The growing influence of artificial intelligence, particularly generative AI, challenges ancient principles of digital humanities (methodological transparency, evidence management, and research reproducibility) by introducing a powerful yet opaque digital infrastructure. While AI accelerates information processing and expands interpretive possibilities, it complicates the traceability of sources and undermines the reproducibility and verifiability of knowledge production (Vitali-Rosati, 2025). Confronted with this opacity, researchers must extend the classical “critique of sources” (Langlois, Seignobos, 1898) to encompass the tools that generate or select them, developing a critique of infrastructures themselves.

This requires critical literacy (Marin, Steinert, 2022): the capacity to assess not just the reliability of outputs produced by opaque models but also the epistemic values and intentions embedded within them. Key questions arise: Who built the model, with what data, and for what purpose? What blind spots or biases are concealed within its design? More broadly, what kinds of historical, social, or cultural questions can such tools meaningfully illuminate—and when might they instead weaken research by bypassing critical engagement with sources?

Library labs, located at the intersection between those who collect and preserve sources and those who interpret them, are particularly well suited for this kind of critical co-construction (Carlin, Laborderie, 2021). By bringing together expertise from curation, data science, and research methodology, these spaces make it possible to document data provenance, uncover corpus bias, negotiate modeling choices, and develop participatory practices that bridge scientific and civic perspectives.

From this interdisciplinary collaboration emerged the Mezanno/Corpusense ecosystem, part of the BnF’s four-year research plan. This open-source tool extracts both the text and structural features of documents and integrates AI modules for content categorization. While its use requires no technical proficiency, it keeps researchers central to the process through three principles: local control of data, modularity of processing chains (layout detection, OCR transcription, structuring via LLMs), and systematic documentation of methodological choices. The goal is not to adopt prepackaged AI services but to adapt them to the standards of the humanities and social sciences: transparency, reproducibility, and openness to collective critique.

Mezanno shows how library labs can become infrastructures of engagement, making visible the technical operations, epistemic decisions, and uncertainties that underlie research. This transparency invites shared critique among researchers, librarians, heritage institutions, and the broader public. Through the example of the BnF DataLab project, this contribution demonstrates how the power of generative AI can be reconciled with the values of critical, participatory, and sustainable research.

Bibliography
Caroline Muller & Frédéric Clavert, Écrire l’histoire. Gestes et expériences à l’ère numérique, Paris, Armand Colin, 2025.
Marcello Vitali-Rosati, “L’IA a-t-elle tué les humanités numériques?”, Humanités numériques et intelligence artificielle: interactions/altercations, Sorbonne Université, 2024. https://ihn.sorbonne-universite.fr/journee-detude-humanites-numeriques-et-intelligence-artificielle-interactions/altercations
Marie Carlin & Arnaud Laborderie, “Le BnF DataLab, un service aux chercheurs en humanités numériques”, Humanités numériques, 2021, 4. ⟨hal-03285816v2⟩
Joseph Chazalon et al., “Reprendre le contrôle des processus de traitement de données de recherche en histoire quantitative : présentation de l’écosystème Mezzano,” DHNord2025, Tourcoing, France. ⟨hal-05421670⟩