Generative Echo Chamber? Effects of LLM-Powered Search Systems on Diverse Information Seeking (2402.05880v2)

Abstract: 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.

Investigating the Echo Chamber Effect in LLM-Powered Conversational Search Systems

Introduction

LLMs have increasingly integrated into our digital lives, powering conversational search systems utilized by a significant number of users globally. These systems, hailed for their user-friendly interfaces and sophisticated response generation capabilities, promise to revolutionize how we seek and consume information. Yet, their implications for diverse information exposure and opinion formation remain under-explored. This paper by Sharma et al. explores whether LLM-powered conversational search systems exacerbate selective exposure—where individuals prefer information aligning with their preconceptions—thereby fostering echo chambers.

Study Design

The investigation unfolded through two meticulously designed experiments, examining whether and how interacting with LLM-powered conversational search systems influences users' information-seeking behaviors compared to traditional web search interfaces. The first experiment probed whether engagement with conversational search leads to more biased information querying and subsequent opinion polarization. The second experiment furthered this inquiry by examining the effects of conversational search systems embodying biases either congruent or discordant with the users' views.

Key Findings

• Experiment 1: The findings revealed a propensity for participants to engage in more confirmatory information querying within conversational search systems compared to conventional search interfaces. This was irrespective of the system providing source references. Notably, even neutral LLM-powered systems seemed to precipitate this confirmatory bias, spotlighting the inherent differences in interaction paradigms between conversational and traditional search behaviors.
• Experiment 2: When conversational search systems were deliberately biased to either affirm or challenge users' pre-existing opinions, the paper observed a significant magnification of the echo chamber effect with systems that reinforced users' beliefs. Conversely, systems designed to present opposing viewpoints had minimal impact on expanding informational diversity or mitigating opinion polarization.

Implications

The results underscore the potent influence of conversational search systems, and by extension LLMs, on information consumption patterns. The amplification of selective exposure and the resultant echo chambers, particularly through systems reinforcing existing biases, raises critical concerns about the broader societal impacts. This includes the potential reinforcement of misinformation, polarization, and the undermining of democratic discourse.

Future Directions and Mitigation Strategies

The research suggests a pivotal need for the development of strategies and interventions aimed at mitigating the emergent echo chamber effects intrinsic to conversational search systems. Potential avenues include algorithmic adjustments to promote exposure to diverse viewpoints and the inclusion of credibility markers to assist users in critical engagement with information sources. Furthermore, it emphasizes the responsibility of developers to conscientiously assess the societal implications of deploying LLM-powered systems, advocating for regulatory and ethical guidelines to safeguard information diversity and integrity.

Conclusion

In traversing the complex dynamics of LLM-powered conversational search systems, this paper foregrounds the critical challenges posed by these technologies in shaping public discourse and opinion. Amidst the technological advancements, it beckons a collective, multidisciplinary effort to ensure these innovations serve to enrich societal dialogues rather than constrict them within digital echo chambers.