This paper presents findings from a collaboration between a scholar and a large language model (ChatGPT) during four months of sustained editorial work on a digital edition of a major nineteenth-century literary-philosophical text, documented in approximately eighty working sessions preserved from a live editorial workflow. The task involved creating 940 new Wikidata items for persons, works, and editions cited in the text while encoding references in TEI-XML, requiring repeated decisions about bibliographic modeling, ontological classification, and authorship attribution.
Analysis of these sessions reveals recurring patterns in how the LLM functioned across the symposium’s three proposed roles. As colleague, the AI accelerated multilingual bibliographic research, searched catalogues, authority files, digitized texts, and Wikidata live, generated structured data and QuickStatements, and helped identify missing works and persons and inconsistencies in the project’s encoding. As mirror, it acted as a soundboard to articulate modeling decisions, making tacit editorial practices visible by forcing them to be restated as rules and operationalized; however, it also copied the project’s existing formalization, including encoding errors, thereby exposing discrepancies between its own reading of the text and the existing markup. While it read the text correctly, its tendency was to trust available encoded structure and metadata more than natural-language meaning. As rival, it asserted competing judgments with unwarranted confidence: over-modeling against the project’s deliberate minimalism and sometimes challenging the critical apparatus of print editions.
The central finding is that AI augments the decision-making process of editorial work and significantly reduces both mechanical data entry and philological and bibliographic research. Its main weakness proved to be the reliable assignment of external identifiers and structured fields, making verification and revision indispensable. The scholar’s corrections—forcing convergence, restating rules, insisting on operational rather than ontological justification—made editorial logic more explicit and systematic than it would have been without the collaboration.
Author: Silvia Stoyanova