Detecting the Aura of an Article – Was the Original Manuscript Seen?

Abstract

This paper investigates whether scholarly texts reveal, through linguistic patterns, whether authors engaged directly with original artefacts or relied on facsimiles, digital or analogue. Drawing on debates about reproduction and materiality, it frames this question as a classification task and evaluates machine learning approaches using both synthetic and real-world data. While initial experiments under controlled conditions suggest that certain linguistic features may correlate with different modes of access, these patterns prove unstable in practice. Key challenges include the lack of reliable ground truth labels, the influence of domain-specific writing conventions, and multilingual variation. In particular, highly standardised descriptive language in specialised fields limits the discriminative power of textual features. The findings indicate that the task is fundamentally ill-posed under standard assumptions, highlighting the limits of data-driven methods in capturing epistemic conditions that are only partially encoded in text.