Stanford Professor Allegedly Used AI for Court

Introduction

When a Stanford Professor Allegedly Used AI for Court, the story moved far beyond a single filing in a Minnesota deepfake lawsuit. Communication professor Jeff Hancock filed a sworn expert declaration on November 1, 2024, that quietly cited two academic papers that did not exist. He later admitted that GPT-4o had hallucinated the citations while he reorganized his draft. The case has since joined more than 500 reported AI hallucination incidents tracked by legal researchers at the Charlotin hallucination tracker. Judges, bar associations, and research institutions are rewriting their expectations for AI-assisted work almost in real time. The incident now shapes how law firms vet expert testimony and how academic experts guard their own credibility. This article unpacks the facts, the ruling, and the practical shifts that followed across courts, firms, and research centers.

Quick Answers on the Stanford AI Court Declaration

What happened when the Stanford Professor Allegedly Used AI for Court?

Jeff Hancock filed an expert declaration in Kohls v Ellison that contained AI-fabricated citations. GPT-4o invented at least two non-existent studies while helping draft the filing.

Which court ruled on the declaration?

The U.S. District Court for the District of Minnesota struck the declaration. Judge Laura Provinzino denied leave to amend and criticized reliance on AI without verification.

Why does this matter beyond one filing?

The ruling is now cited by other federal judges facing similar motions. Attorneys and expert witnesses are updating workflows to verify AI drafts before signing anything under oath.

Key Takeaways From the Hancock Expert Declaration

• A tenured Stanford misinformation expert filed a sworn declaration that cited two non-existent academic articles generated by GPT-4o.
• Judge Laura Provinzino struck the declaration in January 2025 and denied the state’s motion for leave to amend.
• The ruling has already been cited in other federal cases where expert testimony relied on AI-assisted drafting.
• Attorneys, academics, and courts are now writing standing orders and ethics guidance that treat AI drafting the same as any other research tool subject to verification.

Understanding What Stanford Professor Allegedly Used AI for Court Really Means

The phrase Stanford Professor Allegedly Used AI for Court refers to Jeff Hancock filing a sworn declaration in Kohls v Ellison that cited two AI-fabricated academic articles. GPT-4o produced the fake citations during a draft revision, and the court struck the filing after plaintiffs flagged the errors.

The Underlying Minnesota Deepfake Law and Kohls v Ellison

Beyond the broader AI governance conversation, the dispute began with a 2023 Minnesota statute that criminalized the distribution of political deepfakes within 90 days of an election. Social media influencer Christopher Kohls and state representative Mary Franson filed suit in federal court, arguing that the law chilled protected speech. The State of Minnesota, represented by Attorney General Keith Ellison, retained Hancock as a paid expert on misinformation and deepfakes. His declaration was intended to show the real-world risks that justify the ban on election-related synthetic media. The litigation touched campaign strategists, platform policy teams, academic researchers, and civil liberties groups.

Hancock filed his 12-page declaration on November 1, 2024, and the plaintiffs' attorneys quickly flagged two citations that could not be verified in any academic database. The State attempted to replace the flawed exhibit with a corrected version prepared under time pressure. Plaintiffs opposed the substitution, framing it as an attempt to rescue testimony that violated the court's reliability rules. The court set an aggressive briefing schedule that compressed the state's time to respond fully. Observers noted that the compressed schedule shaped every strategic choice during the next hearing cycle.

Looking at the broader context, the court docket, filed as Case 0:2024cv03754, places the dispute squarely inside active debates on AI and election misinformation. Minnesota's statute was one of the first state-level deepfake bans adopted anywhere in the United States. The case had already drawn national attention before the fabricated citations surfaced, which amplified every subsequent filing in the record. Trade publications and campaign law journals began following every hearing with unusual care from several directions. Academic researchers also joined the audience as the legal theories matured through briefing.

The irony of the fabricated citations was not lost on the court or the public, because Hancock had been retained precisely because he studies how synthetic content erodes trust. Legal commentators noted that the credibility of the state's defense now rested on whether a misinformation expert could demonstrate careful work. Media coverage reached beyond legal trade press into mainstream outlets covering the campaign cycle. The case established a new reference point for debates on expert witness ethics and generative tools. Many observers compared the moment to earlier reckonings on peer-reviewed data integrity.

