What does responsible, sustainable AI-assisted historical research require, and who has access to it? This talk approaches the question as an infrastructural challenge. AI-assisted research depends not only on models and subscriptions, but on digitized collections, language resources, repositories, expertise, and the institutional capacity to sustain them. The current turn toward sovereign AI and publicly funded infrastructures make these dependencies especially visible.
Using the 1908 Messina earthquake as a case study of transnational post-disaster discourse, I examine what building a responsible and sustainable AI-assisted historical workflow requires. Analyzing more than 30,000 newspaper pages from Italy, Germany, France, and Austria-Hungary required careful adaptation of models and workflows to historical, linguistic, and cultural differences. Responsibility here means more than keeping models small or data private; it means making historical sources computationally accessible without flattening the differences that give them meaning.
These adaptations are not only safeguards against flattening; they also produce infrastructure. Multilingual annotations, domain-specific evaluation practices, adapted models, and historically grounded datasets can become reusable resources for languages and cultural contexts that existing AI systems represent poorly. Responsible and sustainable AI-assisted research therefore means not only adapting AI to historical material but doing so in ways that leave better data, methods, and expertise for future research across different linguistic and cultural settings.