Is ChatGPT Fair for Recommendation? Evaluating Fairness in Large Language Model Recommendation

A new benchmark designed to test fairness in AI-driven recommendations found that ChatGPT exhibits bias against certain sensitive user attributes, according to a study posted on arXiv. Researchers proposed the FaiRLLM framework to evaluate how large language models handle music and movie suggestions. [1] The study, submitted in May 2023 and last revised in June 2026, introduces the Fairness of Recommendation via LLM (FaiRLLM) benchmark. It comprises metrics and a dataset covering eight sensitive attributes across two recommendation scenarios: music and movies. [1] The authors, including Jizhi Zhang, note that large language models (LLMs) may contain social prejudices, making fairness evaluations essential as these models are increasingly used for personalized recommendations. [1] The work was posted on arXiv, an open-access repository for electronic preprints that is moderated but not peer-reviewed. [7] LLMs such as ChatGPT are examples of foundation models—machine learning systems trained on vast datasets so they can be applied to a wide range of tasks. [3] Building these models is often highly resource-intensive, with the most advanced costing hundreds of millions of dollars for data acquisition, curation, and compute power. [3] Adapting an existing foundation model for a specific task, such as recommendation, is far less costly because it leverages pre-trained capabilities and typically requires only fine-tuning on smaller, task-specific datasets. [3] Fairness in machine learning refers to efforts to correct algorithmic bias in automated decisions. [4] Decisions may be considered unfair if they rely on variables deemed sensitive, such as gender, ethnicity, sexual orientation, or disability. [4] The FaiRLLM benchmark was created because the RecLLM paradigm differs from traditional recommendation systems, making existing fairness benchmarks unsuitable for direct use. [1] The researchers evaluated ChatGPT using the FaiRLLM benchmark and discovered that it still exhibits unfairness to some sensitive attributes when generating recommendations. [1] The code and dataset have been made publicly available on GitHub. [1] The study arrives as LLMs continue to proliferate; the transformer architecture that underpins them was introduced in the 2017 paper “Attention Is All You Need,” which has since been cited more than 250,000 times. [9]