Connecticut Supreme Court: Duty of Competence Requires Independent Verification of Generative AI Citations; Sanctions for “Hallucinated” Authorities

1. Introduction

In TOV Realty, LLC v. Suarez (SC 21183), decided alongside Kosel Equity, LLC v. MacGregor (SC 21184), the Supreme Court of Connecticut confronted—expressly “for the first time”—the professional-responsibility and sanctions consequences of a lawyer’s use of generative artificial intelligence (“generative AI”) that resulted in court filings containing “hallucinated” citations (i.e., fabricated or inaccurately generated case authorities).

Attorney Ian G. Gottlieb and his firm, GLG Law, LLC, filed applications for certification to bring public interest appeals under General Statutes § 52-265a and later filed merits briefs. In both stages, the submissions contained multiple erroneous, unverified citations that arose during an AI-assisted editing process. After amici curiae (represented by the Jerome N. Frank Legal Services Organization) alerted the court to apparent hallucinations, counsel filed errata and corrected briefs. The court then issued a sua sponte show cause order to determine whether sanctions should be imposed for misrepresenting the law through generative AI.

The core issues were: (1) whether submitting filings containing AI-hallucinated citations constitutes a violation of the Rules of Professional Conduct—at minimum, the duty of competence under Rule 1.1; and (2) what sanctions are appropriate when the misconduct is negligent rather than intentional and does not alter the underlying validity of the legal propositions advanced.

2. Summary of the Opinion

The court held that Attorney Gottlieb’s submission of court documents containing seven erroneous and unverified citations—caused by generative AI altering or adding citations—violated Rule 1.1’s duty of competent representation. Although the court found no intent to deceive and noted that the substantive legal propositions remained “well established and correct,” it emphasized that inaccurate citations undermine the integrity of the judicial process and impose real burdens on adversaries and courts.

Invoking its inherent authority to administer justice, the court imposed sanctions tailored to education, deterrence, and institutional accountability:

  • Mandatory CLE: In addition to the annual minimum required by Practice Book § 2-27A, Attorney Gottlieb must complete six additional hours of CLE in ethics and law office management, with three hours specifically addressing generative AI.
  • Monetary remedy via education-support donation: Attorney Gottlieb and GLG Law, LLC must each donate $1000 to the CT Bar Institute, Inc., to support training on legal ethics, law office management, and generative AI’s role in those areas.
  • Compliance reporting: Both must file a compliance report in the docket numbers within six months, with a courtesy copy to the Statewide Grievance Committee.
  • Reciprocal discipline compliance: Attorney Gottlieb must satisfy reciprocal discipline obligations in any other jurisdiction where admitted.
  • Public posting: The order will be posted on the Judicial Branch website; a courtesy copy goes to the Statewide Grievance Committee (expressly “not a referral” for further action).

The court also underscored that law firms share responsibility when partner-level review occurs without policies and procedures addressing responsible generative AI use.

3. Analysis

3.1. Precedents Cited

McCarthy v. United States Drug Enforcement Administration, 171 F.4th 245 (3d Cir. 2026)

The court relied on McCarthy as a contemporary federal appellate statement that courts “uniformly” treat AI-generated erroneous or hallucinated citations as a competence problem. By citing McCarthy (including its footnote collecting authorities), the Connecticut Supreme Court positioned its decision not as an outlier, but as Connecticut’s alignment with an emerging national consensus: attorneys remain responsible for verifying every authority cited, regardless of drafting tool.

Park v. Kim, 91 F.4th 610 (2d Cir. 2024)

Park was cited for the principle that failing to check AI-generated citations violates a duty of “reasonable inquiry” into the accuracy of court filings under Federal Rule of Civil Procedure 11(b)(2). Although Connecticut’s decision rests on Rule of Professional Conduct 1.1, the court used Park to emphasize that AI-hallucination errors are not merely stylistic defects—they implicate baseline certification obligations associated with signing and filing documents.

