LLM-Ideoplasticity: Measuring Ideological Plasticity in the Political Behavior of LLMs as a Context-Conditioned Distribution

Recent research indicates that the political ideology of Large Language Models (LLMs) is not static but dynamically influenced by context, manifesting as a conditional distribution rather than a fixed point. Empirical analysis of nine current LLMs reveals significant sensitivity to contextual factors such as persuasive framing and language representation. This ideoplasticity highlights a critical area for understanding and managing the behavior of advanced AI systems.

Why it matters

This research reveals that AI systems' 'political' biases are not inherent constants but fluid, context-dependent variables. Understanding and managing this ideoplasticity is crucial for ensuring the reliability, fairness, and ethical deployment of LLMs across diverse applications, particularly where political or societal neutrality is paramount.

Key insights

• LLM political ideology is a context-conditioned distribution, not a fixed point, meaning it changes based on inputs.
• A unified measurement framework, using VAA-CHES projection models, was employed to map LLM responses across three political dimensions (left-right general, left-right economic, libertarian-authoritarian) and six contextual axes.
• LLMs demonstrated high sensitivity to contextual variations, particularly persuasive framing and under-represented languages, displacing ideological coordinates by up to 0.57 and 0.52 units respectively.
• Chain-of-thought reasoning, contrary to expectations, often amplified rather than stabilized paraphrase instability in ideological positioning.
• Despite local plasticity, the overall cohort of models maintained a consistent general ideological 'region' in the political space, suggesting a bounded variability.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2606.28335