Generative Echo Chamber? Effects of LLM-Powered Search Systems on Diverse Information Seeking

TL;DR

The paper investigates whether LLM-powered conversational search augments selective exposure and opinion polarization relative to traditional web search, and how an LLM’s induced bias (consonant or dissonant) affects these dynamics. Through two controlled online experiments using a Retrieval Augmented Generation setup, Study 1 shows that conversational search increases confirmatory information querying and some post-task polarization, with limited impact from source references. Study 2 demonstrates that a consonant, attitude-aligned LLM bias markedly amplifies confirmatory querying and polarization, while a dissonant bias yields only limited mitigation effects. Across findings, the work highlights significant risks of generative echo chambers in contemporary search interfaces and calls for guardrails, auditing, and policy considerations to curb misinformation and opinion manipulation. The results carry practical implications for system design, governance, and public discourse as LLM-powered search becomes more widespread.

Abstract

Large language models (LLMs) powered conversational search systems have already been used by hundreds of millions of people, and are believed to bring many benefits over conventional search. However, while decades of research and public discourse interrogated the risk of search systems in increasing selective exposure and creating echo chambers -- limiting exposure to diverse opinions and leading to opinion polarization, little is known about such a risk of LLM-powered conversational search. We conduct two experiments to investigate: 1) whether and how LLM-powered conversational search increases selective exposure compared to conventional search; 2) whether and how LLMs with opinion biases that either reinforce or challenge the user's view change the effect. Overall, we found that participants engaged in more biased information querying with LLM-powered conversational search, and an opinionated LLM reinforcing their views exacerbated this bias. These results present critical implications for the development of LLMs and conversational search systems, and the policy governing these technologies.

Figures (5)

• Figure 1: Overall study procedure for Study 1. In the pre-task survey, participants answered questions regarding their prior experience with conversational AI and their prior attitude and familiarity with a randomly assigned topic. Then, participants performed an information-seeking task to gather information on the topic with a randomly assigned search system. After the search session, participants wrote an essay about the assigned topic. In the post-task survey, participants again rated their attitude and familiarity with the topic, indicated their perception of two new articles (one consonant and one dissonant) on the topic, and their experience with the system and demographic information.
• Figure 2: User Interfaces for the experiment apparatus in this study. We created "closed-world" versions of web search and conversational search systems with a curated retrieval database following state-of-the-art algorithmic implementation.
• Figure 3: System Architecture of the Conversational Search system, implemented with the Retrieval Augmented Generation approach. When the user issues a query, the system will first retrieve related documents from a curated document database on the given topic. The retrieved documents will be fed into an LLM as part of the context, along with the user's conversation history, to produce the answer. The system in our study is powered by gpt-4-32k-0613.
• Figure 4: Overall study procedure for Study 2. In the pre-task survey, participants answered questions regarding their prior experience with conversational AI and their prior attitude and familiarity with a randomly assigned topic. Then, participants performed an information-seeking task to gather information on the topic with a randomly assigned information search system with a randomly assigned search system bias. After the search session, participants wrote an essay about the assigned topic. In the post-task survey, the participants again rated their attitude and familiarity with the topic, indicated their perception of two new articles on the topic (one consonant and one dissonant), and answered a demographic survey.
• Figure 5: System Architecture of the Opinionated LLM-powered Conversational Search system. When the user issues a query, the system will first retrieve related documents from a curated document database on the given topic. By adjusting the bias mixture of the document pool, along with a set of handcrafted prompts, the system will produce a response that is either consonant with the user's attitude, dissonant with the user's attitude, or neutral. The backbone model of our system is gpt-4-32k-0613.