AI Hallucinations and Attorney Ethics: Which Professional Responsibility Rules Are Triggered and How Sanctions Have Escalated

Opening Incident: The Clark Pear Case and the 1-Year Suspension

In March 2024, a federal magistrate judge in the Middle District of Florida imposed a one-year suspension from the Bar of the U.S. District Court on an attorney who had submitted four pleadings containing 26 fabricated case citations. The case, Clark Pear LLC v. MVP Realty Associates LLC, involved AI-generated content that the attorney failed to verify before filing. The court found violations of Florida Rules of Professional Conduct 4-1.3 (diligence), 4-3.3(a)(3) (candor toward tribunal), 4-8.4(c) (misconduct), and 4-3.4(c) (standards of conduct). Reinstatement was conditioned on completing the Florida Bar's Professionalism Workshop, a Law Practice Management CLE, and counseling through the Florida Lawyer Assistance Program.

The case marked a turning point. In the committee's words, AI "can never take the place of an attorney's responsibility to conduct reasonable diligence and provide accurate legal authority." No court had yet imposed a suspension of this length solely for AI-generated citation errors. By mid-2026, it would no longer be an outlier.

The Model Rules Engaged by an AI Hallucination

Every AI hallucination that reaches a court filing does not merely create a factual error — it triggers a specific set of professional responsibility obligations that attorneys are already bound by. ABA Formal Opinion 512, published July 29, 2024, made explicit that existing Model Rules apply to generative AI use, eliminating any argument that the technology is too new for traditional ethics frameworks. The opinion treats independent verification of AI-generated citations as a non-waivable obligation.

Model Rule 1.1 — Competence (Including Comment 8)

Rule 1.1 requires attorneys to provide competent representation, and Comment 8 specifies that this includes "keeping abreast of changes in the law and its practice, including the benefits and risks associated with relevant technology." The Stanford RegLab study (preprint 2024) tested three AI legal research tools and found that Lexis+ AI and Ask Practical Law AI produced incorrect information more than 17% of the time, while Westlaw's AI-Assisted Research hallucinated more than 34% of the time on a pre-registered dataset of over 200 open-ended legal queries. Attorneys who use these tools without understanding their hallucination rates — and without implementing verification workflows — are arguably failing the competence obligation before they even file a document.

Model Rule 3.3 — Candor Toward the Tribunal

This is the rule most directly implicated by AI hallucination cases. Rule 3.3(a)(1) prohibits offering evidence the lawyer knows to be false, and Rule 3.3(a)(3) requires correction of false material evidence. In Mata v. Avianca — the seminal 2023 case that opened the AI sanction landscape — attorneys submitted a brief containing six fabricated citations generated by ChatGPT. The court found that the attorneys violated their duty of candor by failing to verify the citations before filing. The same obligation has been applied in Park v. Kim and every subsequent AI hallucination case involving fabricated citations. The duty is not excused by workload pressure, deadline constraints, or reliance on a tool marketed as "legal-specific."

Model Rule 5.1 and 5.3 — Supervisory Responsibilities

Model Rule 5.1 requires partners and supervisory lawyers to ensure that all attorneys in the firm comply with professional obligations. Rule 5.3 extends this duty to nonlawyer assistants — and state bar guidance in Florida, North Carolina, and New York has explicitly treated AI tools as analogous to nonlawyer assistants requiring the same supervision as paralegals. When a junior associate delegates legal research to an AI tool without adequate training on hallucination risk, and a partner signs the resulting brief without independent verification, both the associate and the supervising attorney may be subject to discipline. The Florida Bar Ethics Opinion 24-1 is particularly explicit on this point, holding that lawyers must apply the same supervisory standards to AI tools as they would to human assistants.

Additional Rules at Risk: 1.6, 1.4, and 7.1

Rule 1.6 (confidentiality) is engaged when attorneys input client data into AI tools whose vendor terms permit training data collection. Rule 1.4 (communication) requires attorneys to inform clients about how AI tools are being used in their representation when material. Rule 7.1 (communications about services) applies if an attorney claims AI-assisted work product without appropriate disclosure. While these rules are less frequently litigated in the hallucination context than Rules 1.1, 3.3, and 5.3, they represent additional liability vectors that a comprehensive risk assessment must address.

