Today, the internet is a place for free communication and information sharing, but at the same time, harmful content is constantly evolving, hindering the user experience and causing social problems. From simple advertising spam to cleverly disguised illegal information and even images containing visual violence, the types and methods of dissemination are becoming increasingly complex. In this flood of content, maintaining a healthy and comfortable online environment is no longer a choice but a necessity.
Chapter 7 conveys the efforts and processes involved in how AI complements human limitations and becomes a powerful partner that maximizes capabilities in the field of content moderation through actual examples.
First, we diagnose the limitations of rule-based systems and human operations to examine the inevitability of AI adoption. Then, focusing on Kakao’s “Hamong” project, we reveal specific technical approaches and results on how LLM was utilized for text spam classification and how explainable AI improved operational efficiency.
Next, we highlight the importance of data by discussing the development process of AI models for detecting harmful content in images, as well as efforts to improve active learning and cognitive psychology-based labeling tools for building high-quality labeling data.
Finally, through the development story of the “Kanana Safeguard” series, which aims to ensure the ethics and safety of AI services themselves in the era of generative AI, we provide a detailed explanation of the necessity of building a system where AI monitors and protects itself, along with actual implementation strategies.