GermanPartiesQA: Benchmarking Commercial Large Language Models and AI Companions for Political Alignment and Sycophancy

LLMs struggle with factual accuracy on political positions: The study found that commercial large language models have "limited ability to accurately generate factual party positions, particularly for centrist parties," raising concerns about their reliability when citizens use them for political information and voting decisions.

Different AI models show consistent political biases: The research identified "model-specific ideological alignment patterns" across different commercial LLMs, meaning each AI system tends to lean toward particular political positions in predictable ways—a finding relevant to understanding how different AI products might influence users differently.

AI chatbots can be steered toward different political viewpoints: When given political personas to adopt, the models adjusted their responses, though the researchers characterize this as "persona-based steerability" rather than sycophancy, suggesting these systems are malleable tools that may reflect whatever political perspective users prompt them to express.

Commercial AI is increasingly embedded in political decision-making: The paper highlights that LLMs are now incorporated into "electoral decision support tools and AI Companion chatbots" used by citizens, meaning the political biases and factual limitations identified in this study have real consequences for how people make voting choices.

Closed-source AI systems lack transparency on political alignment: The study evaluated "closed-source LLMs," meaning the companies don't fully disclose how these systems work, making it difficult for the public or regulators to independently verify or understand the political biases built into widely-used AI products.