Abstract: Large Language Models (LLMs) have the potential to support policymaking, but they have inherent risks. While LLMs have demonstrated strong capabilities in generating recommendations across domains, concerns remain regarding their tendency to reflect Global North perspectives that may not align with developing countries. This study investigates contextual bias in climate policy recommendations generated by LLMs, with a specific focus on their applicability within South Africa. Supporting analyses include exploratory data analysis, sentiment analysis, topic modelling, and semantic analysis on climate-related newspaper articles from South Africa and the United Kingdom. Using scenario-based prompts, three LLMs (Cohere Aya Expanse, Mistral and Meta Llama 3.3 70B) generated policy recommendations informed by the South African and United Kingdom datasets. Generated recommendations are assessed using a contextual bias and feasibility evaluation framework developed from climate justice, algorithmic bias, and policy evaluation literature. A logistic regression classifier was then trained to determine whether generated policies were applicable to South Africa. The findings highlight the challenges of generating region-specific policy recommendations due to limited data and variations in climate change and policy lexicon.