CrowdWorks begins development of generative AI reliability evaluation framework

2024.07.17. The Fair Economy — AI tech company CrowdWorks (President Kim U-seung) announced on the 17th that it has begun the "Generative AI Trustworthiness Evaluation Framework Development" project to lay the foundation for AI reliability.

CrowdWorks recently won the "Practical Approaches to Generative AI Trustworthiness Evaluation" project led by the Korea Telecommunications Technology Association (TTA). It will participate in a consortium with natural language processing solution startup TeddySum (President Ham Yeong-gyun), Professor Kim Han-saem's research team at Yonsei University, and others.

According to CrowdWorks, this project aims to establish guidelines for ensuring the safety and trustworthiness of generative AI, with the goal of "developing and disseminating a generative AI trustworthiness evaluation framework." It seeks to systematize definitions of key risk factors for generative AI, which are defined differently across studies, and develop a practical trustworthiness evaluation framework that considers social and cultural characteristics to minimize the risks and side effects of generative AI.

The scope of the project includes: ▲ systematizing the definition and classification of generative AI risks; ▲ designing a risk-assessment matrix mapping structure; ▲ collecting red-team testing techniques and creating educational materials; ▲ designing evaluation mechanisms; ▲ and conducting pilot applications to assess the practicality of the evaluation approach.

CrowdWorks, leveraging its extensive experience in building training data and validating models for LLM models such as HyperCLOVA X and EXAONE, and its technical expertise in developing the sLLM WorksOne, will be responsible for developing educational materials and demonstrating the generative AI evaluation framework to promote its adoption.

To this end, the company will collect domestic and international red-team testing techniques, derive malicious prompt characteristics and patterns based on data, and prepare a red-team training guide that includes analysis of attack prompting techniques and types. Additionally, it will conduct real attacks on domestic AI models to demonstrate the generative AI trustworthiness evaluation framework guide, and propose improvements and response measures.

CrowdWorks is expanding the breadth and leadership of AI technology based on its unique data technology competitiveness and differentiated expertise built within the AI ecosystem. In 2022, two data labeling technologies developed by CrowdWorks were adopted as group standards by the Telecommunications Technology Association (TTA), and the group standards "Task Design Considerations" and "Result Data Specification Guidelines" have become the foundation for the activation of the domestic data industry.

Kim U-seung, CEO of CrowdWorks, said, "The AI reliability and safety evaluation framework will accelerate AI evaluation standardization and enhance AI interoperability and compatibility, thereby speeding up the development of the AI industry. We will do our best to secure leadership and competitiveness in the AI reliability technology sector and help Korea leap forward as an AI powerhouse."