Computational Detection and Source Attribution of AI-Generated Literary Texts: A Systematic Review of Computational Features, Detection Methods, and Evaluation Designs

Abstract

With the increasing application of large language models in novels, short stories and poetry, identifying AI-generated literary texts and confirming how machines participate in creation has become an important issue in the fields of computational linguistics and digital humanities. This study follows the PRISMA 2020 guidelines and adopts a structured systematic review approach to analyze 21 relevant empirical studies published between 2020 and 2026. The research focuses on the computational characteristics of AI-generated literary texts, computational detection and source-attribution methods, and the impact of generation regime and evaluation design on detection performance. The findings indicate that the existing evidence does not support a universal “AI marker” applicable across languages and genres. A more consistent tendency is that machine-generated texts often show higher homogeneity in the multi-dimensional language and narrative feature space and also retain generator-specific fingerprints. Vocabulary distribution, specific language structure, emotional expression, literary form and narrative organization can provide detection signals, but their effectiveness is substantially influenced by the comparison corpus, prompt strategies, and genre conventions. Supervised classifiers usually perform well in specific datasets, but their cross-domain transferability is still unstable. Probability-based zero-shot methods show considerable variation across poetry and long narrative texts; in contrast, LLM judges perform relatively better when provided with examples, source texts, or provenance information. In general, the detection performance of literary texts is jointly determined by language, genre, text length, form of machine involvement, reference information, and evaluation tasks. Future research should build more representative literary corpora, adopt rigorous data grouping and external evaluation methods, and shift from simple human–AI binary classification to interpretable identification of the degree and type of machine involvement.