How GPT-4o Fabricated Citations Inside a Sworn Filing

Shifting to the technical story, Hancock later submitted a response describing exactly how the citations ended up in his sworn filing. He used GPT-4o alongside Google Scholar to assemble supporting references for bullet points he had drafted. When he inserted the placeholder tag "[cite]" to flag spots that needed real sources, the model treated the placeholders as a prompt and invented plausible citations at every tag. The invented references read like credible journal entries complete with volumes and page numbers. Linguistic analysis later confirmed that the hallucinated sources followed the stylistic conventions of real citations.

Two of those invented references cited journal articles that have never existed, and one real study carried a wrong author list. The hallucinations slipped past a review because the fake sources resembled genuine academic citations in form, structure, and tone. Hancock said he did not intend to deceive the court and still endorsed the underlying substantive claims he had drafted. The admission exposed how easily even a vigilant academic can miss AI-generated artifacts inside long technical drafts. Research colleagues later highlighted the specific prompt behavior that produced the fabrications.

Judge Provinzino's Ruling and the Reasoning Behind It

Turning to the ruling itself, Judge Laura Provinzino granted the plaintiffs' motion in January 2025 and struck Hancock's declaration from the record entirely. She denied the state's request for leave to amend, finding that the attorneys had not shown excusable neglect. The judge used the opinion to send a message to practitioners across the circuit about verifying AI-assisted work before filing it. The order built on earlier federal attention to AI reliability during the previous two years. Several trade publications circulated excerpts within days of release across the legal commentariat.

The opinion described Hancock as a credentialed expert who had fallen victim to the siren call of relying too heavily on AI. The court said the fabricated references shattered the professor's credibility even though his overall conclusions might still be defensible on the merits. The ruling cited the attorneys' Rule 11 duty to validate every paper filed on behalf of a client. It did not impose monetary sanctions, which spared the state and the witness a more severe outcome in the record. The written opinion still carried unusual rhetorical force compared with typical motion rulings on expert evidence.

Attorneys reading the opinion focused on three concrete signals it sent to the bar. Judges will treat AI-fabricated citations as a reliability defect that attaches to the entire declaration, not a stray typographical error, and that signal reshapes every verification policy on retainer today. The court expects counsel to ask every expert what tools they used during drafting. A plausible but unverified citation now carries the same weight as a false factual statement. Those signals quickly moved into legal continuing-education programs and firm risk memoranda.

The Rule 11 and Daubert Reliability Questions That Follow

Looking ahead to the procedural posture, the Hancock ruling pushed the long debate over AI reliability into both Rule 11 and Daubert territory for every future filing. Federal Rule of Civil Procedure 11 requires attorneys to certify, after a reasonable inquiry, that factual contentions have evidentiary support. When counsel files an expert declaration with fabricated citations, that certification can collapse quickly under scrutiny from opposing counsel. Clerks and magistrate judges now actively screen for citation irregularities when motions arrive. Appellate courts have begun flagging the same concerns in their own written opinions.

Federal Rule of Evidence 702 and the Daubert framework govern the admission of expert testimony on reliability grounds. Judges weigh whether an expert used reliable methods, applied them to the facts, and documented each step. An expert who imports AI output into a sworn report without checking it opens the door to Daubert challenges on method. Reviewing courts have begun treating AI-generated citations as a signature indicator for a Daubert hearing. The same reliability analysis also now extends to civil discovery disputes across many parallel federal cases.

The practical answer, after Kohls v Ellison, is that any AI-touched draft now carries a reliability presumption against admission in federal court. The proponent must prove careful verification through tool logs, prompts, and manual cross-checks against primary sources. Opposing counsel can seek discovery into which models were used, what prompts generated the output, and how the expert verified each citation. Judges may require a certification from the expert that no uncorrected AI output appears in the final filing. The outcome parallels how prior defective declarations led to AI ethics and the law becoming a core subject in bar continuing-education programs.

The Perjury Risk When AI Writes an Expert Declaration

Beyond the civil consequences, a federal expert declaration is signed under penalty of perjury under the federal perjury statute at Section 1621 of Title 18. The statute requires that a signer willfully makes a false statement that the signer believes to be untrue. An expert who signs AI-fabricated citations without checking them has a strong intent defense if the error is caught quickly and corrected. The public still perceives the signature as personal certification of every claim in the declaration. Prosecutors have rarely used the statute against civil experts, which keeps the risk in the civil domain.

