| Abstract |
| South Korea is facing a serious mental health crisis, with a suicide rate approximately 2.4 times higher than the Organization for Economic Co-operation and Development average, while mental health service utilization remains low at 12.1%. As Large Language Models (LLMs), including ChatGPT, are increasingly filling this gap, concerns have emerged that LLM sycophancy may foster emotional dependency in a psychological counseling context. This study examines the perceived dependency reflected in user discourse rather than directly measuring sycophancy. A total of 1,476 Naver blog posts were collected using Selenium-based web crawling and analyzed via Latent Dirichlet Allocation topic modeling and KoELECTRA-based sentiment analysis. Five major topics were identified, with discussions primarily focusing on emotional support, empathy, and sharing personal concerns. The dependency-related discourse primarily appeared among users who explicitly recognized their reliance on ChatGPT. Moreover, sentiment analysis indicated that positive sentiment accounted for 65.6% of the data, while the “ChatGPT dependency” keyword group exhibited a relatively high proportion of negative sentiment (37.7%). These findings suggest that although sycophantic LLM responses can provide emotional support, they may also contribute to emotional dependency among some users, thereby providing empirical evidence to support the design of safer AI-based psychological counseling services. |
|
|
| Key Words |
| LLM Counseling, Perceived Dependency, User Discourse, LDA Topic Modeling, Sentiment Analysis, LLM 심리상담, 의존 인식, 사용자 담론, LDA 토픽모델링, 감성분석 |
|
|
 |
|