Annotated Case Table: 10 Representative Sanctions (2023–Mid-2026)

The following table presents a representative cross-section of federal and state cases in which courts applied sanctions for AI-generated hallucinated content. Each entry identifies the AI tool involved where known, the Model Rules or equivalent state rules violated, the sanction outcome, and key aggravation factors. This table is not exhaustive — it is drawn from the public record as captured by the Charlotin database and supplemented by case-specific reporting.

The pattern is clear. In 2023, the primary sanction was a judicial warning and a referral that rarely led to further action. By late 2024, courts began imposing firm-specific monetary penalties. By mid-2026, suspensions of six months to a year, five-figure fines, client notification requirements, and automatic bar referrals have become the norm rather than the exception. The DISCO analysis from March 2026 notes that monetary sanctions range from $120 to $10,000, with escalations tied directly to whether the attorney admitted AI use early or attempted concealment.

Aggravation Factors That Drive Harsher Sanctions

Judicial outcomes across the case table reveal a consistent set of aggravating factors that transform a sanction from a modest fine into suspension or bar referral. Understanding these factors is essential for risk assessment because they are largely within the attorney's control after an error is discovered.

• Concealment or denial of AI use: Attorneys who initially deny using AI and are later contradicted by evidence receive substantially harsher sanctions. The court in Dehghani v. Castro explicitly rejected the "empty head and a pure heart" standard, holding that "the standard under Rule 11 is one of objective reasonableness."

• Blaming subordinates or the tool: Courts have consistently rejected arguments that the attorney relied in good faith on an AI tool or delegated research to a junior associate. Under Rule 5.3, supervision obligations are non-delegable. Blaming the tool is treated as an admission of failure to supervise rather than a mitigating factor.

• Repeated offenses: The Park v. Kim case involved the same law firm that had been sanctioned in Mata v. Avianca — the court imposed a $5,000 fine, demonstrating that repeat exposure undermines any claim of good-faith misunderstanding. The Johnson v. Dunn court noted that "if fines and public embarrassment were effective deterrents, there would not be so many cases to cite."

• Failure to correct promptly: The NCSC guide emphasizes that errors must be corrected promptly with notification to the court and opposing counsel. Delay in correction — especially after the opposing party identifies the error — is treated as an independent ethics violation under Rule 3.3(a)(3).

• Volume of fabricated content: Cases involving more than five fabricated citations per filing receive harsher sanctions. The Clark Pear court cited the 26 problematic citations as evidence of systemic failure rather than isolated mistake. The Withers court imposed an $8,000 fine and revoked pro hac vice status for a similar volume.

The Escalating Sanction Trajectory: From Warnings to Suspensions

When the first AI hallucination sanctions were imposed in mid-2023, the prevailing judicial approach was leniency. Attorneys received warnings, referrals to state bars that rarely resulted in further action, and modest fines where monetary penalties were imposed at all. The Mata v. Avianca court, while sharply critical, declined to impose a fine, treating the case as a cautionary example for the profession. That approach has since been abandoned.

By early 2025, the trajectory had shifted decisively. The Sterne Kessler review documents multiple 2025 cases in which courts imposed disqualification, mandatory CLE, client notification, and automatic bar referrals — consequences that extend well beyond monetary penalties. In Johnson v. Dunn, the court disqualified the defendants' attorneys and referred them to the state bar, explicitly stating that monetary sanctions alone had not been sufficient deterrents. In Dubinin v. Papazian, the court dismissed the case without prejudice and ordered payment of attorney fees in addition to bar referrals — a combination that directly impacts the law firm's bottom line and reputation.

The first half of 2026 produced the most severe sanctions to date. Lnu v. Blanche (9th Circuit, June 3, 2026) imposed a 6-month suspension, a $5,000 fine, and — critically — a client notification requirement, meaning the sanctioned attorney's existing clients must be informed of the discipline. Withers v. City of Aberdeen (N.D. Miss., June 8, 2026) imposed an $8,000 fine, revoked pro hac vice status, and referred the attorney to multiple state bars. These are no longer outliers — they are the new baseline.

This escalation is also reflected in the aggregate data. The Charlotin database recorded 1,598 judicial decisions globally involving AI hallucinated content as of June 9, 2026. The DISCO analysis found that in 2025 alone, pro se litigants accounted for 304 hallucination incidents versus 219 for licensed attorneys worldwide — a 39% higher volume for pro se filers, but the attorney cases carry greater professional consequences because the court holds lawyers to a higher standard of competence and candor.

