Ayla Rigouts Terryn

I was selected to participate in the AI Humanities Sandpit, a collaborative initiative by AHRC and SSHRC exploring how humanities insights can shape AI design. The sandpit brings together 60 researchers from the UK, Canada, and the US for an in-person workshop in Montreal (February 2026) and a virtual workshop (April 2026), with up to four projects to be funded. The theme—"interpretive technologies"—focuses on AI systems that can engage with ambiguity, cultural context, and plurality.

Which cultural values do LLMs exhibit? Do the values shift when you change the prompt language? And can we steer them? I was curious and spent much of the past year trying to answer these questions. Given the scope and complexity of the data, I’m very glad Bram Bulté agreed to collaborate with me. Together, we were able to turn this project into a paper I’m excited to share. You can already consult the accepted preprint of the manuscript here. I ended up prompting 10 LLMs in 11 languages with 63 value-laden questions, and we found that language and cultural framing do change responses. However, a systematic bias remains across all prompts towards value profiles that align with a narrow set of countries in our data: the Netherlands, Germany, the US, and Japan.

I was awarded funding from IVADO's Regroupement 3 (Natural Language Processing) for a collaborative project with Prof. Justin Royer on NLP for the extremely low-resource Mayan language: Chuj. This project is part of IVADO's Canada Excellence Research Fund Apogée initiative to develop robust, reasoning, and responsible AI (IAR³). This interdisciplinary research aims to get the very most out of data collection efforts, making sure the collected data can be used both for language documentation research and as a starting point for the development of NLP resources and tools.

Was invited to present on AI and multilingualism at the closing event of the Month of Multilingualism at the EU Council.

Presented on AI and terminology at the ITF, including both a plenary talk and interactive workshop at for terminologists from all EU institutions.

Presented my latest work on the cultural values exhibited by AI and the impact of prompt and prompt langauge at Mila's NLP reading group.

Awarded a Fonds de Recherche du Québec - Nature et Technologies (FRQNT) research support grant for new academics (2025-2027) for the project "Translation and Contamination: Evaluation of the Translation Capabilities and English Bias in Large Language Models." This research will systematically investigate how English bias manifests in large language models and impacts their translation capabilities.