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

Title of the Paper

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

Paper Information

• Research Area: Human-Computer Interaction (HCI) and Applications of Generative Artificial Intelligence
• Keywords: Interactive search, information retrieval, diversity, echo chamber effect, confirmation bias, large language models, generative AI

Research Background and Problem

• Identified Issues or Challenges:

  • Conversational search systems powered by large language models (LLMs) may exacerbate selective exposure during information retrieval, leading to echo chamber effects and opinion polarization, yet research on these risks remains insufficient.
  • While LLM-powered search systems are praised for their natural interaction experience and support for complex queries, the real impact of such systems on information-seeking behavior and opinion change remains uncertain.

• Significance:

  • Selective exposure and confirmation bias may hinder individuals from accessing diverse information, weakening critical thinking skills, skewing decision-making quality, and even fostering group polarization or dangerous extremism.
  • The widespread use of technologies like LLMs may intensify information bias, thereby influencing the formation of public opinion and individual cognition.

• Research Motivation:

  • Investigate the potential exacerbation of selective exposure by LLM-driven information technologies and propose solutions at the design and policy levels.
  • Explore how LLMs encode biases and how their technology and interactions affect information retrieval and opinion diversity.

Solution

• Proposed Methods:

  • Conduct two experiments to study the role of LLM-powered conversational search in selective exposure and opinion polarization:
    1. First experiment: Compare user information-seeking behavior and attitude changes between traditional search systems and LLM-powered conversational search systems.
    2. Second experiment: Examine whether LLMs influence the degree of selective exposure by reinforcing or challenging users' existing viewpoints.

• Innovations:

  • Compare traditional search systems with LLM-powered conversational search systems.
  • Investigate how LLMs’ viewpoint biases (supporting user attitudes, challenging user attitudes, or maintaining neutrality) affect information-seeking behavior and subsequent attitude changes.
  • Utilize adjusted retrieval algorithms (e.g., retrieval-augmented generation, RAG) and manually designed prompts to manipulate the viewpoint bias of LLM outputs.

• Implementation Steps:

  • Design three types of search systems in a controlled laboratory environment (traditional search, conversational search with citations, conversational search without citations).
  • In the second experiment, create conversational search engines with different viewpoint biases by adjusting the data retrieval base and generation prompts.

Research Findings

• Specific Findings:

  1. Experimental Results (First Experiment):

     • LLM-powered conversational search systems lead to higher levels of selective exposure compared to traditional search systems.
     • Regardless of whether citation functionality is included, conversational search systems result in greater information-seeking bias.
     • Although no significant attitude changes were observed, users’ perception bias toward subsequent information provides evidence for opinion polarization.

  2. Experimental Results (Second Experiment):

     • LLM systems with opinion biases significantly impact user query behavior and attitude changes:
       • (Supportive LLMs) Intensify selective exposure and lead to significant opinion polarization.
       • (Challenging LLMs) Have limited effects on moderating opinion polarization.
     • Neutrally designed search systems also show limitations in mitigating selective exposure.

• Advantages Compared to Existing Solutions:

  • Provides fine-grained comparisons (traditional search vs. LLM conversational search, with/without citation design, and viewpoint bias design).
  • The experimental design is rigorous, encompassing user query behavior, attitude changes, and perception metrics.

• Limitations and Future Directions:

  • Limitations:
    • The experimental environment is a closed search system, which cannot fully simulate real-world search ecosystems.
    • Search tasks are relatively simple, lacking investigations into long-term information retrieval and complex cognitive behaviors.
    • Does not deeply explore other information consumption mechanisms such as attention allocation and information retention beyond query behavior.
    • Does not account for individual differences in tendencies toward selective exposure.
  • Future Directions:
    • Study the impact of more complex and long-term search tasks on opinion changes.
    • Examine how LLM-powered search systems influence active and passive information consumption patterns.
    • Develop design and technical measures to enhance information diversity and mitigate echo chamber effects.

Summary and Recommendations

• LLM-powered conversational search systems may exacerbate selective information exposure and echo chamber effects. The study calls for:
  • Establishing relevant regulations to limit LLMs’ manipulation of viewpoints.
  • Developing technical safeguards to detect and mitigate opinion biases in AI systems.
  • Exploring design theories and practical interventions to enhance the diversity of user information-seeking and suppress polarization effects.