Kohls v. Ellison: a court throws out a Stanford misinformation expert's declaration after GPT-4o invented its citations

The problem

Christopher Kohls and Mary Franson challenged a Minnesota statute restricting AI-generated “deepfakes” in elections, and moved for a preliminary injunction on First Amendment grounds (source). To oppose that motion, Minnesota Attorney General Keith Ellison submitted two expert declarations, one from University of Washington professor Jevin West and one from Stanford professor Jeff Hancock, Director of the Stanford Social Media Lab and a credentialed expert on AI and misinformation (source). The plaintiffs moved to exclude both, and specifically alleged that Professor Hancock had included fabricated material in his declaration (source).

What was built

After the plaintiffs’ motion, the Attorney General’s office contacted Hancock, who admitted that his declaration inadvertently included citations to two non-existent academic articles and incorrectly cited the authors of a third article (source). As the court recorded, these errors originated from Professor Hancock’s use of GPT-4o, a generative AI tool, in drafting his declaration, and “GPT-4o provided Professor Hancock with fake citations to academic articles, which Professor Hancock failed to verify before including them in his declaration” (source). The court underlined the irony: Professor Hancock, “a credentialed expert on the dangers of AI and misinformation, has fallen victim to the siren call of relying too heavily on AI” (source), and the court added that this was “in a case that revolves around the dangers of AI, no less” (source).

The outcome

The citations that did not exist: two fabricated articles, plus a third mis-attributed. Judge Laura M. Provinzino held that the court “cannot accept false statements … in an expert’s declaration submitted under penalty of perjury,” and that Hancock’s “citation to fake, AI-generated sources in his declaration … shatters his credibility with this Court” (source).

What the AI cost the state: the whole declaration. The court ordered that the plaintiffs’ motion to exclude was “GRANTED IN PART AND DENIED IN PART,” granted as to the Hancock Declaration and denied as to the West Declaration (source). Because it declined to consider Hancock’s testimony in deciding the preliminary-injunction motion, the court also held that the Attorney General’s motion for leave to file an amended expert declaration was “DENIED AS MOOT,” so the state could not repair the AI-tainted evidence (source).

The penalty: no monetary sanction. This order imposed no fine; the AI failure cost the party its expert evidence, not money (source). The order observed in general terms that “[t]he consequences of citing fake, AI-generated sources for attorneys and litigants are steep,” citing Mata v. Avianca and other cases (source).

How this was verified

Checked live on 2026-08-14. Every fact here is stated by an independent federal court in a signed, publicly filed order (Document 46, D. Minn., filed 2025-01-10), live-fetched from CourtListener RECAP and byte-tied to a Wayback capture (CDX digest KAG3WUT5VJN2HU6D6LZ3MD4E6PDYR3DN, archived 2025-01-14), so the facts need no client-confirmation call to be true. The court’s language and the GPT-4o-fabricated citations are independently corroborated by two secondaries re-read live on 2026-08-14: Reason / The Volokh Conspiracy and PPC Land, which reproduce the same quotations verbatim. There is no dollar figure to confirm; this is an evidentiary exclusion, and the order imposes no monetary sanction. The honest limit: the court later denied the preliminary injunction and the case proceeded to the Eighth Circuit (No. 25-1300), a downstream posture this file does not track beyond noting it.

Related case files

The same core failure, an officer of the court relying on generative AI and not verifying the output, recurs across the registry with different stakes. In Gauthier v. Goodyear the AI-hallucinated citations came from the attorney’s own brief and drew a $2,000 sanction plus a mandatory generative-AI CLE, the fine that Kohls conspicuously lacks. Lacey v. State Farm shows the same unverified-AI error inside a sophisticated firm, evidence that it is not a novice’s mistake but a discipline failure. Wadsworth v. Walmart is the Rule 11 version of the story, where the price of fake citations was a monetary sanction rather than lost evidence.

Sources

Tier 1 = the court stating its own ruling in a signed order; Tier 2 = independent press naming the parties. Each source was checked live on 2026-08-14.

1. U.S. District Court, D. Minnesota, “Order” in Kohls v. Ellison, No. 24-cv-3754 (LMP/DLM), Document 46, 2025-01-10 (Tier 1, the adjudicating court; live-fetched from CourtListener RECAP, byte-tied to Wayback, CDX digest KAG3WUT5VJN2HU6D6LZ3MD4E6PDYR3DN). https://www.courtlistener.com/docket/69206960/kohls-v-ellison/
2. Reason / The Volokh Conspiracy (Eugene Volokh), “Misinformation Expert’s ‘Citation to Fake, AI-Generated Sources in His Declaration … Shatters His Credibility with This Court’,” 2025-01-10 (Tier 2, independent legal commentary quoting the order at length). https://reason.com/volokh/2025/01/10/misinformation-experts-citation-to-fake-ai-generated-sources-in-his-declaration-shatters-his-credibility-with-this-court/
3. PPC Land, “Minnesota court rejects expert testimony tainted by AI-generated citations,” 2025-01-26 (Tier 2, independent press). https://ppc.land/minnesota-court-rejects-expert-testimony-tainted-by-ai-generated-citations/