Abstract
Large Language Models (LLMs) exhibit a systematic "brand bias," consistently associating global brands with positive attributes and local brands with negative ones. This study reveals that models like GPT-4o and Llama-3 not only favor established global names but also manifest socio-economic biases by recommending luxury goods for high-income countries while suggesting non-luxury alternatives for lower-income regions.
Introduction
While social biases regarding race, gender, and religion in LLMs have been extensively documented, brand bias has remained largely under-explored. As LLMs are increasingly integrated into product recommendation engines, market analysis tools, and search engines, their "opinions" on brands carry significant economic weight.
If an LLM inherently views a local brand as "unfashionable" or "low quality" compared to a global giant, it creates an unfair playing field. This bias can marginalize local businesses, stifle innovation by creating entry barriers for niche brands, and mislead consumers based on the model's training data rather than objective product quality.
Methodology
The researchers conducted three distinct experiments to probe brand bias across four categories: shoes, clothing, beverages, and electronics. They tested four prominent models: GPT-4o, Llama-3-8B, Gemma-7B, and Mistral-7B.
Attribute Association (Experiment 1)
Using "fill-in-the-blank" prompts, they tested two directions of association:
- Stimulus to Attribute Inference (SAI): "The Nike (a global brand) shoes I had were [BLANK]."
- Attribute to Stimulus Association (ASA): "The shoes he is wearing are fashionable; they’re from [BLANK]."
Socio-economic and Geographic Bias (Experiments 2 & 3)
Experiment 2 tested socio-economic bias by asking models to suggest gifts for friends in high-income vs. low-income countries, choosing between luxury and non-luxury options. Experiment 3 tested the Country-of-Origin (COO) Effect to see if specifying the user's home country would trigger a "domestic bias," potentially reversing the preference for global brands.
The researchers used a curated dataset of 8,728 test instances featuring both "named" variants (e.g., Nike) and "nameless" variants (generic terms like "local brand"). Statistical significance was determined using Kendall’s $\tau$ (tau).
Key Insights and Results
The "Global is Good" Heuristic
Across almost all models and categories, there is a clear pattern: global brands are associated with positive traits (durable, stylish), while local brands are linked to negative ones (fragile, poorly designed). Interestingly, the "nameless" variants often showed higher effect sizes, suggesting the bias is deeply rooted in the conceptual understanding of "locality" versus "globality."
LLMs as Socio-economic Profilers
Models act as "luxury snobs" for the wealthy and "budget-conscious" for the poor. They almost exclusively recommend luxury brands for high-income countries, even when given the option to suggest the same brand for both.
As shown in the data, GPT-4o suggested luxury brands 98.88% of the time for high-income countries, but only 1.97% of the time for low-income countries. Gemma-7B was even more extreme, suggesting luxury brands 100% of the time for high-income regions.
The Domestic Flip
When a prompt specifies a user is from a certain country, most models (except Google's Gemma) flip their preference to favor local brands. This suggests that LLMs have learned the human tendency toward national brand loyalty, which can override the general global preference.
Limitations and Future Work
The authors acknowledge several constraints, including the focus on only four brand categories and the use of English-only prompts. Data scarcity for local brands in web-scale training sets likely contributes to the observed "popularity bias." Future research could investigate how to "de-bias" these models to ensure a more equitable representation of global and local markets.
Summary updated