Extraction of Subjective Information from Large Language Models
Atsuya Kobayashi, Saneyasu Yamaguchi · 2024
Large Language Models (LLMs) have been advancing natural language processing technologies for several years. Especially in the last year, generative AI (Artificial Intelligence) models have improved remarkably and attracted attention. Then, there has been a great discussion about these models, such as extracting the knowledge in these models. However, these existing works were mainly for the extraction of objective information. In this paper, we study the extraction of subjective information from LLMs by focusing on GPT, Gemini Pro, and Claude2. We then show that such information extraction can be severely limited in LLMs, but that information extraction is possible from some models.