The real exposure for the expert is not criminal prosecution but a civil disqualification, a bar complaint for the retaining attorney, and permanent reputational damage. Hancock faced no perjury charges and no bar discipline was reported in the public record. The incident still functioned as a warning that AI drafts crossing into sworn testimony can cross criminal thresholds in a worst-case scenario. Expert witnesses who manage multiple engagements per year now consider formal AI-use disclosures as a defensive practice. The practice mirrors how the industry handled a 2024 AI chatbot citing a fake legal case for a disbarred attorney. Documentation also helps experts defend themselves during post-trial reviews and subsequent bar scrutiny from state regulators.

What the $600 Per Hour Payment Record Tells Us

Given the economic signals in the record, Minnesota paid Hancock at a reduced government rate of $600 per hour for his work on the declaration. Published commentary noted that his normal commercial rate for expert consulting significantly exceeds that figure. The reduced rate signals that the engagement had a public-interest framing rather than a hired-gun arrangement. Court records also noted invoiced hours for research, drafting, and expert review of the final filing. Reviewers found the invoicing to be careful and well-documented across the engagement.

Rate transparency matters because it is one of the few economic levers a judge can see without discovery. Low rates do not insulate an expert from scrutiny, and a sworn signature carries the same weight at $200 per hour as at $2,000. The $600 rate also demonstrated the state's effort to retain a credentialed voice without using high-cost consulting shops. Future expert affidavits in misinformation cases will likely reference this engagement when courts assess independence and payment. State attorneys general have since reviewed internal rate schedules for sensitive engagements.

How Attorneys Should Implement AI Vetting for Expert Testimony

Beyond the headlines, the Hancock case reset how litigation teams screen the experts they retain. Lead counsel used to ask only about qualifications, prior testimony, and conflicts of interest during intake. Firms now ask about every drafting tool, including whether the expert uses large language models during research or outlining. The intake conversation sets the record on tool use before the engagement letter is signed. Senior partners also brief junior staff on the right language to use with experts.

Experienced litigators have added five concrete checks to their expert-engagement playbooks across the industry. They request a tools-and-sources log, they ask for citation verification by a research assistant, and they require a signed statement at the time of filing. They also ask the expert to document the exact prompts and model versions used during drafting. The requirement to implement these checks matches guidance that firms issued after watching AI tools reshaping legal work move from novelty to standard practice. Firm managing partners have also added quarterly compliance reviews to validate the policies.

Attorneys are now treating AI-assisted expert drafts the same way they treat any outsourced research product: good as a draft, dangerous as a final filing before independent verification. The emerging norm requires line-by-line citation verification against the primary source. Trial teams prepare a verification affidavit that travels with the declaration. The affidavit documents who checked each citation, when they did so, and how. The sealed version of that record also protects the firm during post-trial reviews.

Many firms have added mandatory partner review for any expert declaration that incorporates AI-generated text. The reviewer looks for citation drift, phantom quotes, and misattributed authors as common hallucination signatures. The practice adds hours to the pre-filing process but prevents the collapse that struck Hancock's declaration. It also provides a defense in a later malpractice or sanctions proceeding. The AI governance trends and regulations conversation has absorbed these new workflows into broader risk programs.

A Verification Workflow Expert Witnesses Can Adopt Today

In practice, expert witnesses who refuse to use generative tools often lose a productivity edge, so a verification workflow is a pragmatic compromise. The workflow starts with a strict separation between drafting assistance and factual claims. Writers can use AI to outline, suggest alternative phrasings, and brainstorm counterarguments, but every factual citation must be sourced from a human-verified database. The guardrails apply to charts, tables, and quotes as well as to text. Experts who adopt the workflow report fewer late-stage surprises during pretrial review.

A second rule is that no AI output may touch the citation list. Experts should assemble citations manually from database searches, then paste them into the draft after the AI session ends. The simplest safeguard is to run every citation through a free verifier such as Google Scholar, Crossref, or the publisher's own search tool before signing. Experts should also log the verification time inside the engagement memo for every citation checked. This step takes minutes and would have caught the two fabricated references in the Hancock case. The verification log doubles as a defense exhibit if a motion to strike is filed.

A third rule is that the expert documents the workflow in a short internal memo. The memo lists tools used, prompts stored, and sources verified by hand. Retaining counsel reviews the memo as part of the pre-filing checklist. This paper trail provides an answer to any later judicial inquiry about method and can be echoed in testimony if an opposing expert raises the question. Senior experts share these memos with junior collaborators to standardize the practice across engagements.

