TORNE-H is a research 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, which aims at using artificial intelligence, and more precisely computer vision, to help with the inventory of uncatalogued heritage iconographic collections. The project is anchored at the Musée des Arts décoratifs, whose collections comprise nearly one million objects, including approximately 700,000 items that remain only partially documented. The project shifts the focus from proof-of-concept experimentation to practice-based integration, asking not only what AI can detect, but how it can be meaningfully embedded into museum documentation and conservation workflows.

Beyond human-in-the-loop, our methodology is grounded in digital humanities principles such as situated interpretation and reflexive data production (Drucker 2020; Villaespesa and Murphy 2021). Rather than treating AI as an autonomous system, computational analysis is framed as a heuristic tool supporting expert inquiry, this approach requiring continuous human oversight. The workflow was developed through interviews and workshops with conservation and documentation professionals, in order to identify concrete use cases where AI could provide assistance without disrupting existing practices. Selected use cases included object recognition, iconographic pattern detection, visual similarity analysis, and assisted metadata enrichment. Several tools were developed or adapted, including the TiamaT pipeline (YOLO-based object detection), the use of Panoptic (a CLIP-based system developed by CERES–Sorbonne), and additional scripts designed to integrate and enrich human-produced documentation, reinforcing the iterative human–machine loop.

Finally, 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. Equally significant was the process of professional acculturation to AI. Collaborative experimentation fostered a shared vocabulary and helped institutions articulate feasible, context-sensitive requirements. Moreover, TORNE-H adopts an ecological perspective on AI by privileging measured impact rather than technical escalation. Instead of large-scale model training or resource-intensive infrastructures, the project relies on targeted datasets, modular pipelines, and the reuse or adaptation of existing models. This approach limits computational and energy costs while remaining responsive to institutional needs. 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, replication of the method across other partner institutions will refine both technical and cultural practices of the proposed workflows. We aim at fostering dialogue among digital humanists, museum professionals, and AI practitioners. The project demonstrates how scholarship can help with meaningfully, sustainably, and responsibly embedding AI tools and processes in heritage practices, thus significantly contributing to the enhancement of the social impact of digital humanities research in cultural heritage institutions.