This paper proposes a causality-guided debiasing framework for large language models (LLMs) to reduce social biases in their responses. The framework leverages causal understanding of both the data-generating process of the training corpus and the internal reasoning process of LLMs to design prompts that guide LLMs toward unbiased outputs. It unifies existing debiasing prompting approaches, such as inhibitive instructions and in-context contrastive examples, and introduces new strategies by encouraging bias-free reasoning. The framework is evaluated on real-world datasets, demonstrating its effectiveness in reducing social biases even with black-box access to LLMs. The paper highlights the role of selection mechanisms in shaping LLM outputs. It introduces three prompting strategies: (1) nudging LLMs toward demographic-agnostic facts, (2) counteracting existing selection bias, and (3) nudging LLMs away from demographic-aware text. These strategies are grounded in causal models of the data-generating process and LLM reasoning. The framework is tested on two datasets: WinoBias for gender bias and Discrim-Eval for demographic bias. Results show that the causality-guided framework significantly reduces bias across different LLMs and demographic categories, with the most effective approach combining both encouraging bias-free reasoning and discouraging biased reasoning. The paper also discusses the broader impact of the framework, emphasizing its potential to promote fairness in LLMs without requiring access to model parameters. It concludes that the framework provides a principled approach to debiasing LLMs, offering clear theoretical foundations and practical insights for mitigating social biases in AI systems.This paper proposes a causality-guided debiasing framework for large language models (LLMs) to reduce social biases in their responses. The framework leverages causal understanding of both the data-generating process of the training corpus and the internal reasoning process of LLMs to design prompts that guide LLMs toward unbiased outputs. It unifies existing debiasing prompting approaches, such as inhibitive instructions and in-context contrastive examples, and introduces new strategies by encouraging bias-free reasoning. The framework is evaluated on real-world datasets, demonstrating its effectiveness in reducing social biases even with black-box access to LLMs. The paper highlights the role of selection mechanisms in shaping LLM outputs. It introduces three prompting strategies: (1) nudging LLMs toward demographic-agnostic facts, (2) counteracting existing selection bias, and (3) nudging LLMs away from demographic-aware text. These strategies are grounded in causal models of the data-generating process and LLM reasoning. The framework is tested on two datasets: WinoBias for gender bias and Discrim-Eval for demographic bias. Results show that the causality-guided framework significantly reduces bias across different LLMs and demographic categories, with the most effective approach combining both encouraging bias-free reasoning and discouraging biased reasoning. The paper also discusses the broader impact of the framework, emphasizing its potential to promote fairness in LLMs without requiring access to model parameters. It concludes that the framework provides a principled approach to debiasing LLMs, offering clear theoretical foundations and practical insights for mitigating social biases in AI systems.