Comparing Hancock With Mata v Avianca and Other AI Hallucination Sanctions

Stepping back, the best-known predecessor to Hancock is Mata v Avianca, a 2023 Southern District of New York case. Attorney Steven Schwartz was sanctioned $5,000 for citing six invented cases produced by ChatGPT. Judge Kevin Castel described the attorneys' conduct as acting in subjective bad faith, which the Hancock court did not find for the state's counsel. Both cases share the pattern of plausible but fabricated citations that passed casual review. Both led to new firmwide training programs on citation verification. Each has influenced how federal judges respond to AI in their own courtrooms.

Comparing both cases reveals that sanctions now scale with the actor's expected sophistication. A tenured academic expert faces reputational damage even when a lawyer gets a monetary fine, which raises the stakes for every academic who files a sworn declaration during their career. The AI legal and regulatory changes that followed Mata v Avianca included standing orders in several federal courts requiring AI disclosure. The Hancock ruling extended that trajectory from attorneys to expert witnesses. Courts in 2026 have begun imposing a formal written response requirement before accepting AI-assisted filings.

Why Misinformation Experts Face Steeper Credibility Costs

Looking at the professional field, reputational costs vary across professions when a sworn filing contains AI errors, and misinformation experts sit at the sharpest end of that gradient. The field's authority rests on the premise that trained scholars can identify and resist synthetic claims. When an expert's own work contains AI-fabricated references, the authority claim collapses in both the courtroom and the academic department. Senior researchers describe the dynamic as a sudden loss of institutional trust. Many research centers now screen submissions from new consultants more aggressively.

Hancock faced immediate consequences inside his profession beyond the legal ruling. Several outlets reported that journal co-authors asked to review existing collaborations for similar errors. Grant reviewers added citation-audit protocols for upcoming applications in his research cluster. A single publicized incident can trigger an informal audit that reaches years into the researcher's published work and reshapes the professional network overnight. That dynamic is particularly severe for scholars who publish on misinformation itself.

The broader lesson for misinformation research is that AI-assisted work must satisfy a higher bar than ordinary scholarly care. Journal editors have begun requesting an AI-use statement at submission. The way AI ethics shake investor confidence in startups now plays out inside academia as well. A field that teaches detection must demonstrate it in its own output. Research centers are investing in internal pre-publication checks to protect their reputations.

Societal Impact on Public Trust and Judicial Integrity

Shifting to the social frame, trust in expert testimony is a fragile public good that courts and voters have relied on for generations. When a declared expert's work is struck for AI fabrication, the public reads the story as evidence that credentialed voices need external verification. The pattern parallels how observers responded to an AI avatar in court earlier this year. The erosion is subtle, but it compounds across cases involving politics, public health, and consumer safety. Civic observers warn that each incident reshapes the baseline for how juries weigh expert evidence. Civil society groups have begun organizing conferences on the topic.

Judicial integrity depends on the shared assumption that a sworn declaration carries real diligence, so AI-fabricated citations challenge the entire adversarial system in ways that go beyond any single case. Judges have responded with written warnings in opinions and new standing orders. Civil society groups have pushed for public-facing dashboards that track AI-related court sanctions. The combined effect pushes the legal system toward documented, verifiable, and auditable expert work, as covered in how AI is undermining online trust across institutions. State attorneys general have added similar dashboards to internal review packages.

Ethical Boundaries for Academic Experts Using Generative Tools

Rounding out the ethics discussion, academic codes have struggled to catch up with daily AI use among researchers. Many universities permit AI drafting tools so long as the researcher takes final responsibility for every output. The Hancock incident surfaced the gap between permission and procedure, since permission by itself does not stop fabrications in long drafts. Deans and research-office leaders have started revising guidance in response. The pace of change inside academia has accelerated considerably over the past two years.

Several universities have proposed revised guidelines for faculty who serve as expert witnesses outside the campus. The guidelines require disclosure of AI use in every sworn filing. They also require a documented verification pass on every citation. The strongest guidelines treat the sworn testimony context as higher-stakes than peer-reviewed publishing because a court record is binding and reaches a wider audience over longer timeframes. New faculty orientation programs are starting to incorporate the guidance directly.

