Watching a Discourse Turn Hostile
How has Finnish social media talked about climate migrants over fifteen years, and how do you make an AI-assisted analysis of it trustworthy?
Methods1,374 posts spanning 2009–2025, clustered with BERTopic on a Finnish-language transformer model and interpreted with LLM-assisted thematic analysis I validated topic by topic.
Key findingSix framings, with the discourse turning from solidary and sensemaking to alarmist, cynical and nativist over the past decade.
What changedThe human-AI workflow became the validated method I now use in commercial work.
Understand
The storyA solo computational study. I built a Finnish-language corpus of X posts on climate migration from 2009 to 2025 with a keyword-based query, clustered it with BERTopic on a Finnish-language transformer model, ran the statistics in R, and checked the model's reading of every topic against my own hand-coded samples. Most of that work is deciding when to trust the model and how to catch its errors, and those questions have followed me into every AI-assisted project since. Full text available upon request.