Research & Publications

The access to non-textual documents, such as historical photographs, still represent a major challenge in accessibility to historical archives and digital libraries. The access to such information has been accelerated through the years thanks to the emergence of multimodal encoders. However, due to the vast training of such models, they present biases than, when deployed on heritage institutions, may result in the distorsion of historical narratives when involving minorities and underrepresented groups. In this paper, we investigate how gender bias in multimodal encoders shifts when exposed to historical documentation. We conduct a large-scale empirical audit of gender bias in CLIP using century-spanning datasets of archival photographs, ranging from controlled yearbook images to unconstrained real-world collections. We leverage a semantic taxonomy of stereotype-related concepts and an entropy-based bias metric, we quantify how gender associations encoded by CLIP vary across decades and semantic domains.

Available soon →

This paper takes the position that the mass adoption of LLMs and multimodal AI systems in educational contexts is unlike other technologies and, above all, an epistemic shift. We believe that the most powerful thing we can do in response is to focus on building critical thinking in the classroom. For this, we propose an educational framework built around critical inquiry and technological sovereignty. The goal is for students to learn to question and understand AI, not just blindly use it. The framework provides five design principles and a cyclical model of critical thinking development which different educational communities can adapt to their own contexts and needs. Drawing from the authors’ teaching experiences, two case studies are presented to examine the framework in practice: a vocational training program in mechanical fabrication, and a digital skills course for adults with disabilities. The paper concludes with a proposal for a collective pedagogical playbook: an open, living resource where educators from Fab Labs, universities, and informal learning contexts can contribute their own methodologies, build upon others’ and document the process. The proposals in this paper are starting points, not rules, and the authors believe they should be co-created with the education community.

Available soon →