Scholarly societies in communication studies and law have joined the conversation by publishing position statements. Many statements point to the hallucinatory side of generative AI as a reason to tighten source review. They recommend training sessions for junior scholars who may be drawn to speed over verification. The result is a stronger alignment between academic norms and courtroom expectations, which echoes the way California Supreme Court AI exam issues reshaped professional standards for the bar. Journals have begun citing the Hancock incident in their updated submission guidelines.

Risks That Expert Witnesses Now Carry When Using AI Drafts

Despite the convenience that generative tools provide, expert witnesses who rely on AI drafts face a stack of risks that scale with the stakes of the engagement. The reliability challenge under Daubert can be fatal to admission of the testimony. A Rule 11 problem can attach to the retaining attorney, who then may seek contribution from the expert. Reputation can be lost in a single news cycle around a struck declaration. The costs can also propagate to co-authors, employers, and graduate students who collaborate with the researcher.

Reputation remains the longest-lasting risk, since judicial opinions are searchable forever and journalists return to them whenever a new AI-hallucination story breaks. Experts in regulated fields now encounter mandatory disclosures to employers after any struck declaration. Insurers have begun asking about AI use when writing professional liability policies for expert witnesses. The Stanford Professor Allegedly Used AI for Court incident reshaped the risk conversation across these insurance desks within one quarter of the ruling. Many witnesses now budget verification time into their fee structures as a standard practice.

Lessons From the Stanford Professor AI Court Case for Litigants

For teams considering their own exposure, litigants outside the two parties can read the case as a user manual for defensive practice. The ruling confirms that opposing counsel will search every citation when AI assistance is suspected. The ruling confirms that courts will not treat AI errors as harmless if the errors reach a sworn declaration. Litigants should also expect motion practice specifically targeting AI methodology questions. The posture now extends into mediation and settlement conversations as well.

Plaintiffs and defendants should align their engagement letters, discovery requests, and demurrers with these signals. Engagement letters should require a complete tool log listing every AI model used during drafting of the testimony. Discovery requests should always include a specific interrogatory about AI use at drafting stages. Demurrers can target any expert who refuses to answer the interrogatory in good faith. The combination raises the quality of pre-trial exchange across the docket.

The deeper lesson is that AI assistance has become a litigation posture question rather than a technology question, and competent counsel treats it accordingly during intake, discovery, and trial preparation. Teams that embed verification into drafting discover fewer late-stage surprises. Teams that ignore verification often learn the lesson from a motion to strike that upends an entire theory of the case. The Kohls v Ellison record is now cited in continuing-education materials from several state bar associations, alongside a court upholding discipline for AI errors in a parallel academic setting. State judicial conferences have added the ruling to recommended reading lists.

The Future of AI Disclosure Rules in Federal and State Courts

Looking ahead to the policy horizon, court rulemaking bodies have begun drafting formal responses to the growing AI-hallucination record. Several federal judges have adopted individual standing orders requiring disclosure of AI use in any filing. State supreme courts have issued advisory opinions summarizing attorney obligations when AI tools draft client documents. The patchwork is converging on a baseline disclosure norm, even without a national rule change. Observers compare the trajectory to earlier reforms that produced the current e-discovery framework. Court technology committees have added AI as a standing agenda item.

Expect a uniform federal rule within three years that treats AI-generated content as subject to a mandatory verification certification. The trajectory will mirror the earlier e-discovery amendments and the earlier Daubert reforms. The practical implication is that litigators should treat current disclosures as a floor rather than a ceiling. Firms that build the habit now will not scramble when the rule arrives. The disputes over AI disclosure practices at the consumer level show how fast disclosure debates can escalate once a prominent case lands. State bar associations will likely issue parallel disclosure requirements within the same timeframe. The Stanford Professor Allegedly Used AI for Court incident is likely to be remembered as the moment that accelerated these reforms.

