<p>With the rapid growth of the credit market, credit scoring becomes increasingly important for credit risk management. Current credit scoring models tend to use both structured data and textual data for credit assessment, and the main difficulty lies in extracting meaningful textual features from unstructured loan texts. ChatGPT is a generative large language model that can proficiently comprehend textual material and perform reasoning tasks. Therefore, this study proposes a ChatGPT-based integrated framework, named as GPT-LGBM. The proposed framework can be divided into two stages: the initial stage involves psychological feature extraction with ChatGPT, followed by credit risk classification using light gradient boosting machine (LightGBM). During the first stage, two paths are designed to extract borrowers’ Big Five personality traits from loan texts without manual annotation. One denoted as Path-D employs ChatGPT directly to analyze all loan texts, which aims at maximizing the benefits of using ChatGPT. The other denoted as Path-K adopts the idea of knowledge distillation to combine ChatGPT and the deep learning model RoBERTa, which can mitigate the risk of data breach by restricting ChatGPT’s access to all loan texts. In the second stage, the Big Five personality traits and structured features are concatenated as inputs for the LightGBM model. Extensive experiments based on the Lending Club data demonstrate that utilizing the ChatGPT-generated psychological features, GPT-LGBM can outperform the benchmark models and retain feature interpretability, making it an effective framework for credit evaluation.</p>