When KeaneAdvisors.AI began building the Attorney AI Sanctions Tracker, one data point stood out immediately: Rule 3.3 — Candor to the Tribunal — accounts for the majority of documented AI-related disciplinary actions in the United States. The pattern is consistent across jurisdictions, practice areas, and firm sizes.
Understanding why this rule fails so frequently is not an academic exercise. It is the single most important risk management question any attorney using AI should be asking.
The Mechanism of Failure
Generative AI tools like ChatGPT, Claude, and similar large language models do not retrieve information — they generate it. When asked to produce a case citation, the model produces text that looks like a case citation. It has no awareness of whether the case exists, whether the citation is accurate, or whether the holding it attributes to the case is real.
This is not a bug that will be fixed in the next version. It is a structural feature of how these models work. They are optimized to produce plausible text, not accurate text. In most contexts, the difference is tolerable. In a legal brief submitted to a federal court, it is not.
The attorneys who have been sanctioned under Rule 3.3 were not, in most cases, trying to deceive the court. They were using AI as they would use a capable research assistant — asking it to find authority for a position, receiving what looked like a well-formatted response with proper citations, and submitting that response without independent verification. The citations didn't exist. The courts noticed.
What the Sanctions Look Like
The sanctions KeaneAdvisors.AI has documented range from formal written reprimands to five-figure fines to mandatory CLE requirements focused on AI literacy. In several cases, courts have ordered attorneys to submit sworn certifications in future filings attesting that they have independently verified every citation. In at least two cases, referrals to state bar disciplinary authorities have been made.
The pattern that emerges from reviewing these orders: courts are not treating AI-assisted citation fabrication as a technical error. They are treating it as a failure of the attorney's fundamental professional obligation to verify the authority they are presenting to the court.
The Supervision Dimension
Rule 3.3 violations that arise from AI use often also implicate Rules 5.1 and 5.3 — the supervision rules. When an associate or paralegal uses AI to generate a brief that a supervising attorney then submits without reviewing the citations, the supervising attorney has a problem. Rule 5.1 makes partners responsible for establishing systems that ensure subordinates comply with professional conduct rules. A firm that allows AI-assisted drafting without a citation verification protocol has no such system.
This is where the governance gap creates compounding exposure. The initial failure is the unverified AI output. The secondary failure is the absence of a supervision protocol that would have caught it.
Building a Rule 3.3-Compliant AI Workflow
A compliant workflow for AI-assisted legal research has three components:
First, every citation generated by AI must be independently verified against a reliable legal database — Westlaw, Lexis, or equivalent — before it appears in any filing. This is not optional and it is not delegable to AI.
Second, the verification step must be documented. Not necessarily in the filing itself, but internally, as part of the firm's record of how AI-generated work product was reviewed. If a court asks, the firm should be able to show the verification occurred.
Third, the attorney who signs the filing must personally verify, or personally supervise the verification of, every citation in that filing. Supervision by AI is not supervision.
These three components are part of what the KeaneAdvisors.AI Phase 0 Assessment measures under the Output Verification and Supervision & Accountability scoring domains.