ESHS/HSS Annual Meeting

Epistemic Values, Reliability and Trust: The Inaccurate Edition

• In Event: Findings in Digital Humanities 2: Modern

English Abstract

Historical-critical editions have long underpinned work with historical sources. Philology has established rigorous methods whose leading principle is textual accuracy. Yet this ideal relies on time-consuming manual labour by expert transcribers and historians, so only a small fraction of sources can be edited in this way.

The scholarly edition of Alexander von Humboldt’s travel notebooks exemplifies this paradigm. By contrast, the field diaries of the German ornithologist Alfred Laubmann (1886–1965; 35 notebooks, 6,381 pages) illustrate the vast written heritage that remains unedited because funding and labour are limited, despite clear relevance for the history of and modern biology.

Now, AI contests the "old" paradigm. We report on prototyping an “AI-born edition” based on automatic handwritten text recognition, LLM-based text constitution, and computational annotation and contextualisation.

This edition is, by design, inaccurate and error-prone. Our hypothesis is that, despite its defects compared to established standards and expectations, it possesses high epistemic value for most research questions and clearly outperforms traditional editions in cost–benefit efficacy.

To test this hypothesis, we (a) juxtapose the automatically generated Laubmann edition with the “accurate” Humboldt edition and (b) evaluate several metrics, including word error rates, recall and precision of annotated entities, and the relevance of semantic shifts. For instance, the AI-born text may contain words absent from the manuscript but close enough in meaning that historical interpretation remains unaffected. Finally, we interrogate what reliability and trust mean in the context of AI-generated scientific content.

Authors

• Malte Rehbein, University of Passau

• Tobias Kraft, Berlin-Brandenburg Academy of Sciences and Humanities

• Tobias Perschl, University of Passau