DH^CH / Paola Lechuga

Paola Lechuga Santin is a Mexican master’s student currently living in Basel, Switzerland, with a strong academic and creative focus on the intersections of migration, cultural heritage, and identity. With a background in new media, animation, film, and photography, her work bridges critical theory and artistic practice to explore how migrational narratives are represented, preserved, and sometimes overlooked—whether in traditional media or through emerging technologies. She is currently writing her master’s thesis, which proposes a theoretical framework for a large language model specifically fine-tuned on Latin American cultural heritage in Europe, with a particular focus on Switzerland. Her research combines her interests in storytelling, community memory, and digital tools, while also addressing broader concerns about bias, inclusion, and representation within AI systems. Paola’s engagement with photography extends beyond artistic expression. She is equally drawn to its scientific, archival, and educational dimensions, seeing it as a medium capable of recording, communicating, and preserving knowledge in powerful and socially meaningful ways.

This talk aims to look at the cultural and technical challenges of building a large language model focused on Latin American heritage in Europe, specifically in Switzerland, using Meta’s LLaMa 3 architecture. Extracting ideas from Digital Humanities and Machine Learning, I intend to examine how different kinds of bias appear when trying to represent immigrant narratives in large-scale AI systems.
Despite increasing efforts to include more diverse voices in digital collections, Latin American narratives, specifically those based in Europe, are still often missing or scattered across many languages and institutions. This lack of representation affects the model’s training, where English and data from the global North are dominant. As a result, important local and community-based knowledge is left out or misrepresented.
This project aims to work on the building of a multilingual dataset made up of oral histories, community archives, academic writing, and cultural records of Latin American communities in Switzerland and Europe. However, even before training the model, key questions arise: What kinds of data are considered representative? Who decides which voices are important? How do we handle code-switching and mixed-language texts within models trained on formalized text?
I will also suggest ways to reduce bias, such as working with communities to label data, using multilingual fine-tuning, and critically evaluating outputs.
Ultimately, I argue that large language models in the Digital Humanities must be treated not as neutral tools but as interpretative infrastructures that reflect and sometimes can distort the cultural memory we entrust to them.

• New media and animation
• Large Language Models and AI
• Film and photography
• Cultural Heritage
• Science, preservation, and education

• Caught in the Feed: How Algorithms Navigate Us Through Digital Culture
• Navigating Bias in AI Models of Migratorial Heritage