Auditing Preferences for Brands and Cultures in LLMs

In recent years, the development of large language models (LLMs) has significantly enhanced the role of AI systems in information retrieval and decision support. LLMs like ChatGPT, Google Gemini, and Meta AI have become the primary means for many people to access information. These systems not only influence individual choices but also potentially affect market fairness and competition. However, existing research has primarily focused on social biases, such as gender, race, and religion, while brand and cultural biases have received relatively less attention. The introduction of the ChoiceEval framework aims to fill this research gap by systematically evaluating geographic bias in LLM brand and cultural preferences, revealing its potential impact on market and cultural diversity.

Geographic biases in LLM brand and cultural preferences may have profound implications for market fairness, competition, and the diversity of information exposure. Especially in AI-driven markets, these biases could lead to systematic economic advantages or disadvantages. Existing research has primarily focused on social biases, while brand and cultural biases have received relatively less attention. Therefore, there is an urgent need for a systematic approach to evaluate geographic bias in LLM brand and cultural preferences to reveal its potential impact on market and cultural diversity.

The core innovation of the ChoiceEval framework lies in its first systematic generation of evaluation questions and assessment of entity-perception bias in AI assistants. Unlike previous research primarily focused on social biases, ChoiceEval focuses on brand and cultural bias evaluation in open-ended recommendation scenarios, providing a scalable foundation for evaluating how AI assistants may shape real-world decisions. Specifically, ChoiceEval achieves this by segmenting users into psychographic profiles and generating diverse prompts, allowing for the quantification of preference and geographic bias. This approach is not only applicable to measuring social biases but extends to evaluating brand and cultural biases, particularly in open-ended recommendation scenarios.

The implementation of the ChoiceEval framework includes the following steps:

• �� User Cluster Definitions: Use psychographic user clusters (e.g., budget-conscious, wellness-focused, convenience) to capture how different user types interact with AI assistants.

• �� Question Generation: Use LLMs to adapt core consumer clusters for each domain, converting their general characteristics into terminology and concerns specific to that decision-making context.

• �� Response Extraction: Query each LLM with the same set of questions and record its responses. Simulate expert evaluation with multiple independent extraction runs to reduce decoding variability and parsing ambiguities.

• �� Bias Analysis: Perform statistical analysis on extracted recommendations to reveal geographic bias in AI assistant brand and cultural preferences.

In experiments, the ChoiceEval framework was applied to Gemini, GPT, and DeepSeek across 10 topics and over 2,000 questions. The experimental design includes:

• �� Datasets: Select 10 topics covering commerce and culture, such as hotel chains, electric cars, and running shoes.

• �� Baselines: Choose Gemini, GPT, and DeepSeek as comparison models to reveal geographic bias in brand and cultural preferences.

• �� Evaluation Metrics: Use normalized top-k choice sets to quantify preference and geographic bias.

• �� Hyperparameters: Generate 23 questions per topic and user cluster pair, resulting in 2,070 questions in total.

Experimental results show that U.S.-developed models Gemini and GPT exhibit marked favoritism toward American entities, while China-developed DeepSeek shows more balanced yet still detectable geographic preferences. These preference patterns persist across user personas, indicating systematic rather than incidental effects. In multiple domains such as hotel chains, electric cars, and running shoes, LLM recommendations are skewed towards American entities despite the existence of competitive global alternatives. Statistical analysis of extracted recommendations reveals geographic bias in AI assistant brand and cultural preferences.

The application scenarios of the ChoiceEval framework include:

• �� Market Auditing: Help researchers and regulators evaluate geographic bias in LLM brand and cultural preferences to reveal its potential impact on market and cultural diversity.

• �� Model Improvement: Provide feedback to AI developers on LLM preferences to improve model fairness and diversity, ensuring diversity and fairness in recommendations.

• �� Consumer Protection: Help consumers identify biases in AI recommendations to make more informed decisions, avoiding being swayed by certain preferences.

The limitations of the ChoiceEval framework include:

• �� Evaluation results may be limited by the selected models and datasets, especially in geographic bias analysis, which may not fully represent preferences across all regions.

• �� Since the framework relies on psychographic user clusters, it may not fully capture the complex behaviors and preferences of all users.

• �� In some cases, LLM recommendations may be influenced by inherent biases in the training data, which ChoiceEval may not fully eliminate. Future directions include expanding the ChoiceEval framework to cover more topics and user clusters, especially in non-English contexts.