If AI fabricates citations, legal work still carries consequences

On September 24, the Court ordered Rich to show cause in Clark v. Social Security, Commissioner of as to why she should not be sanctioned, including with the imposition of a fine, for including factitious citations in a plaintiff’s brief. The CourtListener RECAP dockets: generative AI and sanctions or hallucinations entry says the brief used generative artificial intelligence and that Rich’s response addressed her use of AI in preparing it. The docket entry identifies the filing as Case 2:25-cv-11554-EAS and the court’s action as a show-cause order, not a final sanctions ruling.

The order moves the accountability question in the direction that matters for AI-native legal work: using AI does not transfer responsibility for the finished filing away from the person who submits it. The alleged defect was not simply that AI assisted with drafting. The Court identified factitious citations and required Rich to explain why sanctions should not follow, with a fine among the possible consequences. That is a direct test of whether a lawyer verified the work before placing it before a court. A system that produces legal text without a reliable review step has not eliminated the consequences of bad legal work; it has made the review step more consequential.

The Clark order establishes that one court linked AI-assisted preparation, false citations and possible sanctions in one filing. It does not establish that AI assistance itself warrants punishment, that Rich will ultimately be sanctioned, or that every AI-native legal model produces defective work. The distinction matters. CourtListener: AI and sanctions (opinions) describes Douglas v. Deutsche Bank National Trust Co., Published Order as a cautionary tale about the misuse of AI and its consequences for attorneys, but that language still concerns the conduct addressed in that order. A broader claim would require final orders across more matters showing that verification controls consistently prevent fabricated authorities from reaching courts and that courts hold the responsible legal actor accountable when those controls fail.

The same accountability principle appears in the filing practice described in Macias v. CCADC. CourtListener RECAP dockets: generative AI and sanctions or hallucinations reports that OpenAI Codex assisted with the document and that the signer certified personal checking of the AI-assisted portions, including all citations. That certification is a process representation, not proof that the citations were accurate. It shows what a stronger reading of AI-assisted legal work would require: a person must stand behind the final product, and the filing must make that responsibility observable. Clark shows the cost when the submitted work instead contains factitious citations.

The next decisive fact is the court’s order resolving the Clark show-cause proceeding. A final order imposing a fine or another sanction would confirm that the filing’s defects produced a concrete consequence; an order declining sanctions would narrow the lesson to scrutiny and required explanation. The filing should also be read for whether the court identifies personal review, the AI tool’s role, and the factitious citations as separate grounds for its decision. The position holds either way: AI assistance does not remove the consequences of bad legal work, but the final order will show how far this case carries that principle.