This paper develops and validates a new questionnaire to assess dependence or addiction behavior on large-scale language models (LLMs). In order to overcome the conceptual limitations of previous studies that applied existing behavioral addiction symptoms to the LLM usage environment, we present a new perspective that considers the subtle aspects of LLM and human interaction. The 12-item questionnaire (LLM-D12) developed with 526 participants in the UK was confirmed to consist of two factor structures, 'instrumental dependence' and 'relational dependence', through exploratory and confirmatory factor analysis. Instrumental dependence reflects the degree of dependence on LLMs in decision-making and cognitive tasks, and relational dependence reflects the tendency to perceive LLMs as socially meaningful, perceptive, or companion-like beings. LLM-D12 showed excellent internal consistency and clear discriminant validity, and the conceptual basis and distinction of the two subscales were confirmed through external validation. The psychometric properties and structure of the LLM-D12 scale were interpreted considering that although LLM dependence does not necessarily indicate dysfunction, it may reflect a level of dependence that may be problematic in certain situations.
