed papers in late 2025 with dozens of bad citations, which the court said "appeared 'hallucinated' by a generative artificial intelligence tool." The February 4, 2026 show cause order listed them, including Berg v. Chevrolet Motor Div., 84 Wn.2d 102, "which does not exist."

The fix made it worse. The corrected Daubert motion "replaced some false quotations with real ones but, in doing so, completely changed their meaning," and the court called her explanations "incomplete at best and misleading at worst." Stewart conceded the Rule 11 violation and said she no longer uses AI tools to generate citations. Judge Cartwright still found the conduct "tantamount to bad faith" (order, Dkt. 265).

That is the lesson we'd pin to the wall: the correction needs the same check as the draft.

How to use this tracker

Use the table as the answer when someone on your team says "the AI checked it." In each row a signed filing carried citations nobody had checked against the source. In five of the seven, the lawyer admitted using an AI tool or the court found one had been used; in Doiban counsel said AI drafted only an outline, and in Whiting the question went unanswered.

Lawyers are simply the group whose mistakes end up in a public order. A marketing claim citing a study that doesn't say what the AI said, a regulatory filing with an invented statistic, a client report quoting a source nobody opened: each carries the same flaw, and the person who catches it may be a regulator or a customer instead of a judge. Our FTC AI marketing claims checklist covers the advertising side.

Small teams are more exposed, not less. A large firm has paralegals and review layers that catch some errors by accident. A five-person team that asks an AI tool for a client report and ships it the same afternoon has nothing between the draft and the client unless someone builds the check in on purpose.

A hand holding a pen over a printed document, representing the verification log a named reviewer signs before AI-drafted work leaves the team

Image: Pexels, used under the Pexels License.

A copy-paste AI output verification policy

Paste this into your AI acceptable use policy or your client-deliverable checklist, and change the roles to match your team.

AI OUTPUT VERIFICATION POLICY

1. Any AI-generated citation, statistic, quote, case reference, or named
   source must be independently verified against its primary source before
   it appears in anything that leaves the organization: filings, client
   deliverables, marketing claims, regulatory submissions, or public
   statements.

2. "Independently verified" means a named human opened the primary source
   itself, not another AI summary of it, and confirmed the fact matches.

3. Every external-facing document produced with AI assistance must carry a
   verification log: who checked it, what was checked, and the date.

4. AI-drafted corrections or revisions require the same verification as the
   original draft. A correction is not exempt just because it fixes a
   flagged error; it can introduce new ones.

5. No AI-generated content moves from internal draft to external delivery
   without passing through step 1. This applies regardless of deadline
   pressure.

6. Any AI tool used to produce external-facing content must be listed in
   the team's approved-tools register. Unlisted tools require sign-off
   before use on anything client-facing.

It costs the time it takes to open a source document. Every order in the table describes a filing where nobody did.

Five questions before anything AI-drafted goes out

Did a person open the primary source, or only read the AI's summary of it? These are different activities.
Does the verification log show who checked this, and when? "We're pretty sure someone looked" is a no.
Was the correction checked as carefully as the original? LeDoux shows the correction can do the damage.
Is the tool that produced this on your approved list? If nobody knows which tool drafted a paragraph, nobody can flag the tool.
Would this survive the other side, a regulator or a client running the citations through a search engine? In McGee, opposing counsel found the fake citations before the judge did.

None of these need a legal department or a budget. They need someone with the authority to say "we're not sending this yet."

What we could not confirm

We could not confirm that generative AI produced the Whiting briefs: the court asked, and the lawyers did not respond. We could not see whether the Couvrette contempt question or the Ninth Circuit appeal has been resolved, or whether any fine in the table has been paid. The table lists only orders we could read in full; sealed and unpublished orders are not in it.

Correction (October 20, 2026): The earlier version dated the Couvrette v. Wisnovsky sanction to the first quarter of 2026; the order is dated December 12, 2025, and only the fee order (March 23, 2026) falls in 2026. It dated United States v. McGee October 14, 2025; the order was filed October 10, 2025. It gave the Whiting docket as No. 25-5425; the sanctions opinion is in Nos. 24-5918/5919 and 25-5424. It also called all five cases 2026 sanctions for AI output and said the Whiting lawyers relied on AI summaries; the Whiting court never established AI use. It said Stewart's corrections introduced new false quotations without acknowledging the problem; the order says the corrections swapped in real quotations that changed the meaning, and she later conceded the Rule 11 violation.