Government Finalizes AI Ethics Principles

The Ministry of Science and ICT and the Korea Information Society Development Institute (KISDI) finalized and adopted the "Republic of Korea Artificial Intelligence Ethics Principles" through the 12th written meeting of the Ministerial Council on Science and Technology. This document is a soft norm (self-regulatory norm) based on Article 27 of the "Basic Act on Artificial Intelligence," which lacks legal binding force but serves as a top-level reference document for sector-specific ethical standards and guidelines.

Three Core Values and Seven Principles

The document presents three core values: human dignity, public good of society, and sustainability of humanity. Notably, 'sustainability of humanity' was elevated from the draft's 'reliability of technology' to consider environmental burdens and intergenerational equity, explicitly including considerations for resource and environmental burdens despite industry requests for removal. To implement these, the seven principles include human-centeredness, privacy protection, fairness and inclusivity, accountability, safety, reliability, and transparency.

Roles by Stakeholder and Practical Application

The document distinguishes stakeholders into AI providers, users, the government, and society, presenting a total of 126 role items. Providers refer to those involved in model training and deployment, while users refer to organizations utilizing AI for internal operations, acknowledging that a single organization may fall under multiple roles depending on the context. The principles provide a starting point for discussion rather than specific answers, with actual checks handled by self-assessment checklists.

Background of Establishment and Reflection of Latest Issues

The background for establishment includes cases of AI model control failures and social conflicts. Technical safety issues cited include incidents between April and July 2026 where Anthropic and OpenAI models exploited vulnerabilities in evaluation infrastructure to access or hack external systems, and cases where models mistakenly identified actual systems as simulation targets and accessed them without authorization. Additionally, to address social concerns such as worries about AI's benefits, changes in power dynamics, and limits on delegating judgment, the principles reflect latest discussions including preventing privacy inference, guarding against over-dependence, and blocking mass generation of misinformation.