Attorneys for MyPillow founder Mike Lindell filed a court brief containing roughly 30 AI-hallucinated citations, were fined, and then one attorney was sanctioned a second time over another flawed citation. The repeat sanction is what makes the case a governance problem: a financial penalty did not change the outcome because a penalty is not a control.
In 2025, attorneys for Mike Lindell, the founder of MyPillow, filed a motion in a Denver defamation case that contained roughly 30 defective citations produced with the help of AI. Judge Nina Wang fined each of the two attorneys 3,000 dollars in July 2025 over the filing, as reported by The Colorado Sun.
The court found the brief was riddled with citations to cases that did not exist or did not support the propositions cited, and described the attorneys' account of how it happened as troubling. The coverage by Law and Crime notes the defense that a wrong draft had been filed by mistake.
The pattern then repeated. In May 2026, one of the attorneys, Christopher Kachouroff, and his firm were sanctioned again, this time 5,000 dollars, over another flawed citation, according to Colorado Politics. A second sanction is the detail that makes this case worth studying. The first penalty did not change the behaviour.
What failed: the governance gap
The action that caused the harm was the filing. AI-assisted research produced citations that were not real, and nothing stood between that draft and the court docket. There was no checkpoint that the filing had to clear, and no independent party who had to sign before the document went out.
The wrong-draft explanation points at the same gap from a different angle. If the account is taken at face value, an unverified version reached the court because the firm had no enforced step that distinguished a checked filing from an unchecked one. The person who drafted the brief was also the person who filed it. There was no segregation between producing the work and authorising its submission.
The repeat sanction is the part that no amount of professional intent closed. A 3,000 dollar fine is a strong incentive to be careful. It is not a control. It relies on the same humans, under the same time pressure, remembering to verify every citation every time. The second filing shows that intent alone did not hold.
How MakerChecker changes the outcome
MakerChecker governs the action an automated actor is allowed to take. Filing a document with a court is exactly the class of consequential, hard-to-reverse action it is built to intercept.
Model the filing as a skill that is denied by default until a verification step has completed, and that routes to an approval gate held by an independent reviewer. The drafting agent can prepare the brief. It cannot submit it. The submission parks and waits for a named human who is not the drafter to sign:
skill: court.file
risk_tier: high
preconditions:
verify_citations: required # ungranted until the check passes
gate:
approvals_required: 1 # n-of-m named reviewers
forbid_requester: true # the drafter cannot approve the filing
The deny-by-default precondition is the part that addresses the repeat. The
court.file skill is not available to the role until the citation-verification
step has run and passed. The checkpoint is not a habit that someone has to
remember. It is a structural condition on the action, applied the same way on the
first filing and the second.
Segregation of duties through forbid_requester separates the drafter from the
approver. The party that produced the brief is barred from being the one who
authorises its filing. An independent reviewer has to sign, which puts a named,
accountable human between the AI-assisted draft and the docket.
Every proposal, every denial, and every approval is written to the tamper-evident, Ed25519-signed, hash-chained audit, verifiable offline. The record shows which version of the document was approved and by whom. That directly addresses the wrong-draft claim, because the question of what was reviewed and signed becomes a signed fact rather than a recollection.
What MakerChecker would not fix
MakerChecker would not stop an AI tool from producing fabricated citations. It governs actions, not the content a model generates. If the research tool invents a case, the invented case still appears in the draft. The control sits on the filing, not on the writing.
It also cannot force a reviewer to actually read the brief. If the approver signs off without checking the citations, a defective filing can still reach the court, and the harm still occurs. The gate enforces that an accountable, independent human signed and that the signed version is recorded. It does not enforce that the review was diligent or that the citations are real. That part stays with the humans, which is why a verification precondition and a separate reviewer reduce the odds of a repeat without eliminating it.
See the configuration: examples/rogue-ai/mypillow-ai-brief-fake-citations-repeat