Authorship Verification with AI-Generated Mimicry: A Study with Bronte-Style Poems

Objectives: Determining authorship is crucial in digital humanities, forensics, and research paper attribution. Authorship Verification (AV), an open-set problem, determines whether the same individual authored the provided pair of texts. Methods: This study examines the poetry of the Brontë sisters, which demonstrates considerable thematic and linguistic similarity, therefore presenting a critical problem for AV. The difficulty of the challenge is further exacerbated by the inclusion of Artificial Intelligence (AI)-generated poems emulating the writing style of the Brontë sisters. This study examines the optimal representative feature aspect when trained using both complete poems and excerpts. Findings: The proposed AV framework utilizes a Siamese network that uses Manhattan distance-based similarity assessment with optimal parameters. In the absence of a comprehensive context, the semantic and structural features of poetry significantly enhance AV performance, surpassing quantitative lexical components by 36%. The overall ability of the baseline and deployed models to distinguish between original and AI-generated poems is examined using the sub-task of closed-set Authorship Attribution (AA). This study establishes a basis for future research on AI plagiarism and academic integrity, offering significant insights for further investigation. Also, AV with AI-generated poems highlights that AI can mimic the authors only to some extent, and that is well depicted with the accuracy of the demonstrated method. Novelty: A computationally and literarily challenging AV task, involving poems, has been carried out (including AI-generated poems). To the best of our knowledge, a few works on poem authorship exist, but none systematically on AV on poem excerpts.