Recent advancements in generative artificial intelligence (AI) and natural language processing (NLP) have provided quantitative social scientists with unprecedented opportunities to collect, process, and analyze massive-scale observational data. In this study, with assistance of ChatGPT, I am able to assess real-time labor market demand and construct skill mismatch measures by analyzing millions of job postings scrapped from prominent online platforms in Hong Kong. Such demand-side job analysis is considered the optimal method for determining the matched (or required) education credentials for specific occupations, a feat previously unattainable due to the limitations of conventional machine learning (ML) methods. Most of these ML methods depend on teams of human experts to manually label large volumes of data for model training, a process that is time-consuming, labor-intensive, and cognitively demanding. However, the findings of this study indicate a promising avenue for future research. General foundation models like GPT-4 show impressive capabilities on handling job posting data without being explicitly trained. Notably, when fine-tuned, GPT-3.5 achieves GPT-4-level performance (over 90%) on automated data labeling and skill identification benchmarks, at just one-tenth of the cost. More importantly, this study posits that large language models (LLM) can attain state-of-the-art performance on information extraction and classification tasks, which surpasses not only the accuracy of smaller, supervised models but also the efficiency of well-trained human experts.