Reproduce “Bias Beware: The Impact of Cognitive Biases on LLM-Driven Product Recommendations” on OpenResearch

About this paper

Giorgos Filandrianos, Angeliki Dimitriou, Maria Lymperaiou, Konstantinos Thomas, Giorgos Stamou

This research systematically investigates how cognitive biases embedded in product descriptions influence large language model (LLM) product recommendations. It demonstrates that LLMs are susceptible to subtle psychological manipulations, with biases like social proof and discount framing increasing product visibility, while, counter-intuitively, scarcity and exclusivity decreased it across models and datasets.

Key insights

• LLMs are demonstrably susceptible to cognitive biases embedded in product descriptions, with these subtle cues often overriding factual product attributes in recommendations.
• Surprisingly, biases effective in human marketing (e.g., scarcity, exclusivity) can have an inverse, negative effect on LLM recommendations, reducing product visibility.
• The susceptibility to these biases does not correlate with LLM size or the presence of explicit reasoning modules, and an agnostic defense strategy proved largely ineffective.

Summary by alphaXiv.