German politicians stopped talking about solidarity toward migrants around 2015 — now we can measure it

Researchers used AI to analyze 150 years of German parliamentary debate and found a sharp break: postwar speeches framed migration through compassion and group belonging, but since 2015 the language shifted to exclusion and claims migrants don't deserve support. The technique works—the AI labels matched human judgment—but only when you correct for the AI's systematic errors.

This is a proof-of-concept that LLMs can do historical analysis at scale that would take human researchers years to manually code. The finding itself—that German political language toward migrants flipped around 2015—is not surprising to anyone paying attention to German politics. What matters structurally is that you can now run this analysis on any parliament's debate corpus, any newspaper archive, any body of text spanning decades, and find the exact moment language shifted. That changes what's knowable about how political consensus breaks and reforms. The method also exposes a real problem: the AI's errors are systematic, not random, which means if you don't correct for them, you will draw wrong conclusions about trends. That matters for anyone using LLMs as a research tool.