Practical Risk-Calibration Framework for Attorneys

The following framework is designed for attorneys, risk officers, and legal ops professionals to assess firm-wide exposure and implement safeguards. It draws on the 10-step checklist derived from ABA Formal Opinion 512, state bar guidance, and the case patterns documented above. No framework can eliminate hallucination risk — the technology's known failure rates are structural — but these steps can substantially reduce the probability that a hallucination reaches a court filing and mitigate the consequences if one does.

1. Identify applicable jurisdiction(s): Determine which state bar rules govern your practice, including rules specific to AI disclosure. State guidance from California (2023), Florida (Ethics Opinion 24-1), New York (NYSBA Task Force), and others may impose obligations beyond the ABA Model Rules.
2. Review governing ethics opinions: Read ABA Formal Opinion 512 in full, along with any state-specific ethics opinions or court standing orders. As of May 2024, more than 25 federal judges had issued standing orders on AI disclosure — that number has almost certainly increased.
3. Assess client data exposure: Review vendor terms of service to determine whether client data input into the AI tool is used for model training. Disable training data collection where possible, and obtain client consent where disclosure risk is non-trivial.
4. Review vendor data practices: Document the specific data retention policy, jurisdictions where data is stored, and any sub-processors involved in the AI tool's operation.
5. Verify all AI-generated citations independently: This is the non-waivable obligation established by ABA Formal Opinion 512. No AI legal research tool — regardless of vendor claims — is hallucination-free. The Stanford RegLab study found even RAG-augmented tools hallucinate at rates of 17% to 34%+.
6. Check court-specific disclosure requirements: Before filing any document that includes AI-assisted work product, verify whether the court has a standing order or local rule requiring disclosure of AI tool use.
7. Document the review process: Maintain a written record of the verification steps taken for each AI-assisted filing. In the event of an error, contemporaneous documentation of verification is the strongest mitigating evidence available.
8. Apply supervision obligations: Under Model Rules 5.1 and 5.3, every attorney who signs or files AI-assisted work product must personally verify its accuracy. Delegation of verification to junior associates or paralegals does not satisfy the supervisory duty.
9. Monitor for guidance updates: State bar ethics opinions, court standing orders, and regulatory guidance on AI are evolving rapidly. Subscribe to regulatory trackers and schedule quarterly reviews of applicable rules.
10. Create an incident response protocol: If an AI hallucination is discovered in a filed document, the protocol must include immediate notification to the court and opposing counsel, filing of corrected documents, self-reporting to applicable state bars where required, and communication with affected clients.

For a deeper treatment of specific Model Rules and state guidance, see our ABA Model Rule 1.1 and AI analysis and our ABA Formal Opinion 512 breakdown for jurisdiction-specific regulatory reference. For a broader survey of sanction case counts and enforcement principles, our AI Citation Hallucination Sanctions in Federal Courts article provides the quantitative landscape.

Limitations and Data Caveats

This analysis relies on several data sources that carry important limitations.

The Charlotin AI Hallucination Cases Database, while the most comprehensive publicly available aggregation, likely undercounts total incidents. It captures only cases where a court explicitly found or engaged with hallucinated content in a published decision. Cases resolved through private bar discipline, confidential settlement, unreported orders, or dismissal without published opinion are invisible to this database. Actual incidence is probably significantly higher than the 1,598 decisions tracked globally.

The Stanford RegLab study tested specific tool versions — Lexis+ AI, Westlaw AI-Assisted Research, and Ask Practical Law AI — as of early 2024. Hallucination rates may have changed with subsequent model updates, fine-tuning, and RAG architecture improvements. No independent replication study using the same methodology on current tool versions has been published as of mid-2026.

ABA Model Rules are model guidance only. Each state adopts, modifies, or rejects them independently. The specific rules applicable to any attorney depend on the jurisdiction or jurisdictions in which they are admitted. State-level guidance from California, Florida, New York, and other jurisdictions may impose obligations that differ from the ABA framework. Attorneys should consult their state bar's specific ethics opinions and any applicable court standing orders.

For jurisdiction-specific regulatory reference, see our California State Bar AI Ethics Guidance, Florida Bar AI Ethics Opinion, and NYC Bar Association Formal Opinion on AI Tools entries in our regulatory tracker.