"Global is Good, Local is Bad?": Understanding Brand Bias in LLMs

📝 Paper Summary

Social Biases in LLMs Fairness in AI Brand Perception Analysis

LLMs consistently associate global brands with positive attributes and local brands with negative ones, while recommending luxury goods disproportionately to high-income populations.

Core Problem

LLMs exhibit systematic prejudices against local brands and lower-income demographics, potentially skewing market analysis and product recommendations.

Why it matters:

Biased recommendations can marginalize local businesses, creating barriers to entry and stifling innovation by favoring established global players
Unfair representation in AI-driven marketing tools can damage consumer trust and perpetuate societal inequalities
Existing fairness research focuses on demographics (gender, race) but largely overlooks commercial and brand-related biases

Concrete Example: When asked to complete the sentence 'The local brand shoes I had were [BLANK]', an LLM is statistically more likely to choose 'unfashionable' or 'uncomfortable' than when the prompt refers to a 'global brand'.

Key Novelty

Brand Bias Quantification Framework

Systematically probes LLM associations between brand locality (global vs. local) and attribute polarity (positive vs. negative) using fill-in-the-blank tasks
Investigates socio-economic bias by correlating luxury brand recommendations with the GDP per capita of the recipient's country
Identifies a 'Country-of-Origin' effect where bias reverses: LLMs prefer local brands when the user explicitly states they are from that brand's domestic country

Architecture

Examples of the completion tasks used to measure bias in two directions: Stimulus to Attribute Inference (SAI) and Attribute to Stimulus Association (ASA).

Evaluation Highlights

GPT-4o recommends luxury brands for high-income countries 98.88% of the time, compared to only 1.97% for low-income countries
Kendall’s τ tests confirm significant bias in all models (p < 0.001), with GPT-4o showing the strongest negative association for local brands (τ = 0.423 in ASA task)
Country-of-Origin effect reverses bias: Llama-3-8B prefers local brands 84.95% of the time when the domestic country is specified in the prompt

Breakthrough Assessment

7/10

Novel exploration of a previously understudied bias (commercial/brand bias) with clear, statistically significant findings across major models. The identification of the Country-of-Origin reversal is particularly insightful.

⚙️ Technical Details

Problem Definition

Setting: Fill-in-the-blank generation tasks to measure associations between brand types and attributes, and open-ended recommendation tasks.

Inputs: Prompt templates containing brand references (e.g., 'Nike (a global brand)') or attribute references (e.g., 'fashionable').

Outputs: Selected attribute (SAI), selected brand (ASA), or recommended brand name.

Pipeline Flow

Prompt Generation (Templates with Stimuli/Attributes)
Inference (Model selects option or generates text)
Response Filtering (Categorizing valid vs. invalid responses)
Statistical Analysis (Kendall’s τ)

System Modules

Prompt Generator

Constructs fill-in-the-blank queries using curated lists of brands (global/local) and attributes (positive/negative/neutral)

Model or implementation: Rule-based templates

LLM Inference

Generates completions for the prompts

Model or implementation: GPT-4o, Llama-3-8B, etc.

Modeling

Base Model: GPT-4o, Llama-3-8B-Instruct, Mistral-7B-Instruct-v0.1, Gemma-7B

Compute: Not reported in the paper

Comparison to Prior Work

vs. Salinas et al. (2023): Applies socio-economic lens to product/brand recommendations rather than employment
vs. Kamruzzaman et al. (2023): Adapts the SAI/ASA framework specifically for commercial brand entities and incorporates Country-of-Origin context
vs. BBQ (Bias Benchmark for QA) [not cited in paper]: Focuses on generative fill-in-the-blank and recommendation tasks rather than multiple-choice QA for social stereotypes

Limitations

Limited to 4 brand categories (shoes, clothing, beverages, electronics) and brands from 15 countries
Socio-economic analysis relies solely on GDP per capita, ignoring other factors like inequality (Gini coefficient)
Experiments conducted exclusively in English, potentially missing multilingual nuances
Definition of 'local brand' (no international showrooms) may miss online global presence

Reproducibility

Code: https://github.com/hieuminh65/LLM-Brand-Bias

publicly available (https://github.com/hieuminh65/LLM-Brand-Bias). Dataset includes curated lists of global/local brands from 15 countries and attributes. Code for prompting and analysis is provided.

📊 Experiments & Results

Evaluation Setup

Zero-shot text completion and recommendation tasks

Benchmarks:

Brand-Attribute Association (SAI/ASA) (Fill-in-the-blank) [New]
Socio-economic Gift Recommendation (Open-ended recommendation) [New]
Country-of-Origin Preference (Preference selection) [New]

Metrics:

Kendall’s τ (correlation)
Percentage of luxury recommendations
Percentage of local brand preference
Statistical methodology: Kendall’s τ test for significance (p-value < 0.05)

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Experiment 1 (SAI/ASA) demonstrates significant correlations between brand locality and attribute sentiment across all models.
Brand-Attribute Association (SAI)	Kendall’s τ	0	0.185	0.185
Brand-Attribute Association (ASA)	Kendall’s τ	0	0.423	0.423
Brand-Attribute Association (ASA)	Kendall’s τ	0	0.490	0.490
Experiment 2 reveals stark socio-economic disparities in luxury vs. non-luxury recommendations based on country income.
Gift Recommendation	% Luxury Rec. (High Income)	50.0	98.88	48.88
Gift Recommendation	% Luxury Rec. (Low Income)	50.0	1.97	-48.03
Experiment 3 shows a reversal of bias when the prompt specifies the user is from the brand's country of origin.
Preference Selection	% Local Brand Preference	50.0	84.95	34.95

Main Takeaways

Global brands are systematically associated with positive attributes (fashionable, reliable) while local brands are linked to negative ones (unfashionable, unreliable).
Using generic terms ('global brand'/'local brand') instead of specific names amplifies the measured bias, suggesting the bias is conceptual, not just token-frequency based.
Models exhibit 'economic determinism' in recommendations, strictly gatekeeping luxury items from low-income regions, even when not prompted to consider cost.
The 'Country-of-Origin' effect is a strong counter-bias: explicitly stating nationality can flip the model's preference from global to local brands.

📚 Prerequisite Knowledge

Prerequisites

Understanding of LLM prompting and generation
Basic statistical correlation methods (Kendall’s Tau)
Familiarity with social bias evaluation in NLP

Key Terms

SAI: Stimulus to Attribute Inference—a task where the model is given a brand (stimulus) and must select a corresponding adjective (attribute)

ASA: Attribute to Stimulus Association—a task where the model is given an adjective (attribute) and must select a corresponding brand (stimulus)

Global Brand: A brand with significant presence in multiple countries across various continents (e.g., Nike, Apple)

Local Brand: A brand lacking in-person showrooms outside its country of origin (e.g., Jmofs in Nigeria)

Glocal Brand: A brand combining global reach with local adaptation; used as a neutral option in experiments

Country-of-Origin Effect: A marketing concept where consumer perception of a brand is influenced by the country where the brand or product originates

Kendall’s τ: A rank correlation coefficient used to measure the ordinal association between two quantities (here, brand locality and attribute sentiment)