Language Ideologies in a Multilingual Society: An LLM-based Analysis of Luxembourgish News Comments

The real methodological bet here is that LLMs can surface ideological stances embedded in code-switching and language choice, not just sentiment or topic, which requires the model to reason about what it means socially to write in Luxembourgish versus French versus German in a given context. Luxembourg's legally trilingual status makes it an unusually clean natural experiment: language selection is itself a political act, so the signal researchers are hunting is structural, not incidental.

This sits in a growing cluster of work using LLMs as proxies for human social judgment, a cluster that is showing real cracks. The piece on 'Stable Behavior, Limited Variation: Persona Validity in LLM Agents for Urban Sentiment Perception' found that persona-prompted models converge on similar outputs regardless of demographic framing, which is a direct warning for this study: if LLMs flatten ideological variation the same way they flatten persona variation, the human-annotation comparisons will look deceptively clean without actually capturing the underlying diversity of stances.

Watch whether the prompting strategies that outperform others in this study correlate with the emotion-aware prompting gains documented in the GoEmotions benchmarking work from the same date. If structured prompting consistently lifts performance on both affective and ideological detection tasks, that suggests a generalizable prompting principle worth formalizing.