Cojom v. Roblen, LLC, 2025 WL 3205930 (D. Conn. November 17, 2025)

Cojom reinforced two points that Connecticut adopted explicitly: (1) the “oversight” is more than “sloppy lawyering,” because it “imperils the integrity of our judicial process”; and (2) sanctions are warranted even where misconduct is negligent. The court’s quotation of Cojom signaled a judiciary-centered rationale: the harm is systemic (diverting judicial resources, corrupting adversarial assumptions), not merely inter-party.

State v. Coleman, 280 N.E.3d 1042 (Ohio App. 2026)

State v. Coleman provided an especially close analogue because it involved AI-generated inaccuracies in a court filing (there, transcript quotations produced by uploading case-related materials to a generative AI application). Connecticut borrowed Coleman’s broader framing: even absent intent, reckless or careless reliance on AI-generated content “strikes at the foundation of the adversarial system.” The court used this to justify sanctions grounded in institutional integrity rather than punitive intent.

Garner v. Kadince, Inc., 571 P.3d 812 (Utah App. 2025)

Garner supplied the court’s most concrete articulation of harm: opposing parties waste time and money, judicial resources are diverted, and clients may lose valid arguments. By quoting Garner, the court connected the “fake citations” problem to practical litigation costs and client-service harms—supporting sanctions even where the underlying legal propositions are correct.

Lafferty v. Jones, 236 Conn. App. 672 (2025)

Lafferty was cited for Connecticut’s use of the American Bar Association’s Standards for Imposing Lawyer Sanctions and the role of aggravating and mitigating factors (Standards 9.22 and 9.32). The citation anchored the court’s sanction methodology in familiar Connecticut disciplinary analysis: structured consideration of culpability, harm, cooperation, remorse, and history.

United States v. Heppner, 820 F. Supp. 3d 292 (S.D.N.Y. 2026)

Although not essential to the ultimate holding, Heppner was important for the court’s warning that “competence” in using generative AI also includes understanding implications for attorney-client privilege, confidentiality, and work product—especially “when a public access platform is involved.” This expands the doctrinal takeaway beyond citation verification: a competent attorney must evaluate the data-security and privilege risks of the tool itself.

Kosel Equity, LLC v. MacGregor, 354 Conn. 842 (2026) and TOV Realty, LLC v. Suarez, 354 Conn. 745 (2026)

These merits decisions were referenced to clarify posture: the sanction proceeding followed the court’s disposition of the underlying appeals, underscoring that the sanctions inquiry was about integrity of filings and professional obligations, not about altering case outcomes.

3.2. Legal Reasoning

The court’s reasoning proceeds in four steps:

  1. Identify the misconduct and its source: Attorney Gottlieb performed research using LexisNexis and verified citations with Shepard’s. He then used ChatGPT to improve organization and quality. ChatGPT “added new case citations or altered existing case citations,” and these changes were not re-verified before filing.
  2. Fix the minimum ethical violation: At the show cause hearing, Attorney Gottlieb admitted his conduct violated Rule 1.1, and the court agreed. The court noted possible applicability of Rules 1.6, 3.3, 5.1, and 8.4, but explicitly declined to decide those issues because Rule 1.1 sufficed.
  3. Explain why negligence still warrants sanctions: Drawing on the cited authorities, the court emphasized that fabricated citations waste adversary and judicial resources and threaten the adversarial system’s core assumptions—regardless of intent. The court expressly found no intent to deceive, but treated the competence lapse as sanctionable due to the systemic risk.
  4. Calibrate sanctions using aggravating/mitigating factors and institutional needs: Applying the ABA Standards framework (Standards 9.22 and 9.32), the court weighed mitigation (no dishonest motive, cooperation, contrition, no discipline history, negligent use of new technology, corrections filed) against the seriousness of misleading citations. The sanctions emphasize: (a) education and competency building (CLE, including AI-specific hours); (b) profession-wide remediation (donations to support training); (c) accountability and monitoring (compliance report, public posting); and (d) regulatory consistency (reciprocal discipline obligations).

