AI-Generated Fake Cases: What Recent Sanctions Teach Lawyers
The legal AI risk that once looked like an emerging issue has become a professional responsibility problem. Courts are no longer asking only whether lawyers used AI. They are asking whether lawyers verified the work before filing, advising, or relying on it.
The professional-risk pattern is now clear
Lawyers have always had to verify their authorities. What changed is the speed and confidence with which generative AI can create plausible legal text that looks finished before it has been checked. A fake case citation can now arrive inside a polished argument, an edited brief, a draft email, a research note, or a copied passage from a tool that felt safer because it was marketed for legal work.
The recent incidents matter because they are not all the same. Some involve fabricated cases. Some involve real case names paired with wrong numbers or propositions. Some involve fake quotations from real cases. Some involve senior lawyers being held responsible for junior lawyers or law clerks. One important lesson follows: the risk is not simply "ChatGPT made up a case." The risk is an uncontrolled workflow in which legal authority becomes decorative instead of verified.
Recent incidents lawyers should know
In May 2026, Reuters reported that a federal magistrate judge in San Francisco sanctioned the managing partner of a California law firm after a junior lawyer filed an AI-assisted brief containing a false citation in Hill v. Workday. The supervising lawyer was admonished, fined $1,001, and required to complete training on supervising attorneys and ethical AI use. The court wrote that managing lawyers must take reasonable steps to ensure lawyers in the firm make ethical representations to the court and that, at minimum, a supervising lawyer should read pleadings and check citations for accuracy (1).
In March 2026, the U.S. Court of Appeals for the Sixth Circuit sanctioned two lawyers in Whiting v. City of Athens. Reuters reported that the court found more than two dozen fake citations and factual misrepresentations bearing hallmarks of AI hallucinations. The lawyers were ordered to reimburse the city for appeal work and to pay $15,000 each to the court (2).
In April 2026, the Los Angeles Times reported that the State Bar of California had filed disciplinary charges against attorneys accused of submitting AI-generated filings with nonexistent or irrelevant legal decisions. The same report noted that disciplinary measures had been approved against another California attorney who had submitted nonexistent and erroneous citations in a 2025 federal filing (3).
In November 2025, an Oregon federal judge declined to impose formal sanctions on Buchalter lawyers after a filing contained two fake AI-generated citations. The court accepted remedial measures, including a $5,000 donation to a legal aid organization, internal safeguards, and an offer to reimburse fees. That incident is useful for lawyers because the explanation was not "we asked AI to do legal research." One lawyer said he used AI to improve the writing of a document after doing his own research, and the tool inserted hallucinated citations that he failed to catch (4).
In September 2025, California's Second District Court of Appeal fined attorney Amir Mostafavi $10,000 after an opening brief included fabricated material. CalMatters reported that the opinion said 21 of 23 case quotations in the brief were fake, and the court published the decision as a warning that no filing should include citations the responsible attorney has not personally read and verified (5).
In July 2025, a federal judge in Colorado fined two attorneys representing Mike Lindell $3,000 each after a filing in a defamation case contained more than two dozen mistakes, including citations to cases that did not exist. NPR reported that the court found the lawyers had violated the certification obligation that legal claims be well grounded in law (6).
In June 2025, London's High Court warned that lawyers who cite nonexistent cases through misuse of AI may face contempt of court or, in the most serious cases, criminal consequences. The warning followed two cases in which written arguments referred to fake case law. The judge emphasized that public confidence and the administration of justice are at stake when AI is misused (7).
And in Utah, attorney Richard Bednar was sanctioned after a filing included fabricated legal authority from ChatGPT, including a nonexistent case called Royer v. Nelson. The Utah Court of Appeals stressed that AI may be a legal research tool, but every attorney has an ongoing duty to review and ensure the accuracy of court filings (8).
Why this keeps happening
These incidents are not only about technology. They are about workflow design, supervision, incentives, and review culture.
First, generative AI is optimized to complete patterns. In legal text, a citation is part of the pattern. If the model is allowed to draft freely, it may create an authority-shaped sentence before the workflow has confirmed that the authority exists and supports the proposition.
Second, lawyers often use AI under time pressure. The draft looks useful, the tone is professional, and the cited cases appear plausible. The more polished the answer, the easier it is to skip the slow act of reading each authority.
Third, some lawyers treat editing as lower risk than research. Recent incidents show that this is unsafe. A tool asked to improve style may still add, alter, or distort legal references unless the workflow prevents it.
Fourth, real citations can still be wrong. A case may exist but not support the proposition. A quote may be fabricated. A citation may combine a real case name with the wrong docket or jurisdiction. A source may be outdated, limited, or procedurally irrelevant. Responsible legal AI must guard against all of these failure modes, not only the obvious fake case.
