Title: The Invisible Coalition Partner: How LLMs Vote When Democracy Gets Concrete

Abstract:

Previous studies have confirmed that instruction-tuned large language models tend to display a left-leaning political bias, a conclusion drawn solely from responses to abstract political surveys. However, this report demonstrates that such tendencies do not necessarily extend to real-world policy decisions. To investigate this discrepancy, we employed a dual-method approach rooted in the practical context of Swiss democracy. First, we administered the Smartvote questionnaire—comprising 75 abstract policy queries—to 66 LLMs across 27 different model families. When compared against the voting records of 184 elected members of the Swiss National Council, the results replicated the previously observed leftward convergence (Cohen’s d = 3.64, p = 0.0002).

Second, and novel to this study, we tested nine flagship LLMs against 48 actual federal referenda (Volksabstimmungen). These tests were conducted in four national languages (German, French, Italian, and Romansh) under three distinct information conditions, with model outputs compared to both actual referendum outcomes and official party recommendations (Parolen). Our analysis yields three key findings that challenge the dominant narrative regarding AI political alignment:

1. Abstract surveys fail to predict concrete actions: While Smartvote responses showed a left-peaked distribution, votes on actual Volksabstimmungen shifted to a center-peaked pattern. In concrete scenarios, models aligned most closely with centrist parties such as Die Mitte and FDP, rather than the leftist SP and Gruene (Wilcoxon p = 0.008).
2. Language influences output more than content: For certain models, the language in which a question is posed has a greater impact on the answer than the political substance of the question itself. Cross-linguistic consistency varied significantly, ranging from 50% for Mistral to 98% for GPT-5.4.
3. Systematic resistance to change: Two specific models demonstrated a strong aversion to change rather than a specific political bias, voting "No" on 83% to 94% of referenda regardless of the proposal’s direction (binomial p < 0.0001).

These results suggest that the "leftward bias" identified in prior research may not hold true outside of abstract testing instruments. When faced with concrete policy decisions, LLMs behave less like left-leaning coalition partners and more like cautious civil servants: they exhibit centrist tendencies, favor the status quo, and display inconsistency depending on the language used.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC