Towards Explainable AI in Advertising Compliance: A Logic-Driven Two-Stage Multi-Agent System

Makoto Takamatsu, Modong Tan, Ryotaro Sugiura, Mioko Hirose, Naoya Takizawa

Keywords: Advertising_Compliance/Multi_Agent_System/Legal_Risk_Assessment/Logical_Analysis/Integration_Agent

TL;DR: AI with Logic of Thought improved ad compliance checking, raising accuracy from 72% to 90% and F1 score from 60% to 85% for Japanese law violations.

Abstract: This study addresses the challenges of transparency and explainability in AI-driven advertising compliance review by proposing an autonomous AI agent that integrates Logic of Thought (LoT) to clarify logical propositions with Reflection Prompting (RP) for self-correction mechanisms. The proposed agent identifies legal violations in the advertising content and expresses them as logical formulae, enabling visualization of reasoning processes previously unattainable in conventional models. We evaluated the agent using 100 experimental cases based on administrative guidance and court decisions related to Japan's Act against Unjustifiable Premiums and Misleading Representations and the Unfair Competition Prevention Act. The results showed significant performance improvements, with accuracy increasing from 72\% to 90\% in identifying the relevant laws and number of articles, and the F1 score improving from 60\% to 85\%. The results of this study suggest the potential for flexible adaptation to various legal frameworks across different countries, indicating applicability beyond advertising reviews for general products, financial products, and pharmaceuticals. It also highlights the possibility of application to other tasks requiring explainable and detailed legal compliance assessments.

Submission Number: 13