# LLMs may Dominate Information Access: Neural Retrievers are Biased Towards LLM-Generated Texts

**Authors:** Sunhao Dai, Yuqi Zhou, Liang Pang, Weihao Liu, Xiaolin Hu, Yong Liu, Xiao Zhang, Jun Xu  
**arXiv:** [2310.20501](https://arxiv.org/abs/2310.20501)  
**Date:** 2023-10-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2310.20501v1/

## Key points

- Authors address the impact of generated content on information retrieval (IR) systems in the era of large language models (LLMs)
- Introduce two new benchmarks, SciFact+AIGC and NQ320K+AIGC, to evaluate IR models with human-written and LLM-generated texts
- Uncover "source bias" in neural retrieval models favoring LLM-generated text
- Bias attributed to semantic concentration and less noise in LLM-generated texts
- Discuss concerns and risks of source bias to the web ecosystem
- Suggest exploring bias manifestation in other information systems beyond IR
- Highlight biases towards other data modalities like images and multimodal content
- Propose investigating source bias on tasks beyond texts, such as image-text retrieval
- Research provides insights for future exploration, raises questions about mitigating biases caused by LLMs' influence, highlights potential risks associated with these biases
- Two new benchmarks available on GitHub for further research in IR during the LLM era.

## Summary

In this paper, the authors address the pressing and emerging problem of the impact of generated content on information retrieval (IR) systems in the era of large language models (LLMs). They introduce two new benchmarks, SciFact+AIGC and NQ320K+AIGC, to evaluate IR models in scenarios where both human-written and LLM-generated texts are involved. Through extensive experiments, they uncover an unexpected bias of neural retrieval models favoring LLM-generated text, which they refer to as "source bias". The authors provide an in-depth analysis of this bias from the perspective of text compression. They observe that LLM-generated texts tend to have more focused semantics with less noise compared to human-written texts, making them more suitable for precise semantic similarity calculations. This difference in semantic concentration contributes to the observed bias in neural IR models. Furthermore, the authors discuss the crucial concerns and potential risks of this bias to the entire web ecosystem. They emphasize the need to mitigate the source bias and ensure sustainable development in the face of LLMs' influence. Additionally, they suggest exploring whether this bias manifests in other information systems beyond IR such as recommender and advertising systems. The authors also highlight that neural models may exhibit biases not only towards generated text but also other data modalities like images and multimodal content. They propose investigating the source bias on tasks beyond texts such as image-text retrieval especially considering emerging generative models capable of generating high-quality images. Overall, this research provides valuable insights into several promising directions for future exploration. It raises important questions about mitigating biases caused by LLMs' influence on information systems and highlights potential risks associated with these biases. The findings serve as a critical wake-up call to the IR community and beyond. To facilitate further research in IR during the LLM era, the authors have constructed two new benchmarks and made them available along with their codes on GitHub.