AI Hallucination Risk and Attorney Professional Responsibility: What the Sanctions Record Shows

The professional responsibility problem with AI hallucination is not hypothetical. Since the Mata v. Avianca sanctions order in June 2023 — where a federal judge in the Southern District of New York sanctioned attorneys for submitting a brief containing six fabricated case citations generated by ChatGPT — courts across multiple jurisdictions have imposed monetary sanctions, required remedial filings, and in a small number of cases referred matters to state bar disciplinary authorities.

What makes this pattern professionally significant is not just the sanctions themselves. It is the reasoning courts have used to get there: reliance on AI output, without independent verification, violates the attorney's pre-existing duty of competence under Model Rule 1.1. The AI tool is treated as irrelevant to the analysis. The question courts ask is whether the attorney reviewed what they signed.

The Competence Duty and What It Actually Requires

ABA Model Rule 1.1 requires that a lawyer provide competent representation, which includes "the legal knowledge, skill, thoroughness and preparation reasonably necessary for the representation." Comment 8 to that rule was amended in 2012 to add that competence includes keeping abreast of "changes in the law and its practice, including the benefits and risks associated with relevant technology."

That comment has become the primary textual hook for bar ethics opinions addressing AI use. The argument runs straightforwardly: if AI tools are part of legal practice, attorneys must understand their failure modes well enough to supervise their output. Hallucination — the generation of plausible-sounding but factually incorrect content, including fabricated case citations — is a documented, well-publicized failure mode. An attorney who does not know this, or who knows it but submits AI output without checking, has arguably failed the competence standard.

The Sanctions Pattern: What Courts Have Actually Done

The Mata v. Avianca order (Mata v. Avianca, Inc., No. 22-CV-1461 (PKC), S.D.N.Y. June 22, 2023) is the anchor case. Judge P. Kevin Castel imposed $5,000 in sanctions and required the attorneys to send copies of the order to the clients whose case had been harmed. The opinion documented the attorneys' failure to verify any of the six AI-generated citations — none of which existed — and rejected the defense that they had been deceived by the AI tool.

Since then, analogous orders have appeared in Texas, California, and federal courts in several other circuits. The pattern is consistent: the court identifies fabricated citations in a brief or motion, the attorney acknowledges using an AI tool, and the court frames the violation as a Rule 11 or local rule certification failure rather than as a novel AI-specific wrong. The underlying professional responsibility analysis — that the attorney certified the accuracy of something they had not checked — is the same as it would be for any other unverified factual claim.

What State Bar Ethics Opinions Have Said

State bar ethics opinions on AI use have proliferated since 2023. The California State Bar issued Practical Guidance for the Use of Generative AI in the Practice of Law in November 2023. The Florida Bar issued an opinion in January 2025. New York's ethics committee published guidance in 2024. Most of these opinions share a common structure: they do not prohibit AI use, but they impose verification obligations that effectively require attorneys to treat AI output as a draft requiring independent confirmation.

The California guidance is notable for its specificity. It identifies hallucination as a known risk and states that attorneys using generative AI for legal research must verify citations against authoritative legal databases before relying on them in any filing or client communication. It also addresses confidentiality under Rule 1.6, noting that submitting client matter details to a third-party AI service may constitute a disclosure requiring client consent.

The Florida Bar's opinion similarly requires verification of AI-generated legal authority but goes further on supervision: it states that a supervising attorney is responsible for the AI-assisted work product of subordinate attorneys and non-lawyer staff, tracking the existing supervisory obligations under Rules 5.1 and 5.3.

The Gap Between Guidance and Enforcement

Bar ethics opinions are guidance, not binding rules. Their practical effect depends on whether disciplinary authorities treat them as defining the standard of conduct for competence purposes. So far, most documented sanctions have come from courts applying Rule 11 and local certification rules — not from state bar disciplinary proceedings. That may reflect the relative speed of court enforcement versus bar disciplinary timelines, or it may reflect a deliberate choice by bar authorities to let courts take the lead.

The 5th Circuit's 2024 referral in Park v. Kim is worth watching. If the state bar disciplinary authority acts on a court referral specifically tied to AI hallucination, it would establish a precedent path from AI error to formal bar discipline — a step beyond monetary sanctions.

The Supervision Problem: Law Firms and Delegated AI Use

Most sanctions orders have targeted the signing attorney. But the professional responsibility structure for AI use in law firms is more complicated than that. When a junior associate or paralegal uses an AI tool to draft a brief section or compile a research memo, and a partner signs the resulting document without reviewing the underlying citations, the supervisory failure is the partner's — not just the associate's.

Model Rules 5.1 and 5.3 already address this. Rule 5.1 requires partners and supervising attorneys to make reasonable efforts to ensure that firm lawyers' conduct conforms to the Rules. Rule 5.3 extends the same obligation to non-lawyer staff. The AI tool itself is not an actor under these rules — but the person who uses it and the person who supervises that use are both within scope.

Candor to the Tribunal: Rule 3.3 Implications

The Mata opinion invoked Rule 3.3 — candor toward the tribunal — alongside Rule 1.1. This is significant because Rule 3.3(a)(1) prohibits an attorney from making a false statement of fact or law to a tribunal. Submitting a brief that cites nonexistent cases is, at minimum, a false statement of law.

The intent question matters here. Courts have generally not found that attorneys who submitted AI-fabricated citations intended to deceive — they found that the attorneys were careless. But Rule 3.3 does not require intent for a violation to occur. The statement is either accurate or it is not. Whether the inaccuracy was produced by an AI tool, a research error, or a typo does not change the character of the submission.

