Introduction
As Large Language Models (LLMs) like GPT-3.5 become integrated into e-commerce platforms to generate millions of product descriptions, they do more than list specifications; they shape consumer perceptions of what is "normal" or "appropriate" for different genders. While algorithmic bias in hiring and facial recognition has been extensively studied, e-commerce remains a relatively unexplored domain. This research investigates how AI-generated descriptions embed subtle gender biases, ranging from stereotypical adjectives to disparities in persuasive language.
Methodology
The researchers employed a multi-stage, data-driven approach to identify and categorize these biases. The process began with an initial dataset of 10,000 descriptions generated by GPT-3.5 for eBay listings. This dataset underwent a rigorous filtering process involving human annotation and automated flagging to isolate potentially biased content.
Following the flagging stage, a panel of four experts in Diversity, Equity, and Inclusion (DE&I) and AI red teaming conducted a qualitative review. This expert feedback was synthesized into a new taxonomy of e-commerce-specific algorithmic harms. Finally, the researchers validated these findings through quantitative keyword detection and counterfactual analysis, comparing descriptions for identical products where only the target gender was varied.
A Taxonomy of E-commerce Bias
The study identifies six distinct categories of bias that traditional, general-purpose frameworks often overlook. These categories are grouped under three high-level themes: Exclusionary Norms, Stereotyping and Objectification, and Disparate Performance.
1. Exclusionary Norms
- Body Size Assumptions: Language that positions specific body types as the "default" (e.g., "fits most women" or "ideal for regular-sized women"), implicitly marginalizing those outside these narrow bounds.
- Target Group Exclusion: Emphasis on a target group that explicitly excludes certain genders (e.g., "designed exclusively for men").
2. Stereotyping and Objectification
- Target Group Assumptions: The AI often "hallucinates" gendered contexts not present in the input data, such as describing a baby bottle as "perfect for moms on-the-go."
- Bias in Advertised Features: A stark divide exists in feature highlighting. Women’s clothing descriptions frequently focus on appearance and objectification (e.g., "flattering silhouettes"), while men’s clothing focuses on utility and durability (e.g., "rugged," "comfortable").
- Product-Activity Associations: Stereotypical links between products and activities, such as associating a purse with "running errands."
3. Disparate Performance
- Persuasion Disparities: The research uncovered a "persuasion gap," where GPT-3.5 was 5.5 percentage points more likely to use aggressive "call-to-action" language (e.g., "Order now!") for men’s products compared to women’s.
Quantitative Validation and Linguistic Fingerprinting
To prove that these biases are systematic, the authors conducted a counterfactual test. They trained a simple bigram classifier to predict the intended gender of a description where all explicit gendered terms (like "man" or "woman") were masked. The classifier achieved over 90% accuracy, demonstrating that the AI's stylistic and vocabulary choices are deeply and systematically gendered.
Key predictive keywords further illustrated this divide: for children's clothing, terms like "adventure" and "toughest" were highly predictive of products for boys, while "fashion" and "flattering" were predictive of products for girls. Additionally, exclusionary language regarding body size appeared in 10.1% of GPT-3.5 descriptions and 14.3% of descriptions from a specialized internal e-commerce model.
Discussion and Limitations
While this study provides a robust framework for understanding e-commerce bias, it is subject to several limitations. The analysis was restricted to a binary (men/women) view of gender, reflecting the current categorization of major e-commerce platforms. Furthermore, the human annotators were demographically similar, and the scope was limited to English-language descriptions on a single platform.
Future work should expand this methodology to investigate intersectional biases—how gender interacts with race, religion, or disability—and explore how LLMs handle non-binary and gender-diverse identities in a commercial context.
Summary updated