A notable aspect of the reasoning is the court’s allocation of responsibility beyond the individual lawyer. Although Attorney Gottlieb accepted full responsibility, the court stated that “the law firm shares responsibility” because partner review occurred without policies and procedures governing responsible AI use. This point resonates with supervisory-duty concepts (even though the court did not formally rule on Rule 5.1).

3.3. Impact

This order establishes a clear Connecticut Supreme Court marker on generative AI in filings:

  • Rule 1.1 competence explicitly encompasses AI verification: Submitting AI-altered citations without independent verification constitutes incompetent representation in Connecticut, even if the legal propositions are substantively correct and there was no intent to mislead.
  • Sanctions will be real and public: The court imposed education requirements, significant donations, compliance reporting, and public posting, signalling that AI-citation failures are treated as integrity-threatening misconduct, not harmless error.
  • Institutional alignment with new Practice Book rules: The court connected its sanction decision to newly adopted statewide practice rules addressing generative AI risk—Practice Book § 4-9(b) (requiring independent verification of citations, legal authorities, or evidence produced by generative AI), Practice Book § 4-2(b) (signing constitutes certification of compliance), and the appellate counterparts Practice Book § 62-6(d) and § 85-2(11) (making noncompliance sanctionable). The decision reinforces that these rules are not aspirational; they are enforceable norms.
  • Firm governance expectations: By faulting the firm’s lack of AI policies/procedures, the decision encourages (and effectively pressures) Connecticut firms to implement AI governance: tool vetting, confidentiality protocols, verification checklists, training, and supervision.
  • Privilege/confidentiality risk elevated: The court’s competence discussion highlights that AI usage implicates confidentiality and privilege, particularly with “public access” platforms—pointing toward future disciplinary or evidentiary disputes when lawyers upload client materials to AI tools.

4. Complex Concepts Simplified

  • “Hallucinated” citations: When a generative AI system outputs a case name, citation, or quotation that does not exist or is incorrect, often presented with confidence. It is not merely a typo; it can be an invented authority.
  • Rule 1.1 (competence): Requires “legal knowledge, skill, thoroughness and preparation reasonably necessary” for the work. Here, “thoroughness” includes independently verifying every citation after AI involvement.
  • Sua sponte show cause order: The court initiated the sanctions inquiry on its own (not at a party’s request) and required the lawyer/firm to explain why sanctions should not be imposed.
  • Errata sheets: Formal correction filings used to fix errors in previously submitted documents.
  • Amici curiae: “Friends of the court” who are not parties but provide information or argument; here, they helped identify the hallucinated citations.
  • Reciprocal discipline: If a lawyer is sanctioned in one jurisdiction, other jurisdictions where the lawyer is admitted may impose corresponding discipline or require reporting/compliance steps.
  • Confidentiality/privilege/work product concerns with AI: If client information is entered into certain AI platforms, it may be stored, processed, or disclosed in ways that threaten confidentiality or privilege protections.

5. Conclusion

The Connecticut Supreme Court’s order in TOV Realty, LLC v. Suarez (and the companion Kosel Equity, LLC v. MacGregor) crystallizes a practical rule for modern practice: lawyers may use generative AI, but they remain personally responsible for the truth and accuracy of every citation and representation in filed documents. AI-assisted “hallucinations” are a competence violation under Rule 1.1 and warrant meaningful, public sanctions even absent intent to deceive.

By pairing individualized discipline (AI-focused CLE, monetary contributions supporting professional education, compliance reporting) with express reference to newly adopted Practice Book provisions requiring independent verification, the court both corrects the immediate misconduct and sets a forward-looking compliance baseline for Connecticut practitioners and firms navigating generative AI in litigation.