Over the past decade, digital humanities research has increasingly explored artificial intelligence and computer vision for cultural heritage analysis, particularly for large-scale image collections (Villaespesa and Murphy 2021). These approaches have demonstrated the potential of automated visual analysis for tasks such as object detection, visual similarity, and thematic clustering. However, much of this work prioritises scalability and pattern extraction, often at the expense of interpretability and integration into scholarly or professional workflows. Recent DH scholarship has called for more reflexive and situated uses of AI when dealing with heterogeneous and historically complex materials (Arnold and Tilton 2023). From this perspective, data are understood as cultural artefacts shaped by curatorial practices and institutional histories. This has led to growing interest in human-in-the-loop approaches, in which algorithmic outputs are treated as provisional and subject to expert validation. In the GLAM sector, several studies have also highlighted the limitations of large, general purpose models, such as LLMs or vision-language models, when applied to specialised heritage collections, due to vocabulary mismatch, lack of contextual grounding, and risks of misclassification.

Building on these discussions, TORNE-H (Traitement d'Objets par Reconnaissance Numérique en Environnement Humain) is a research and development project led by the École nationale des chartes-PSL, in partnership with the French Ministry of Culture, the Musée des Arts décoratifs (MAD), the Bibliothèque nationale de France (BnF), and the Musée d'Orsay. Shifting the focus from proof-of-concept experimentation to practice-based integration, it asks not only what AI can detect, but how it can be meaningfully embedded into museum documentation and conservation workflows. It addresses shared institutional challenges related to the documentation, analysis, and accessibility of large and heterogeneous heritage collections (Bermès and Charpier, 2024).

The project's experimental core is anchored at the Musée des Arts décoratifs, whose department of drawings, wall papers and pictures alone comprises approximately 700,000 items that remain partially documented or not yet inventoried. This context provides a concrete testing ground for evaluating AI-based tools under real operational constraints. A concrete example illustrates the project's approach: for the Jean Royère collection, a YOLO-based object detection model was trained on decorative motifs, while a lightweight Python script using regular expressions extracted structured data from the designer's own textual descriptions. Both outputs were then merged into a unified dataset ready for ingesting into the museum's documentation system – demonstrating that meaningful integration can be achieved using low-resources algorithms. TORNE-H adopts a use-case-driven approach based on interviews and workshops with museum professionals, ensuring alignment with professional expertise, documentation standards, and long-term sustainability. Complementary experimentation with the BnF and the Musée d'Orsay enables cross-institutional comparison and supports the transferability of methods.

Grounded in DH principles of situated interpretation and reflexive data production (Drucker 2020), TORNE-H frames computational analysis as a heuristic tool supporting expert inquiry rather than replacing it. Technically, the project relies on computer vision pipelines combining object detection, image segmentation, and feature extraction, applied to carefully selected subsets of collections rather than large-scale model training. Two core tools structure this workflow: TiamaT, a YOLO-based iterative object detection protocol (Charpier 2025), and Panoptic, a CLIP-based image similarity exploration system (Bouté et al. 2024). Outputs are presented through explanatory interfaces exposing confidence levels and intermediate results, enabling experts to evaluate and correct algorithmic suggestions, ensuring traceability, accountability, and alignment with professional standards.

A major outcome of TORNE-H has been the identification of shared use cases across partner institutions. Through sustained dialogue with museum professionals, the project clarified recurring needs related to documentation backlogs, collection prioritisation, and realistic expectations regarding AI, proving essential in avoiding speculative or unsustainable applications. Operationally, the project aimed at enabling the rapid indexing of the Jean Royère collection (18,000 items), demonstrating that computer vision tools can support early-stage documentation tasks at scale. While outputs do not replace scholarly description, they provide structured entry points that facilitate expert analysis and reduce initial engagement time with large visual corpora. Equally significant was the process of professional acculturation to AI: collaborative experimentation fostered a shared vocabulary and helped institutions articulate feasible, context-sensitive requirements.

Human engagement is a foundational principle of TORNE-H, which also adopts an ecological perspective on AI by privileging measured impact rather than technical escalation. Instead of large-scale model training or resource-intensive infrastructures, it relies on targeted datasets, modular pipelines, and the reuse or adaptation of existing models, limiting computational and energy costs while remaining responsive to institutional needs. Ecological engagement is thus framed not only as energy efficiency, but as a methodological choice aligned with the material and epistemic constraints of heritage institutions. Museum professionals are involved at every stage of the process: defining use cases, shaping annotation categories, evaluating outputs, and validating results. The project’s human-in-the-loop architecture ensures that AI outputs remain suggestions rather than authoritative statements. Algorithmic bias is addressed through systematic expert validation at each iteration: results are assessed by heritage professionals rather than by the models themselves, a critical safeguard given that accurate interpretation of heritage collections requires fine-grained disciplinary knowledge that pre-trained models cannot substitute.

TORNE-H explicitly addresses ethical concerns related to the deployment of AI in cultural institutions. Particular attention is paid to professions potentially at risk of technological substitution, such as documentation, mediation, or curatorial work. Rather than pursuing automation for its own sake, the project integrates a reflexive dimension, in which proposed uses of AI are discussed, evaluated, and sometimes rejected. Ethical engagement, in this context, involves preserving professional agency, ensuring transparency in algorithmic processes, and avoiding applications that could undermine the quality, integrity, or social role of heritage work.

As the project continues, further evaluation across institutions will refine both technical and organisational aspects of the proposed workflows, with the aim of articulating a transferable framework for responsible AI integration in heritage institutions. We hope this contribution will foster dialogue among digital humanists, museum professionals, and AI practitioners on how AI can be meaningfully, sustainably, and responsibly embedded in heritage practices, and on the forms of engagement such technologies should support.