Technological advances are commonplace and there is nothing inherently improper about using a reliable artificial intelligence tool for assistance. But existing rules impose a gatekeeping role on attorneys, who must ensure the accuracy of their filings.

Mata v. Avianca, Inc., No. 22-CV-1461 (PKC), S.D.N.Y. June 22, 2023

Confidentiality Exposure: The Other Hallucination Risk

Citation fabrication is the most visible AI hallucination risk, but it is not the only one with professional responsibility consequences. A less-discussed failure mode is hallucinated facts — AI output that mischaracterizes the contents of a contract, misstates a client's position, or invents procedural history. These errors may not be caught as easily as a nonexistent case citation, because they do not generate a "case not found" error when someone tries to verify them.

There is also the confidentiality dimension. Attorneys who submit client documents to a cloud-based AI tool may be sharing confidential information with a third party. Whether this triggers Rule 1.6 depends on the service agreement and the nature of the data, but several bar opinions — including California's 2023 guidance — have flagged this explicitly. The risk is not just that the AI hallucinates; it is that the attorney may have created a disclosure problem in the process of using the tool.

What Verification Actually Looks Like in Practice

Bar opinions and court orders establish that verification is required. They are less specific about what verification must include. Based on the documented sanctions record and the structure of the ethics opinions, a defensible verification workflow for AI-assisted legal research includes at minimum:

1. Confirming each cited case exists in an authoritative legal database (Westlaw, Lexis, or equivalent) before the document is filed.
2. Confirming that the cited case actually stands for the proposition it is cited for — not just that the case exists.
3. Checking that cited cases have not been overruled or distinguished in ways that undermine the argument.
4. Reviewing AI-generated factual summaries of documents against the source documents, not just the AI summary.
5. Assigning the verification step to a named individual — not leaving it as an assumed background task.

Steps 1 and 3 are the ones most directly addressed by the sanctions record. Step 2 is arguably more important — a case can exist and be cited for something it does not actually hold — but it requires more judgment and is harder to systematize. Step 5 is a firm management question, not just a legal research question.

Court-Imposed AI Disclosure Requirements

A separate but related development is the emergence of standing orders requiring attorneys to disclose AI use in filings. By early 2025, a significant number of federal district court judges had issued standing orders addressing AI — some requiring disclosure of any generative AI use in brief preparation, others requiring certification that AI-generated content had been reviewed and verified.

These orders vary considerably. Some require disclosure only if generative AI was used to draft substantive legal argument. Others cover any AI-assisted research. A few explicitly exempt AI-powered legal research platforms (like Westlaw or Lexis) from the disclosure requirement, treating them differently from general-purpose large language models.

Where the Regulatory Framework Currently Stands

As of mid-2026, the professional responsibility framework for AI hallucination risk in US legal practice is built on existing rules applied to new facts — not on AI-specific rules. Model Rule 1.1 (competence), Rule 3.3 (candor), Rules 5.1 and 5.3 (supervision), and Rule 1.6 (confidentiality) are doing the work. The ABA has issued formal ethics opinions and standing committee guidance addressing AI, but has not amended the Model Rules to add AI-specific provisions.

State bars have moved faster than the ABA on ethics opinions, but the opinions themselves vary in specificity and binding force. California's guidance is detailed; some other states have issued only brief general statements. There is no uniform national standard for what AI verification looks like, how AI use must be disclosed to clients, or whether specific AI tools are permissible for specific tasks.

The Malpractice Exposure Question

Court sanctions and bar discipline are public consequences. Malpractice liability is the private one. An attorney who submits AI-fabricated citations that harm a client's case — by causing a motion to be stricken, by undermining the attorney's credibility with the court, or by causing a filing deadline to be missed during the corrective process — has potentially caused compensable harm.

Documented malpractice claims specifically attributed to AI hallucination remain rare in the public record as of mid-2026, likely because most such matters settle confidentially. But the legal theory is straightforward: if the standard of care requires citation verification, and the attorney failed to verify, and the client suffered harm as a result, the elements of a negligence claim are present.

Legal malpractice insurers have begun asking about AI use in renewal applications. Some carriers have issued guidance recommending specific verification protocols as a condition of coverage for AI-assisted work product. This is an early signal that the insurance market is starting to price AI hallucination risk — which in turn creates an economic incentive for firms to document their verification workflows.

What Remains Unsettled

Several questions that practitioners actually face do not yet have clear answers in the ethics opinions or case law.

• Does using a legal-specific AI research platform (such as Westlaw's AI features or Lexis+ AI) satisfy the verification obligation, or must citations still be independently confirmed against the underlying database?
• How should attorneys handle AI-generated factual summaries of documents — not citations, but characterizations of what a contract or statute says? The verification standard for factual claims is less clearly articulated than for legal citations.
• What disclosure, if any, is owed to clients when AI tools are used in their matters? Some bar opinions address this; others do not. The ABA's formal opinion on AI (Formal Opinion 512, issued in July 2024) discusses the issue but does not mandate client disclosure in all circumstances.
• How does the competence obligation apply to attorneys who rely on AI for transactional work — contract drafting, due diligence checklists — where errors may not be discovered until enforcement or dispute?

These gaps mean that attorneys and firms are currently operating with partial guidance. The safest position — treating all AI output as unverified draft material requiring independent review before use in any client-facing context — is also the most conservative and the most resource-intensive. Whether that standard is actually what competence requires, or whether a more calibrated approach is defensible, is a question the profession has not yet fully resolved.