This master’s thesis investigates how cultural biases manifest in the outputs of Large Language Models (LLM) trained on Western versus Chinese text corpora, particularly in response to politically and socially sensitive prompts. Despite the increasing adoption of LLMs in global contexts, current models often reflect culturally skewed patterns due to imbalances in training data.

The study aims to answer three core research questions regarding sentiment, rhetorical framing, and reasoning patterns in LLM-generated responses, and whether prompt engineering can reveal latent cultural assumptions. To achieve this, the research adopts a Design Science Research (DSR) methodology and develops a novel comparative framework grounded in Hofstede’s and the GLOBE cultural dimensions, augmented by contemporary sentiment and argumentation metrics. A selection of six LLMs, three Western-based and three Chinese-based, will be tested using culturally sensitive prompts, with outputs analysed for sentiment, linguistic structure, and reasoning logic.

The study contributes both to theory, by extending cultural bias research in Natural Language Processing, and to practice, by offering a reproducible framework for bias evaluation in AI applications. Findings will support the development of culturally inclusive AI systems and highlight critical considerations for cross-cultural model alignment.