Courts across the country have responded inconsistently to the use of AI to prepare court filings. Some judges have released individual standing orders, with some requiring disclosure of any AI usage, banning AI outright, or simply cautioning attorneys on the dangers of not verifying AI output. Sanctions vary just as widely, ranging from mere warnings to fines of nearly $60,000. This fragmented approach has done little to stop litigants and lawyers from filing fake AI-generated citations. A nationwide rule with a greater deterrent effect is needed.
Courts have generally taken two competing approaches on the role of AI in legal practice. The first treats AI as just another research tool that is sufficiently regulated by existing professional conduct rules. The second treats AI as a novel technology that requires new obligations beyond existing ones. The consequence of this divide is that attorneys who appear before judges with different positions face uncertainty about whether they can even use AI prior to judicial assignment. This disagreement also touches on whether new regulation is even necessary.
The “just another tool” camp holds that existing obligations are sufficient to govern attorneys’ use of AI in their filings. The Illinois Supreme Court is illustrative: it released guidance stating that judges should refrain from requiring attorney disclosure of AI use. The court believes that the Rules of Professional Conduct are sufficient to govern AI usage. However, this guidance was nonbinding, and several Illinois judges have required AI use disclosures. This intra-state conflict shows that even within a single jurisdiction, there can be inconsistent policies.
The “novel technology” camp believes existing rules are insufficient to regulate AI in the legal profession, and new mandatory guidance is necessary. The response has spanned from simple disclosure to outright bans on the use of AI. One advantage of these policies is greater uniformity: in both the U.S. District Court for Nebraska and the U.S. District Court for Hawaii, there are district-wide requirements for attorneys to disclose AI use and certify that the AI output has been verified. Although the policies are consistent, they still do not prevent hallucinated cases from entering the record—noncompliance with disclosure requirements can only be discovered after the documents have been filed.
The most restrictive rules from the “novel technology” camp are outright bans on using AI. For example, in the U.S. District Court for the Northern District of Illinois, Judge Sharon Coleman’s policy prohibits using AI to “draft memoranda.” This rule is ineffective. The scope of “draft[ing]” memoranda is unclear, and an attorney could still use AI to conduct research before writing a memorandum themselves and filing it with hallucinated cases from their research. In such a scenario, the risks of using AI have not been mitigated since fake cases could still enter the record.
The important conclusion from surveying courts nationwide is that disclosure requirements of all kinds are not working. In the last three years, over 1,500 instances of AI hallucinations have been caught by courts or opposing counsel across the United States, and over 650 of those are from 2026 alone. And these are just the instances where the hallucination was caught; the number of actual instances is likely far greater.
Disclosure requirements fail because they do not meaningfully change the behavior of attorneys. Many courts require attorneys to certify that all citations are accurate, so in theory, both the accuracy of citations and the transparency of attorneys are addressed. But, these requirements do not practically add anything beyond the existing requirements of FRCP 11. Rule 11 has required attorneys to certify that their claims are “warranted by existing law” for over decades. An attorney who recklessly submits unverified citations already violates Rule 11 and would not meaningfully be deterred by another “disclosure” requirement. Disclosure certifications do not change the underlying incentives or behavior.
To better combat hallucinated cases entering the record, a “citation authentication” requirement should be implemented nationwide. Under this requirement, any attorney filing with the court would be required to submit a copy of every case cited, referred to as the authentication folder. Court e-filing platforms could be modified to automatically compare the citations in the filing with the attached cases and reject any filing where a citation lacks a case in the authentication folder. A hallucinated citation cannot be produced and would have no text, so it could never pass this authentication.
This solution creates a second layer of protection that does not solely depend on the submitting attorney’s diligence. It also makes Rule 11 obligations enforceable at the time of filing, instead of after the hallucinations enter the record.
The most significant objection to this proposal is the burden it places on litigants and the courts. This requirement would, of course, require more time on the part of litigants and would increase the size of the record. But requiring citation authentication would not unduly burden attorneys because it is accessible without advanced legal research tools, and attorneys are already required to check that their citations are accurate. It is a simple step to download each case as it is accessed in the research and drafting processes.
And for the courts, this process can be automated. Existing e-filing platforms like CM/ECF could be updated to automatically compare each citation from the table of authorities to the cases in the authentication folder. Courts currently face significant administrative costs from retroactively correcting parts of the record that rely on hallucinated authority. There is also a reputational cost to consider: both courts and the legal system in general are facing rising public concerns on their legitimacy and fairness. These negative effects would be heavily reduced by preventing hallucinations from entering the record beforehand.
Citation authentication would not solve every problem created by AI in legal practice, but it would directly address one of the most concrete and recurring harms: fabricated authorities entering the record. By requiring cited cases to be submitted with the filing and enabling automated review at the point of submission, the proposal would shift courts from after-the-fact sanctions to front-end prevention. In doing so, it would create a modest but meaningful safeguard against hallucinated citations while preserving lawyers’ ability to use AI responsibly.
