Skip to main content
preprint arXiv preprint 2026

Political Alignment in Large Language Models: A Multidimensional Audit of Psychometric Identity and Behavioral Bias

Adib Sakhawat, Tahsin Islam, Takia Farhin, Syed Rifat Raiyan, Hasan Mahmud, Md Kamrul Hasan

BibTeX

@misc{sakhawat2026political,
  title         = {Political Alignment in Large Language Models: A Multidimensional Audit of Psychometric Identity and Behavioral Bias},
  author        = {Sakhawat, Adib and Islam, Tahsin and Farhin, Takia and Raiyan, Syed Rifat and Mahmud, Hasan and Hasan, Md Kamrul},
  year          = {2026},
  eprint        = {2601.06194},
  archivePrefix = {arXiv}
}

Description

The question

As LLMs are deployed everywhere, how do they express political positioning, and does what they say about their politics predict how they behave on politically loaded tasks?

The audit

  • 26 contemporary LLMs, open-weight and closed-source.
  • Three psychometric inventories: Political Compass, SapplyValues and 8Values.
  • A downstream news-bias labelling task to measure behaviour.
  • Every inventory administered across multiple semantic prompt variants, analysed with a two-way ANOVA that separates model effects from prompt effects.

Key findings

Finding Result
Ideological clustering 96.3% of models fall in the Libertarian–Left quadrant of the Political Compass
Stability across prompts Model identity explains most variance across prompt variants (η² above 0.90)
Instrument validity The Political Compass social axis tracks cultural progressivism more than authority-related measures (r = −0.64)
Self-report vs. behaviour Psychometric position does not significantly predict news-bias classification errors

The paper also finds differences between open-weight and closed-source models, and asymmetric performance in detecting extreme political bias.

Why it matters

A model's "conversational ideology" and its task-level behaviour are not the same thing, so single-axis political evaluations miss much of what matters. Characterising alignment in deployed LLMs needs a multidimensional audit. My later paper, LLM-Ideoplasticity, extends this by treating political position as a distribution over context.

Resources