Recovering bullying appraisals from retrospective narratives: Agreement of individual human readers, GPT-5, and DeepSeek-V3 with Chinese narrators’ self-reports

Large language models (LLMs) have demonstrated advanced capabilities in processing and interpreting complex human narratives, as well as generating human-like decisions. However, little is known about their ability to approximate narrator-reported appraisals in bullying narratives and their performance compared to individual human readers. This study compares LLMs with individual human readers in recovering narrators’ self-reported bullying appraisals. We designed a four-group comparative study involving 1873 Chinese adults (31.55% males). Participants from Group A ( n = 965) served as the narrators who provided first-person accounts of bullying experiences and self-reports of appraisals. We evaluated how accurately Group B (individual human readers; n = 908), Group C (GPT-5), and Group D (DeepSeek-V3) reproduced the appraisals, using Group A narrators’ self-reported bullying appraisals as the first-person criterion. Accuracy analyses, paired statistical comparisons, logistic regression, and descriptive comparisons of response distributions were performed. The results indicated that LLMs generally matched or exceeded individual human readers in recovering first-person appraisals of bullying content. GPT-5 also showed partial similarity to Groups A and B in associations between appraisals and defending behavior. Because explicit and implicit appraisal cues were not coded separately, the observed agreement may reflect both extraction of stated information and inference from indirect narrative cues. Severity ratings showed a model-specific pattern: GPT-5 assigned higher severity ratings than individual human readers, whereas DeepSeek-V3 assigned lower and more mid-range severity ratings. Our findings suggested that LLMs may assist with research coding of bullying-related appraisals from retrospective narratives. Further work should examine whether this approach extends beyond the Chinese sample and prompt conditions.