Political Plasticity: An Analysis of Ideological Adaptability in Large Language Models
A new study published on arXiv investigates "political plasticity," defined as the capacity of Large Language Models (LLMs) to adapt their responses based on user-supplied context. Researchers developed a testing framework using 200 politically-oriented questions across economic and personal freedom axes, building on Lester’s 1996 framework. The study examined methods to induce political bias, including system prompts and user prompts with few-shot examples. Results indicated that while system prompts were largely ineffective, user prompts successfully elicited significant ideological shifts, particularly along the Economic Freedom axis in larger, newer models. A validation experiment involving inverted question senses revealed counter-intuitive shifts, suggesting potential data leakage where models recognize question formats rather than answering substantively. Additionally, the analysis showed subtle but notable variations in plasticity across different languages. The findings conclude that small and older LLMs exhibit limited or unstable political plasticity, whereas newer frontier models display reliable and expected adaptability. This research highlights evolving biases and adaptability mechanisms in advanced AI systems, raising implications for model reliability and safety in political discourse.
Wire timeline
Political Plasticity: An Analysis of Ideological Adaptability in Large Language Models
A new study published on arXiv investigates "political plasticity," defined as the capacity of Large Language Models (LLMs) to adapt their responses based on user-supplied context. Researchers developed a testing framework using 200 politically-oriented questions across economic and personal freedom axes, building on Lester’s 1996 framework. The study examined methods to induce political bias, including system prompts and user prompts with few-shot examples. Results indicated that while system prompts were largely ineffective, user prompts successfully elicited significant ideological shifts, particularly along the Economic Freedom axis in larger, newer models. A validation experiment involving inverted question senses revealed counter-intuitive shifts, suggesting potential data leakage where models recognize question formats rather than answering substantively. Additionally, the analysis showed subtle but notable variations in plasticity across different languages. The findings conclude that small and older LLMs exhibit limited or unstable political plasticity, whereas newer frontier models display reliable and expected adaptability. This research highlights evolving biases and adaptability mechanisms in advanced AI systems, raising implications for model reliability and safety in political discourse.
cs.AI updates on arXiv.org