AI Safety Assessment

The Korea AI Safety Institute aims to systematically define and evaluate AI risks that require national-level management to ensure the safety of advanced AI. We carry out the following main tasks:

  • AI risk identification and classification

  • AI safety framework development

  • AI safety evaluation

  • Establishment of infrastructure for AI safety evaluation

  • AI risk identification and classification

    • Based on global discussions*, we define AI risks that require intensive management at the national level.

      Examples of global discussions:

      1. AI models that could potentially support the development, production, or acquisition of chemical or biological weapons
      2. AI models that could potentially circumvent human oversight by bypassing safety mechanisms, engaging in manipulation and deception, or autonomously replicating without explicit human consent

  • AI safety framework development

    • In collaboration with industries, universities, and research institutes, we develop AI safety evaluation frameworks (including indicators, standards, and methods) tailored to different AI risks, while also devising measures for safety evaluation and risk mitigation.

      If a certain AI technology has the potential to pose serious risks, we develop and share a reliable evaluation framework for the model.

  • AI safety evaluation and infrastructure establishment

    • We support AI safety evaluation by establishing an evaluation infrastructure, including building AI safety evaluation datasets and developing evaluation tools.

      We collect extensive empirical data related to potential risks posed by AI and leverage it to develop and implement tools and infrastructure that facilitate effective evaluations.