What Published Federal and State Court Orders Actually Say About AI-Assisted Legal Research Tools and the Duty to Verify Citations: A Doctrinal Map Through October 2026

Executive Summary

The judicial record on attorney responsibility for AI-generated citations has evolved from isolated sanctions orders into a growing body of administrative standing orders, revised local rules, and published opinions. Courts have moved through three recognizable phases: initial crisis response following Mata v. Avianca, consolidation into local-rule infrastructure, and tentative doctrinal elaboration. What the record shows clearly, however, is that no jurisdiction has yet published a holding that distinguishes between retrieval-augmented AI systems with live database access and purely generative systems—a doctrinal gap that will likely define the next phase of litigation.

1. The Foundational Sanctions Record

The controlling precedent in federal court remains Mata v. Avianca, Inc., No. 1:22-cv-01461 (S.D.N.Y.), in which Judge P. Kevin Castel imposed $5,000 in sanctions jointly and severally on plaintiff's attorneys Peter LoDuca and Steven Schwartz and their firm, Levidow, Levidow & Oberman on June 22, 2023. The order rested on Federal Rule of Civil Procedure 11's subjective bad faith standard. The factual predicate was submission of ChatGPT-generated case citations that were entirely fictitious; the attorneys had not independently verified those citations against any authoritative legal database before filing. Judge Castel's order also required the sanctioned attorneys to transmit the opinion to their client and to each sitting judge whose name had been falsely attached to a fabricated opinion. No disciplinary referral was issued under the order, and the court did not frame the ruling under Model Rule 3.3.

Mata did not resolve, and did not purport to resolve, the question of what standard of verification satisfies Rule 11. It established only that submitting AI-generated citations with no verification at all falls below the minimum required by the existing candor and certification framework.

The Second Circuit added a distinct dimension in Park v. Kim, 91 F.4th 610 (2d Cir. 2024), where the court referred an attorney to the circuit's grievance panel after encountering an AI-fabricated citation in a brief. Park v. Kim is significant because the referral mechanism—as opposed to a monetary sanction—signals that the circuits view AI hallucination failures as potentially implicating professional responsibility rules rather than solely the civil Rule 11 framework.

At the state disciplinary level, the Colorado case of People v. Crabill (Colo. PDJ, November 2023) resulted in the suspension of attorney Zachariah Crabill over ChatGPT-fabricated citations, with the Presiding Disciplinary Judge finding violations of competence, diligence, and honesty obligations. Crabill is the clearest instance of a state disciplinary body treating an AI-hallucination failure as a substantive professional-conduct violation rather than a procedural irregularity.

Tracking the aggregate volume of related judicial activity, Damien Charlotin's public AI Hallucination Cases database had grown to over a thousand decisions globally by early 2026, with more than a thousand U.S. cases documented by mid-2026—a figure that makes clear these are no longer isolated incidents confined to a handful of high-profile matters.

2. From Disclosure to Affirmative Verification: Local Rules and Standing Orders

The more consequential structural development since Mata has been the administrative proliferation of standing orders and local rule amendments that move beyond mere disclosure to impose affirmative verification duties.

The Northern District of Texas adopted Local Rule 7.2(f), effective September 2, 2025, which requires attorneys to disclose on the first page of any brief prepared with generative AI that such a tool was used; a brief filed without that disclosure certifies that no part of it was prepared using generative AI. This is a substantive verification obligation, not a disclosure-only requirement. Attorneys in that district cannot satisfy the rule by disclosing AI use and stopping there; the rule demands that AI-generated citations be checked against primary sources.

A Bloomberg Law tracker identified 39 federal judges with active AI-related standing orders. The content of those orders varies considerably. Some require disclosure only. Others—consistent with the N.D. Texas model—require certification of independent verification. The heterogeneity of these orders means that practitioners must treat each assignment as jurisdiction-specific; there is no single federal standard.

