On January 1, 2025, the Illinois Supreme Court adopted a statewide AI policy that included a single sentence destined to become the most-cited line in the disclosure debate: “Disclosure of AI use should not be required in a pleading.” A year later, on January 15, 2026, the Chief Judge of Miami-Dade County’s Eleventh Judicial Circuit issued Administrative Order No. 26-04, requiring every attorney and self-represented litigant who uses generative AI in preparing any court document to “disclose such use on the face of the filing” and certify that “all factual assertions, legal authority, and citations have been independently reviewed and verified for accuracy.” Palm Beach and Broward counties followed. A lawyer practicing in both states now faces a regime in which identical conduct requires mandatory certification and risks sanctions in one jurisdiction and requires nothing at all in the other.
The current picture
The split runs across every level of the judiciary. Illinois took the most permissive position in the country. Chief Justice Mary Jane Theis framed the reasoning in terms of sufficiency: “our current rules are sufficient to govern its use.” California is moving in the opposite direction—the State Bar’s Committee on Professional Responsibility and Conduct approved proposed amendments to six Rules of Professional Conduct in March 2026, including a requirement that lawyers “independently review, verify, and exercise professional judgment regarding any output generated by the technology” used in representing a client. New York has no statewide rule; individual judges have issued part rules whose requirements, as a Greenberg Traurig review found, “vary wildly” from courtroom to courtroom within the same courthouse.
At the federal level, the picture is fractured along individual-judge lines. Judge Brantley Starr of the Northern District of Texas issued the first widely discussed standing order in 2023, requiring a binary certification—no AI was used, or AI-generated content was verified. Judge Evelyn Padin in the District of New Jersey requires lawyers to identify the specific tool, describe which sections it assisted with, and certify human review. Hundreds of federal judges have now issued individual AI orders. Bloomberg Law and Law360 maintain trackers that keep growing. The Judicial Conference has not adopted a national policy, and none appears imminent.
The case against disclosure
The arguments against mandatory disclosure deserve to be taken seriously. The most influential articulation comes from Judge Paul Grimm (ret.), Maura Grossman, and Daniel Brown, whose 2023 Judicature article asked whether disclosure and certification orders are “really necessary.” Their answer was no—because the existing framework already prohibits the conduct these orders target. Rule 11 requires attorneys to certify factual and legal accuracy. Model Rule 3.3 requires candor. Model Rule 1.1, as amended by Comment 8, requires technological competence. A lawyer who files fabricated citations has violated these rules regardless of whether a standing order separately requires disclosure. Adding a disclosure obligation, on this account, creates redundancy without additional deterrence.
Grossman, Grimm, and Brown also flag the orders’ overbreadth. Some require disclosure of “any AI tool”—a category that sweeps in Westlaw’s AI-assisted research, Grammarly, and Microsoft Word’s predictive text. Others target “generative AI” without defining the term. The “mosaic of inconsistent, individual standing orders” imposes compliance costs that fall hardest on solo practitioners and small firms, and the inconsistency across courtrooms is itself a source of confusion and expense.
I think the redundancy point is correct as far as it goes. The Q1 2026 sanctions data bears it out: not one of the sanctioned lawyers was caught because of a disclosure requirement. They were caught because someone—opposing counsel, a law clerk, the judge—checked the citations. And the overbreadth concern is sharpening as AI embeds into every research and drafting platform. “Did you use AI?” is rapidly becoming a question to which the only honest answer is “yes, at every step,” and requiring disclosure of all of it produces boilerplate that tells the reviewing court nothing useful—the legal equivalent of a cookie-consent banner everyone clicks through without reading.
But I think the anti-disclosure position misidentifies what disclosure is for. Both Grossman, Grimm, and Brown and the Illinois Supreme Court treat disclosure as an additional compliance obligation imposed on the filing attorney—a rule about the lawyer’s conduct. If that were all disclosure is, the redundancy argument would be decisive. Existing rules already deter the conduct. Another rule does not deter it more.
What disclosure tells the person on the other end
I want to push back on the assumption that disclosure is primarily about regulating the filer. Its more important function is informational: it tells the person reviewing the work something that person needs to know in order to review effectively.
Consider what it looks like to review an associate’s research memorandum when you know the associate used AI to generate an initial draft. You can ask what tool was used, what it was asked to do, how the results were verified, and which parts of the final product reflect the associate’s own analysis. These questions are diagnostic. They let you assess whether the associate exercised the judgment the memorandum represents, or accepted the tool’s output without the independent evaluation the work requires. You can focus your verification on the specific contributions the tool made—checking whether the cited cases exist and say what the memo claims, whether the analytical framework reflects the associate’s assessment of the law or reproduces the model’s pattern-matched synthesis. These are answerable questions, and answering them takes a fraction of the time that comprehensive output review demands.
