An Investigation of Linguistic Biases in LLM-Based Recommendations

1. Problem Statement

The paper investigates linguistic biases in Large Language Model (LLM) based recommendation systems. Specifically, it examines whether LLMs alter their recommendations based on dialectal variations (surface-level syntactic and form differences) even when the semantic intent of the user query remains identical.

The authors argue that recommendation systems can infringe upon descriptive autonomy by steering user behavior toward pre-determined outcomes based on latent linguistic cues (e.g., dialect) rather than explicit user preferences. While biases related to gender, race, and age are well-studied, biases arising from dialect (e.g., Southern American English vs. Indian English) and code-switching (Hindi-English mixing) in recommendation contexts remain under-explored.

2. Methodology

Experimental Design

The study employs a cold-start setting (zero-shot prompting without fine-tuning on user interaction data) to isolate inherent biases within the models.

• Tasks: Restaurant recommendations and Product recommendations.
• Dialects Tested:
  1. Southern American English (AE): Selected as the largest US regional accent group.
  2. Indian English (IE): Standardized Indian English.
  3. Code-Switched Hindi-English (CS): A mix of Hindi and English (e.g., "Arre yaar, I'm dying of hunger...").
• Datasets:
  • Restaurants: Yelp Open Dataset (2023). Subsets of "American" and "Indian" cuisine restaurants were balanced.
  • Products: Walmart Product Reviews Dataset (2020). Categories included Sports, Health, Personal Care, Clothing, Home, Beauty, and Exercise.
• Prompting Strategy:
  • LLMs were provided with a balanced list of 100 items (50 American/50 Indian restaurants or 100 products per category).
  • The LLM was asked to select the top-20 recommendations from this list.
  • Prompts were varied only by dialect (e.g., "Where's a good place to eat?" vs. "Where can I get a good place to eat?" vs. "Kahan milta hai tasty food?").
  • Generalization: To ensure robustness, the study used 20 different random seeds to generate unique balanced lists for each query/dialect combination, resulting in 600 lists per task.

Models Evaluated

Three model families were tested with two size variants each to analyze the effect of parameter size:

1. Llama-3.1: 8B and 70B parameters.
2. GPT-OSS: 20B and 120B parameters.
3. Mistral: 7B-instruct and mistral-small-3.1 (24B).

Statistical Analysis

• Dependent Variable: Aggregate counts of recommended items per cuisine type (for restaurants) or product category (for products).
• Independent Variables: Dialect type (AE, IE, CS), Model Size, and their interactions.
• Model: Linear Mixed-Effects Regression (using lmerTest in R).
  • Fixed Effects: Dialect, Model Size, and interactions.
  • Random Effects: Question ID and Seed ID.
• Testing: Likelihood ratio tests for fixed effects, followed by post-hoc pairwise comparisons (Bonferroni-corrected) of estimated marginal means.

3. Key Contributions

• Novelty in Bias Detection: This is one of the first studies to specifically isolate dialectal and code-switching biases in LLM recommendation systems, distinct from demographic or content-based biases.
• Cold-Start Analysis: By avoiding fine-tuning, the study reveals inherent biases present in pre-trained foundation models before any remedial training or context-provision measures are applied.
• Granular Statistical Evidence: The use of mixed-effects models across 20 seeds provides statistically significant evidence that surface-level linguistic variations significantly alter recommendation distributions.

4. Key Results

Restaurant Recommendations

• Dialect Sensitivity: All tested models showed statistically significant sensitivity to dialect.
  • Llama-3.1: Showed the highest sensitivity. The 70B model recommended significantly more Indian restaurants when prompted with Code-Switched (CS) and Indian English (IE) prompts compared to American English (AE).
  • Mistral: The mistral-small-3.1 (24B) model showed sensitivity, recommending more Indian restaurants with IE prompts than AE prompts.
  • GPT-OSS: Showed the least sensitivity, likely due to superior reasoning capabilities that allow the model to derive semantic intent despite surface variations.
• Model Size Trends: No universal trend was found where larger models were always more or less biased. However, larger models in the Mistral and Llama families tended to recommend fewer Indian restaurants with AE prompts compared to their smaller counterparts, but more with CS/IE prompts.

Product Recommendations

• Category Shifts: Dialect influenced the type of product categories recommended.
  • Beauty & Home: Larger models (Llama-3.1-70B, Mistral-small-3.1) recommended more Beauty products with IE prompts. Smaller models (Llama-3.1-8B, Mistral-7B) recommended more Home products with CS prompts.
  • Code-Switching Impact: The Llama-3.1-70B model was particularly sensitive to CS prompts, recommending more products in 4 out of 7 categories (Sports, Home, Clothing, Exercise) compared to AE/IE prompts.
• Model Family Differences:
  • Llama-3.1: Exhibited the most significant dialect-based and model-size-based differences across all categories.
  • GPT-OSS: Showed the fewest differences, reinforcing the hypothesis that better reasoning capabilities mitigate surface-level linguistic bias.

5. Significance and Implications

• Ethical Concerns (Descriptive Autonomy): The findings suggest that LLMs may infer user identity and preferences based purely on linguistic form (dialect), potentially reinforcing stereotypes and narrowing the diversity of choices presented to users without their consent.
• System Design: The study challenges the assumption that scaling model size alone eliminates bias. In fact, larger models (like Llama-3.1-70B) can exhibit stronger dialect-dependent biases in specific contexts.
• Recommendation Fairness: If recommendation systems adapt outputs based on dialect alone, users may receive constrained or stereotyped suggestions. This highlights a critical need for dialect-invariant behavior in AI systems.
• Future Directions: The authors suggest mitigation strategies such as prompt normalization, dialect-aware training, or fairness-constrained decoding to ensure recommendations are driven by user intent rather than linguistic artifacts.

Conclusion

The paper demonstrates that dialect alone is a significant factor in LLM-based recommendation outcomes. Even with identical semantic intent, prompts in Indian English or Code-Switched Hindi-English systematically yield different recommendation distributions compared to American English. This bias is most pronounced in the Llama-3.1 family and varies by product category, indicating that current LLMs encode latent associations between linguistic form and cultural/demographic preferences that require active mitigation.