Abstract

AI hallucination — the generation of plausible but factually incorrect outputs by large language models — poses a distinctive challenge for legal practice. Unlike factual errors in other professional contexts, hallucinated legal citations, fabricated case holdings, and invented statutory provisions can be submitted to courts, relied upon by clients, and incorporated into transactional documents with potentially severe consequences. The cases of Mata v. Avianca and Park v. Kim have brought attorney sanctions for AI-generated hallucinations into public view, but they represent only the visible surface of a broader systemic problem.

This paper examines the hallucination problem through the lens of legal ethics and professional responsibility. We analyze how the duty of competence under Model Rule 1.1, the duty of candor toward tribunals under Model Rule 3.3, and the duty of supervision under Model Rule 5.1 interact with the probabilistic and non-deterministic nature of large language model outputs. We argue that hallucination is not merely a technical defect to be engineered away but a structural feature of current AI architectures that requires affirmative professional responsibility responses.

We propose a framework of AI output verification obligations calibrated to the stakes of the matter, the nature of the AI-generated content, and the availability of verification tools. The paper concludes with model protocols for attorney verification of AI outputs and recommendations for bar associations developing guidance on AI use in legal practice.

Full Paper

This paper is published on the Social Science Research Network (SSRN). To read the full text, download the PDF, or cite this work, please visit the SSRN abstract page:

→ Read the Full Paper on SSRN (Abstract ID: 6736661)

Published: May 2026 — Authors: Austin, Morris & Das — View all research papers