A Study on the Framework for Evaluating the Ethics and Trustworthiness of Generative AI

Current AI evaluation methods are inadequate: The paper finds that existing evaluation approaches focus primarily on performance and accuracy metrics, but fail to address critical ethical concerns like bias, harmfulness, copyright infringement, privacy violations, and "hallucination" (when AI generates false information). This gap means we're not properly assessing the real-world harms generative AI systems can cause.

A comprehensive evaluation framework is needed: The study proposes assessing generative AI across 11 key dimensions including fairness, transparency, accountability, safety, privacy, accuracy, consistency, robustness, explainability, copyright protection, and source traceability. This multi-dimensional approach recognizes that trustworthiness requires evaluating far more than just technical performance.

Human-centered criteria must be central: The paper emphasizes that evaluation frameworks should incorporate "human-centered criteria that also reflect social impact," moving beyond purely technical assessments to consider how AI systems affect people and society. This represents a fundamental shift in how we should think about AI quality.

Global policy approaches vary significantly: The comparative analysis of AI ethics policies across South Korea, the United States, the European Union, and China reveals different regulatory priorities and approaches, suggesting there's no single universal standard yet for responsible AI governance.

Ethics assessment must span the entire AI lifecycle: Rather than evaluating AI systems only at deployment, the framework applies across the full development cycle, enabling organizations to identify and manage ethical risks throughout design, development, and real-world use.