The practical standard for lawyers
The practical standard is simple to state and hard to operationalize: no legal proposition should rely on an authority that has not been verified for existence, relevance, jurisdiction, currentness, and proposition support.
For court filings, that means personally reading the authorities relied on. For advice and internal research, it means keeping the path from claim to source visible enough that a lawyer can challenge it. For law firm leadership, it means supervising how AI is used by associates, trainees, clerks, paralegals, and outside vendors. Delegating the prompt does not delegate professional responsibility.
Disclosure rules vary by court and jurisdiction, but verification does not. Whether a filing requires an AI disclosure or not, the lawyer remains responsible for the paper submitted, the authority cited, and the representation made to the court or client (9).
How LexVera is designed to prevent this problem
LexVera was built around a conservative premise: legal AI should not be allowed to invent the legal support layer. A useful answer is not enough. The answer must be reviewable, source-grounded, and honest about uncertainty.
Our workflows begin with retrieved legal and matter sources rather than model memory. The system distinguishes public legal authority from client documents, firm knowledge, and user assumptions, so those materials are not flattened into a single undifferentiated answer.
For citation-sensitive work, LexVera uses a verified citation pool before final drafting. In plain terms, citations are not treated as decorative text that the model can freely compose. Final citations are tied back to source records and validation checks. If a citation cannot be verified, the safer behavior is to remove it, mark the limitation, or ask for review instead of presenting a confident but unsupported reference.
LexVera also checks more than whether a citation exists. Legal authority has to be usable for the point being made. That is why our reliability model includes jurisdiction, date, source type, procedural posture, and current-law signals. A real case that is outdated, limited, or not on point should not be promoted as strong support.
The platform is intentionally lawyer-facing. It shows the materials behind the answer, preserves review paths, and makes uncertainty visible. The goal is not to replace professional judgment. It is to reduce the chance that professional judgment is applied to a polished but unsupported draft.
What makes a legal AI workflow safer
- Source grounding: the answer should be built from retrieved materials, not from model memory alone.
- Verified citation handling: citations should come from controlled source records and be checked before final output.
- Proposition support: the cited source should actually support the sentence that cites it.
- Current-law awareness: the workflow should flag outdated, limited, interim, or uncertain authorities.
- Source separation: legal authority, client facts, firm knowledge, and assumptions should be labeled differently.
- Conservative failure behavior: when verification fails, the system should warn, qualify, or omit rather than improvise.
- Reviewable output: lawyers should be able to inspect sources, export work, and correct weak support quickly.
- Supervision fit: partners and team leaders should be able to set expectations for AI use across matters.
Questions lawyers should ask any AI vendor
Vendor diligence should move past generic claims about accuracy. Lawyers should ask operational questions that map to the incidents courts are now seeing.
- Can the model freely invent citations, or are citations generated from verified source records?
- What happens when the system cannot verify a citation?
- Does the workflow test whether the source supports the proposition, or only whether the source exists?
- Can it distinguish legal authority from uploaded client documents and prior firm work?
- Does it flag current-law uncertainty, jurisdiction mismatch, and procedural posture?
- Can users inspect the materials that shaped the answer?
- Can firm administrators define responsible-use rules for court-facing and client-facing work?
- Does the system preserve enough audit context to investigate mistakes?
Practical implication
The recent sanctions do not mean lawyers should avoid AI. They mean lawyers should reject uncontrolled AI. The professional problem is not that software helped draft a paragraph. It is that the legal support for that paragraph was not verified before it was used.
Legal AI should make lawyers faster at finding, testing, and explaining authority. It should not make it easier to file unsupported legal assertions. The most reliable systems slow down at the exact point where the legal profession has always required care: the point where a claim becomes a representation supported by authority.
The question is no longer whether AI can write like a lawyer. The question is whether the workflow forces every legal claim back through sources a lawyer can read, verify, and defend.
Resources and further reading
- Reuters: US judge says senior lawyers must pay for mistakes by subordinates using AI tools
- Reuters: US appeals court fines lawyers $30,000 in latest AI-related sanction
- Los Angeles Times: Attorneys used AI to write court filings, cited fake legal decisions, State Bar alleges
- Reuters: Law firm escapes sanctions over AI-generated case citations
- CalMatters: California issues historic fine over lawyer's ChatGPT fabrications
- NPR: A recent high-profile case of AI hallucination serves as a stark warning
- Reuters: Lawyers face sanctions for citing fake cases with AI, warns UK judge
- The Guardian: US lawyer sanctioned after being caught using ChatGPT for court brief
- American Bar Association: Formal Opinion 512 on generative AI tools