LLM Study Traces 150 Years of German Migration Rhetoric, From Postwar Inclusion to Recent Backlash

Researchers have leveraged large language models to conduct a sweeping analysis of German political discourse on migration spanning over a century and a half. By automating the annotation of parliamentary debates, the team was able to track how attitudes toward migrants have evolved from the postwar era through today, revealing a sharp reversal in recent years.

Decades of Solidarity Give Way to Backlash#

The study examined three major periods of German parliamentary history: the Reichstag (1867–1933), the West German Bundestag (1949–1990), and the reunified Bundestag (1990–2025). The findings show that the immediate postwar period (1949–1957) was marked by relatively strong expressions of solidarity with migrants, especially in compassion-based and group-based forms. However, this pattern shifted sharply after 2015, when anti-solidarity sentiment began to rise noticeably in parliamentary speech.

Validating LLM Analysis at Scale#

Analyzing such vast historical archives would traditionally demand painstaking manual review by human annotators. The research team tested multiple large language models—including GPT-5 and gpt-oss-120B—to determine whether they could reliably classify different types of solidarity and anti-solidarity language. The strongest models achieved accuracy levels comparable to human inter-rater agreement, though the researchers identified systematic biases in how the models made their classifications.

To address these shortcomings, the team combined soft-label outputs from the models with a statistical technique called Design-based Supervised Learning (DSL), which helped correct for systematic errors and produced more reliable trend estimates over the long term.

Implications for Social Science Research#

The work demonstrates that LLMs can meaningfully assist in large-scale automated text analysis for social science, but the output requires careful scrutiny. The researchers emphasize that when systematic errors are present in model predictions, they must be statistically adjusted to avoid skewing historical inferences. This approach offers a potential template for researchers seeking to apply AI to the analysis of extensive documentary archives.