Digital Distortions: Auditing Bias in LLM-based Career Recommendations

We released one LLM generated recommendation dataset and one real-world employment dataset. The LLM generated dataset contains experiments over 9 social groups, 4 college degrees, 17 degree fields, 12 universities, 2 target work locations, and 383 occupation categories, resulting in 14,688 scenarios in total. The real employment dataset is retrieved from LinkedIn via API. The retrieved dataset contains approximately 295,000 individuals and a total of 1,330,000 records. For each individual, we select only their first three occupations within the five years following graduation to focus on early career choices. We rank occupations under the given attribute conditions by calculating their frequencies within the dataset, thereby creating an occupation ranking dataset. This real-world occupation ranking dataset is open-sourced to support future research.