AI without conscience is but the ruin of the soul

---

“Science sans conscience n’est que ruine de l’âme”: worn as it may be, this quotation from Rabelais has never been more pertinent than it is today, including in the humanities and social sciences, as language models become ever more present in our practices, from the most mundane to the most exceptional. With AI as we define it today (machine and deep learning), we historians have carried on along the path we have been following with computing and “the digital” for several decades: an integration — sometimes too discreet (in the sense of undocumented) — of new technologies into the historian’s workshop, as Caroline Muller (Rennes 2) and I have recently discussed (see Écrire l’histoire. Gestes et expériences à l’ère numérique). In less than four years, and though it was used before, AI has thus become a methodological and epistemological issue for history. 

To reflect on this issue, the 8th Conference on Digital Humanities and Digital History (15–16 June 2026), co-organised with the German Historical Institute Washington, the Chair for Digital History at Humboldt-Universität zu Berlin and NFDI4Memory, alternated more reflexive and critical contributions with contributions more oriented towards the historians’ toolkit. “Alternated” may not even be the right word: many contributions, like the two workshops, drew on experience, on practice, on tools as the starting point for critical reflection on the use of AI — by their authors, and by historians more generally.

Opening the conference, Anna Neovesky raised the question of historians’ skills, observing that AI, in various forms and at every stage of the historian’s work, is already integrated into the way historians work. The question of historians’ digital literacy, with or without AI, is a minefield, more polemical than one might think: Anna Neovesky handled it brilliantly by grounding her reasoning in an actual practice of AI and in reference to a now well-established tradition of digital hermeneutics. 

Without being able to do justice to the whole of a rich conference, one can mention — besides our own contribution on an AI assistant for the peer review of Journal of Digital History articles — the use of knowledge graphs (Sara Alimenti, Regina Lupi, Giuseppe Liotta, Carla Binucci, Emilio di Giacomo, Giulio Biondi, Fabrizio Grosso, Marco Legittimo, Paola Venuti (University of Perugia) , “Augmenting Historical Research through AI: A Case Study on Norcia (Italy) 1859”), vocabularies and interpretation (Arlinde C.E. Vrooman, “Constructing Semantically Coherent and Interpretable Historical Vocabularies from Domain-Specific Corpora”), the hard question of quotations across sources (Jeri Wieringa, Bennett Nagtegaal, Rebecca Sutton Koeser, Laure Thompson, Hao Tan, Mary Naydan, Edward Baring (Princeton University), “Embed and Seek: Tracking Quotations Across Sources”), handwritten text recognition (HTR) and its use on colonial manuscripts (“Adapting AI for Marginalised Archives: Developing Handwritten Text Recognition Models for Colonial Paraguayan Manuscripts”, Guillaume Candela, Cardiff University), re-OCRing with vision language models (“Beyond ‘Dirty OCR’: Evaluating VLM-Driven Re-OCR Pipelines for Legacy Digital Collections in Latin America”, Jairo Antonio Melo Flórez, University of California, Santa Barbara Library), the question of infrastructure and reproducibility (“AI as Infrastructure – Building a Reproducible Research Data Cycle for Historical Sources”, Clemens Beck, Daniel Motz & Johannes Mitschunas (Friedrich Schiller University Jena)), the integration of LLMs into historical scholarship (“Augmented Hermeneutics: An LLM-Based Evaluative Workflow for Semantic Annotation in Historical Texts”, Torsten Hiltmann (Chair of Digital History, Humboldt-Universität zu Berlin; NFDI4memory)). All these contributions illustrate a happy — in the sense of welcome and necessary — tension between experience and reflexivity, between practice and ethics. Some add a more political dimension to this tension: Candela and Melo Flórez cannot dissociate their tools from the risk of epistemic erasure. 

This necessary tension also ran through contributions centred more on a form of historicisation of current AI practices, for instance within the European Union (Nele Elina Prinz, “A History of Simulation in the European Union: Politics of Simulation or Simulation of Politics?”, Copenhagen Business School), or on the relationship between AI and collective memory (Anders Ravn Sørensen, Copenhagen Business School, “From Archive to Algorithm: Generative AI and the Transformation of Collective Visual Memory”) — a contribution best complemented by further reading (Hoskins, or Gensburger & Clavert). One might add Bermès’, Hervieu’s and Chemla-Romeu-Santos’ contribution on the Skyblog archives, another reminder that the web’s vernacular past is itself becoming both an object and a casualty of these transformations (“A Sensitive Approach to Web Archives through AI: The Skyblog Experiment”, Emmanuelle Bermès (École nationale des chartes), Marina Hervieu (École nationale des chartes, Bibliothèque nationale de France), Axel Chemla—Romeu-Santos (Université Paris 8)). 

In the end, the expression put forward by Todd Presner in the second keynote — uncritical fabulation — perhaps best captures this ‘happy’ tension: what we need today is to historicise reasoning by statistical patterns, and to draw out its consequences for narrative theory.