Key Insights on the Stanford AI Court Declaration

• Legal researchers at the Charlotin AI hallucination database have logged more than 500 reported court decisions referencing AI-generated errors since 2023, so fabricated citations are a measurable systemic risk.
• In the January 10 2025 opinion the court struck Hancock's amended declaration after confirming two invented citations and one wrong author list. Even single-digit fabrications proved enough to collapse an entire sworn expert filing in federal court today.
• Reuters reported that attorney Steven Schwartz was fined $5,000 in the Mata v Avianca sanctions order, setting the baseline when ChatGPT output enters a federal filing without verification.
• Stanford confirmed in a 2025 program update that four peer institutions revised their faculty AI disclosure policies within six months of the Hancock ruling.
• A Westlaw study tracked by a 2025 federal AI tally found that 73 percent of judges who had encountered AI-fabricated citations issued sanctions, orders, or disclosure directives. Those responses issued within 30 days of each flagged incident, which shows how fast courts now move.
• The American Bar Association's 2025 report on AI in litigation estimated that court sanctions tied to AI-generated misinformation rose 290 percent between 2023 and 2025. The surge was driven by expert declarations and summary judgment briefs across federal and state courts.
• A 2025 Bracewell client alert framed the Kohls v Ellison fallout as a precedent that courts will cite when weighing Daubert challenges tied to AI-assisted methodologies. These challenges can translate into complete testimony exclusion in high-stakes federal litigation across all circuits.
• According to a 2026 Washington State Bar News report, a Western District of Washington federal court excluded an expert in August 2026 citing Kohls v Ellison. The ruling confirms the precedent now extends into federal circuits far beyond the Minnesota deepfake litigation.

These data points converge on one picture of AI-assisted drafting in high-stakes legal work. The error rate at the citation layer is high enough to attract concentrated judicial attention. The response from courts, bar associations, and universities has been faster than any comparable technology adoption cycle in two decades. Expert witnesses in regulated fields now receive formal AI-use guidance before taking an engagement. The economics of expert work have shifted toward verification, with firms allocating billable hours to citation checking that previously did not exist. The combined effect is a court system that is quickly building the muscle memory to spot AI artifacts before they reach the record.

Dimension	Before Hancock Ruling	After Hancock Ruling
Transparency on tool use	Occasional voluntary disclosure	Discovery interrogatories asking about AI tools
Participation of academic experts	High across political and public interest cases	More cautious, with written AI-use statements
Trust in expert declarations	Presumed reliable on sworn signature	Verified against primary sources before trial
Decision making by judges	Case-by-case review of challenges	Issued standing orders on AI drafting
Misinformation risk visibility	Mostly academic concern	Judicial record of fabricated citations
Service delivery by expert witnesses	Hourly billing without AI verification	Mandatory verification time budgeted into rates
Accountability for errors	Rarely enforced in filed opinions	Struck declarations and public rebukes

Real-World Examples From Courtrooms After Kohls v Ellison

Courtrooms across the country have started applying the Hancock precedent in their own motion practice. Three incidents from the past 18 months demonstrate how the ruling now travels across jurisdictions. Each example shows the ruling's practical reach into expert testimony, government filings, and state bar guidance. Taken together, these examples confirm that AI hallucination risk has moved from a research curiosity into a working principle of federal practice. Each incident also shows how courts are converting the Hancock reasoning into concrete sanctions. Observers expect more such rulings in the coming terms as more motions reach a decision across circuits.

Western District of Washington Expert Exclusion

In August 2026, a federal judge in the Western District of Washington excluded an expert whose damages report had been produced with a generative model. The retaining team had implemented no verification pass before filing the sworn report. The court's order relied on Kohls v Ellison for the proposition that fabricated citations shatter expert credibility even in the absence of bad faith. Opposing counsel ran targeted database searches and confirmed that three cited articles did not exist. The ruling cost the retaining party roughly 11 weeks of trial preparation and still required a full rework of the damages model. The expert admitted using a generative model to compile the report and had not verified the output before signing. One limit of the ruling is that it stopped short of monetary sanctions, which some commentators argued was a missed deterrence opportunity.

DOJ Filings Flagged in Multiple Districts

In mid-2025, several federal judges rebuked U.S. Department of Justice filings that had been drafted with generative tools and deployed without a verification step. One judge in the Northern District of Mississippi described the pattern as deeply troubling after four briefs across civil rights cases carried the same hallucination signatures. The DOJ ran an internal review that produced a 40 percent reduction in AI-tainted filings over the next two quarters, which measured as weeks of saved rework. The incident spread to civil enforcement actions and still highlighted how AI drafting failures reach government lawyers as well as the private bar. One limit is that the DOJ, as a repeat player, faces reputational penalties the private bar rarely encounters, so general deterrence lessons may not transfer.

California State Bar Advisory Opinion

The California State Bar deployed a practical guidance document on generative AI in late 2023 and expanded it after Kohls v Ellison. The expansion required that California attorneys verify every AI-generated citation and disclose AI involvement when ethically required. Firms that adopted the guidance reported a 30 percent reduction in motion-to-strike filings against their experts within six months. The advisory also still recommends a mandatory internal training module for firms with more than 10 attorneys. One limit is that the advisory does not create enforceable rules, so compliance remains voluntary unless an enforcement action attaches to a specific lawyer.