Now consider the same memorandum without disclosure. You do not know whether AI contributed. You have two choices: verify everything, or verify selectively and accept the risk that unverified portions contain errors. In practice, most partners choose the second option, because the first is incompatible with the time constraints under which legal work is produced and reviewed. The result is that undisclosed AI use shifts risk from the person who used the tool to the person who reviews the work—without that person’s knowledge or consent.
The same dynamic plays out whenever someone reviews another person’s AI-assisted legal work. A judge who knows a brief was AI-assisted can direct a clerk to run the citations through Westlaw before investing time in the substantive arguments. A law professor who knows a student used AI to research an appellate brief can structure the assessment around the student’s ability to explain and defend the cited authorities, rather than spending the assessment period determining whether the authorities exist. In each case, disclosure converts an open-ended verification problem into a structured supervision task. The distinction between interrogating process and verifying output determines how far a reviewer’s limited time and attention can reach.
The Q1 2026 sanctions cases illustrate the cost when that information is missing. In Couvrette v. Wisnovsky, fabricated citations escalated across three briefs over five months—two in the first, seven in the second, sixteen in the third—even after opposing counsel flagged the earlier errors. The court found persuasive evidence that the attorney’s client had drafted the briefs using AI and that the attorney filed them without verification. Opposing counsel bore the burden of catching errors the filing attorney never checked. Had the filings carried a disclosure and certification, the court would have known from the first brief that AI was involved, and the certification—that all citations had been “independently reviewed and verified”—would have given the court something concrete to test when the verification proved nonexistent.
Without disclosure, the people reviewing AI-assisted work default to suspicion or trust, and neither serves them well. Treating every filing as potentially AI-generated and verifying accordingly imposes costs that scale with the volume of work reviewed. Assuming filings reflect human judgment unless given reason to doubt it is the posture that produced the sanctions the profession is now cataloguing at an accelerating rate.
Why the patchwork looks the way it does
I think the patchwork is a symptom of a conceptual problem. Illinois and Florida disagree about whether disclosure should be required, but they share a framing: both treat it as a question about the filing attorney’s obligations. Illinois says the obligation is unnecessary. Florida says it is essential. Neither asks what disclosure enables the judge, the partner, or the client on the receiving end to do.
That framing explains why hundreds of federal judges have each independently designed a different disclosure regime. No coordinating body has articulated a coherent rationale for what disclosure is supposed to accomplish, so each judge has improvised. A judge in the Northern District of Texas requires a binary certification. A judge in the District of New Jersey requires identification of the specific tool, the specific sections, and the verification methodology. The difference does not reflect a difference in what these judges need to evaluate briefs effectively. It reflects a difference in how two judges, working without guidance, conceptualized a problem on their own.
A disclosure framework designed around supervision rather than compliance would look different from most of the standing orders now in force. It would not ask whether AI was used—a question that loses meaning as AI embeds into every tool in the workflow. It would ask what role AI played in the specific work product and what the human did to evaluate the portions AI contributed. The distinction maps onto the framework I described in an earlier post about delegating tasks rather than judgment: a supervision-oriented regime would distinguish between AI used for tasks that do not require independent verification (formatting, document organization) and AI used for tasks requiring the filer’s professional judgment (legal analysis, factual characterization, strategic recommendations), and it would require the filer to identify the latter and describe the verification steps taken.
Judge Padin’s standing order comes closest to this model. Her order asks what I think are the right questions—not “did you use AI?” but “where did AI contribute, and what did you do about it?” California’s proposed Rule 1.1 amendment moves in a related direction, requiring independent verification of all AI-generated output as a component of competence. But a rule requiring verification and a rule requiring disclosure of the verification process serve different functions. A lawyer can be required to verify without being required to tell the court how the verification was done, and the court is worse off for the gap.
The practical stakes
The seventeen court decisions in a single day in March 2026 noting suspected AI hallucinations represent the visible fraction of undisclosed AI use that produced detectable errors. The invisible fraction—filings where AI-generated analysis passed without scrutiny because no one knew to scrutinize it—is, by definition, uncounted. Disclosure does not eliminate the need for verification. But it tells the judge, the partner, or the professor where to look, and in a system where no one has time to verify everything, knowing where to look is most of the problem.
This post draws on the Illinois Supreme Court Policy on Artificial Intelligence (effective January 1, 2025); the Eleventh Judicial Circuit of Florida’s Administrative Order No. 26-04 (January 15, 2026); Maura R. Grossman, Paul W. Grimm, and Daniel G. Brown, Is Disclosure and Certification of the Use of Generative AI Really Necessary?, 107 Judicature 2 (2023); the California State Bar’s proposed amendments to the Rules of Professional Conduct (public comment period closed May 4, 2026); the ABA Model Rules of Professional Conduct and Formal Opinion 512; and the Vermont Law Review’s analysis of mandatory AI disclosures and uniform standards. The supervision and delegation frameworks referenced throughout are developed in prior posts on judgment delegation and verification.