The Judicial Conference of the United States, as of October 2026, has not adopted a uniform AI-filing certification rule applicable across federal courts. The Administrative Office's interim guidance issued July 31, 2025, concerns the judiciary's internal use of AI tools rather than attorney obligations in filings. Proposed Evidence Rule 707 represents the Conference's primary pending AI-related rulemaking, but it addresses the admissibility of AI-generated evidence rather than citation verification duties.

3. The Rule 1.1 Competence Analogy in Published Opinions and Ethics Guidance

Courts have not yet issued a published holding that formally ties AI citation verification to Model Rule 1.1 by name. The doctrinal connection, however, has been made explicit in ethics opinions that inform how courts are likely to interpret attorney conduct.

ABA Formal Opinion 512, issued July 29, 2024—the ABA's first formal opinion specifically addressing generative AI—covers competence, confidentiality, communication, candor, supervision, and reasonable-fee obligations. On the competence dimension, the opinion requires attorneys using AI tools to understand the tool's limitations sufficiently to evaluate its outputs, including verifying that any cited authority actually exists, says what the AI reports, and remains good law. The opinion does not use the phrase "independent verification" but the operative standard is functionally identical: the attorney, not the AI, bears the professional responsibility for the accuracy of the product submitted to the tribunal.

Texas Professional Ethics Committee Opinion 705, issued February 2025, applies substantially the same framework. It treats the use of a generative AI tool as creating a supervisory obligation analogous to supervising a junior associate: the supervising attorney must review outputs with the level of care that the stakes of the matter require, and citation verification is treated as a baseline element of that review.

Virginia LEO 1901 reaches a related conclusion regarding attorney fees: it addresses the reasonableness of fees when AI tools reduce time spent on a task, and recognizes that identifying AI hallucinations and integrating only validated material is part of the skill lawyers exercise when using such tools—implying a competence obligation to catch those errors before submission.

Taken together, this body of ethics authority constructs a consistent rule: the duty of competence under Model Rule 1.1, as interpreted for AI-assisted research, requires affirmative verification of any AI-generated citation against an authoritative primary source before that citation is submitted to any tribunal. This is not a new duty; it is the existing duty of citation accuracy as applied to a new class of research tools.

4. The Unreached Question: Retrieval-Augmented Versus Generative Systems

The most significant doctrinal gap in the current record is the absence of any published opinion distinguishing between AI research tools that operate through retrieval-augmented generation—querying a live, authoritative legal database and grounding responses in retrieved documents—and purely generative systems that synthesize responses from training-data patterns without real-time database access.

The distinction matters because the hallucination risk profile differs fundamentally between these architectures. A retrieval-augmented system tied to a verified legal database can, in principle, surface only citations that exist in that database. A purely generative system with no grounding mechanism produces citations from statistical pattern-matching and is substantially more prone to fabrication.

No court order through October 2026 has formally recognized this distinction as bearing on the attorney's verification obligation. The standing orders reviewed—including the N.D. Texas Local Rule 7.2(f) and the AI-related standing orders tracked by Bloomberg Law—apply a uniform verification requirement without differentiating by tool architecture. ABA Formal Opinion 512 similarly does not distinguish between these categories.

This doctrinal flatness likely reflects the practical difficulty of adjudicating tool-architecture claims in real time. Courts do not have a ready mechanism to audit which AI tool an attorney used or to verify claims about its architecture.

Conclusion

The judicial and regulatory record through October 2026 establishes a clear baseline: independent verification of any AI-generated citation is an affirmative professional obligation, enforceable through Rule 11, local rules, state disciplinary proceedings, and the ethics opinions that courts use to calibrate the standard of care. The architecture of the AI tool used does not, under current authority, modify that obligation. Courts have been consistent in treating this as a competence and candor problem rather than a technology problem—and the sanctions issued, from Mata v. Avianca through People v. Crabill, demonstrate that the enforcement mechanism already exists. The remaining doctrinal work is elaboration, not foundation-building.