Verification separates AI assistance from accountable legal work

On September 24, a CourtListener RECAP docket entry in Clark v. Social Security, Commissioner of reported that generative artificial intelligence was used in preparing a plaintiff’s brief. The court ordered Rich to show cause why she should not be sanctioned, including with a fine, for including factitious citations in that brief, according to CourtListener RECAP. The docket also says Rich responded to the show-cause order and that the brief involved AI use. The immediate event is therefore not an AI product launch or a policy announcement. It is a court demanding an account of a filed legal document after citations in that document were identified as factitious.

That matters because AI assistance does not move responsibility away from the filing. The Clark order puts the consequence at the point where legal work becomes part of a case: the brief was filed, its citations were challenged, and the court required a response from the person connected to it. CourtListener: AI and sanctions (opinions) describes Douglas v. Deutsche Bank National Trust Co. as a cautionary tale about the misuse of artificial intelligence and its consequences for attorneys. The practical implication for AI-native legal work is direct. A system can reduce the labor of producing a brief, but the finished work still has to survive ordinary legal scrutiny. Verification and accountability are not add-ons to the model; they are conditions for the model to deliver usable legal work.

The Clark order establishes a narrower point than a general verdict on AI in legal practice. It establishes that a court responded to factitious citations in an AI-assisted brief with a show-cause process that could include a fine. It does not establish that every AI-assisted brief contains bad citations, that AI assistance caused every defect in the filing, or that an AI-native legal provider has failed as a business. A stronger reading would require more than a single filing: it would require a pattern of finished work in which AI-generated errors repeatedly escape review, or a showing that the provider’s process leaves no accountable human or institutional check before filing. That stronger showing is not supplied by this order. But the narrower proposition is enough to reject the idea that automation erases the consequences of defective legal work. The sanction risk attaches to the work product and the filing process, not to whether the first draft came from a person or a machine.

The next observable fact is the court’s disposition after the show-cause process in Clark v. Social Security, Commissioner of: an order imposing a sanction, declining one, or explaining what review the filing received would distinguish a live consequence from a preliminary demand for explanation. A useful comparison is already visible in Macias v. CCADC, where CourtListener RECAP reports that OpenAI Codex assisted in preparing a document and that the signer certified personal checking of the AI-assisted portions, including citations. That certification does not prove the document was accurate, and the Clark docket does not show that such a process would have prevented the challenged citations. We will treat a final order and the quality of the filed work as the test: AI-native legal delivery earns credibility only when its verification process is visible in the result.