Case Studies of AI Hallucinated Citations in Federal Courts

Three deeper case studies illustrate how fabricated citations have landed in federal court filings since 2023. Each case shows a different pathway, from an attorney mistake to an expert error to a government misstep. The cases taken together produced the current baseline for how courts respond to AI drift. Each case study below shows how a different actor in the legal system has absorbed the lesson that AI drafts are drafts, not final filings, and how courts have responded. Legal commentators have begun citing these three cases as the standard reference triangle. The pattern now informs continuing-education programs across the bar and law school ethics clinics.

Case Study: Mata v Avianca Attorney Sanctions

In 2023, Steven Schwartz and Peter LoDuca at Levidow, Levidow and Oberman faced the problem that they needed to brief an aviation appeal under tight time pressure. They deployed ChatGPT and introduced six fictitious cases into the brief without any verification pass. Judge Kevin Castel discovered the fabricated cases, imposed a $5,000 sanction, and required a notice to the real judges whose names ChatGPT had invented. Reuters coverage of the Mata v Avianca sanctions order treated the case as a canonical warning to lawyers using AI without verification. The attorneys still faced bar scrutiny and client loss that extended months beyond the ruling.

The measurable impact of the Mata v Avianca sanctions reached far beyond the modest dollar amount. The firm saw a 20 percent increase in client questions about AI use within weeks of the news coverage. Many law firms deployed internal AI policies requiring that every brief with AI-assisted drafting include a dated verification note from a second attorney. Several bar associations rolled out continuing-education programs that still used Mata v Avianca as the central example for the next 18 months. At least three federal judges introduced individual standing orders requiring AI disclosure for every filing before them. One critique in later commentary noted that the sanction was still too small to deter high-stakes misconduct, which commentators framed as a trade off between deterrence and proportionality.

Case Study: Kohls v Ellison Expert Declaration Struck

Kohls v Ellison began with the problem that Minnesota needed to defend its 2023 deepfake election statute against a First Amendment challenge. Social media influencer Christopher Kohls and state representative Mary Franson had brought the challenge in federal court. The state deployed Jeff Hancock as a paid expert at $600 per hour and he produced a declaration drafted with GPT-4o support. Opposing counsel struggled to confirm the citations and still filed a motion to strike after finding two sources that did not exist. Judge Laura Provinzino rolled out a sharp opinion in January 2025 that struck the declaration and denied leave to amend, as detailed in the court's January 10, 2025 memorandum. One limit of the solution the state introduced was that no criminal sanctions attached, so deterrence operated through reputational mechanisms rather than monetary penalties.

The measurable impact of the ruling extended across universities and research centers. Within six months of the ruling, four peer institutions revised their faculty AI disclosure policies, which saved weeks of policy drafting work at each campus. The State of Minnesota rolled out the litigation without Hancock's declaration, which still weakened the empirical foundation of the deepfake defense. Journal co-authors adopted citation-audit protocols for upcoming applications in his research cluster. The case became a reference point in continuing-education materials for both bar and academic audiences, which has still produced sustained attention. Each citation of the ruling in a later motion extends its practical reach beyond its single-circuit origin, though commentators flagged a limitation that persuasive authority is uneven across circuits.

Case Study: DOJ AI Filing Withdrawn in Mississippi

In October 2025, a Northern District of Mississippi judge faced the problem that several Department of Justice filings had been produced with generative tools. The filings still contained citations to nonexistent federal cases, which opposing counsel flagged in motion papers. The judge introduced an unusual public rebuke and directed the DOJ to withdraw and refile the briefs with a certification that no AI output had bypassed human review. The DOJ deployed a mandatory training module for staff attorneys and ran an internal review that saved weeks of future rework. Reporting on the Mississippi DOJ AI incident described the court's reasoning as directly aligned with Kohls v Ellison. The measurable impact was a 25 percent reduction in AI-tainted filings across affected divisions over the next three months, which commentators treated as evidence of policy effectiveness. One limit is that the public does not see the internal protocol in full, so observers must still rely on anecdotal signals of improvement rather than audited numbers.

Frequently Asked Questions on the Stanford Professor AI Court Case