Large language models (LLMs) powered conversational search systems have been widely adopted, but their impact on information diversity and selective exposure remains underexplored. This study investigates how LLM-powered conversational search systems compare to conventional search systems in terms of selective exposure and opinion polarization. Two experiments were conducted: the first compared information-seeking behaviors and attitude changes between conventional search and LLM-powered conversational search systems; the second explored how opinion-biased LLMs affect selective exposure. Results showed that participants engaged in more biased information querying with LLM-powered systems, and opinionated LLMs reinforced existing views, exacerbating bias. These findings highlight the need for careful design and regulation of LLM-powered search systems to mitigate echo chamber effects and promote information diversity. The study underscores the importance of addressing biases in LLMs and conversational search systems to ensure balanced information consumption and reduce opinion polarization.Large language models (LLMs) powered conversational search systems have been widely adopted, but their impact on information diversity and selective exposure remains underexplored. This study investigates how LLM-powered conversational search systems compare to conventional search systems in terms of selective exposure and opinion polarization. Two experiments were conducted: the first compared information-seeking behaviors and attitude changes between conventional search and LLM-powered conversational search systems; the second explored how opinion-biased LLMs affect selective exposure. Results showed that participants engaged in more biased information querying with LLM-powered systems, and opinionated LLMs reinforced existing views, exacerbating bias. These findings highlight the need for careful design and regulation of LLM-powered search systems to mitigate echo chamber effects and promote information diversity. The study underscores the importance of addressing biases in LLMs and conversational search systems to ensure balanced information consumption and reduce opinion polarization.