Exploring attribution bias in LLMs: Social influences and prompt-based mitigation

The growing use of Large Language Models (LLMs) into high-stakes domains such as education, law, and healthcare has increased the need to evaluate their alignment and potential biases in causal reasoning. Grounded in attribution theory and computational social science, this study examines whether LLMs exhibit human-like attributional biases, particularly self-serving bias and the fundamental attribution error. We evaluated five representative LLMs through a controlled experimental framework that manipulates event valence, attributional perspectives, and four key social dimensions: gender, age, education, and regional origin. Multivariate statistical analyses reveal that while LLMs consistently manifest self-serving bias, the fundamental attribution error is not uniformly observed across models. Notably, these attributional patterns are significantly modulated by age, education, and region, whereas gender-related effects remain negligible. The findings further highlight substantial inter-model variability, suggesting that technical heterogeneity inherently shapes the intensity of bias manifestation. Crucially, prompt-based interventions proved effective in reducing attributional extremity, thereby fostering greater fairness and neutrality in model outputs. These insights clarify the mechanisms by which LLMs internalize and reproduce social cognition patterns, offering a theoretical roadmap for socially responsible prompt design and algorithmic accountability in the ethical deployment of generative AI.