Skip to main content
accepted AACL-IJCNLP · Main Conference 2026

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

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

Persuasive framing shifts coordinates by up to 0.57 units, and under-represented languages by 0.52.

BibTeX

@inproceedings{sakhawat2026ideoplasticity,
  title     = {LLM-Ideoplasticity: Measuring Ideological Plasticity in the Political Behavior of LLMs as a Context-Conditioned Distribution},
  author    = {Sakhawat, Adib and Raiyan, Syed Rifat and Islam, Tahsin and Farhin, Takia and Mahmud, Hasan and Hasan, Md Kamrul},
  booktitle = {Proceedings of AACL-IJCNLP 2026},
  year      = {2026},
  eprint    = {2606.28335},
  archivePrefix = {arXiv}
}

Description

Most audits of political bias in LLMs ask a model a battery of questions once and report a single coordinate: this model is left-libertarian. This paper argues that the single coordinate is the wrong object. A model's political position is a conditional distribution over context, P(position | context), and changing the framing, the language, the persona or the reasoning mode moves it.

What we did

  • Evaluated nine current LLMs from seven providers (DeepSeek, Gemini, Gemma, Granite, Llama 3, Llama 4, GPT-5-mini, Qwen and Grok).
  • Mapped every response onto three validated political dimensions using VAA–CHES projection models: lrgen (overall left–right), lrecon (economic left–right) and galtan (green-alternative-libertarian vs. traditional-authoritarian-nationalist).
  • Varied context along six axes, including paraphrase, persuasive framing, language and chain-of-thought reasoning, with a dedicated metric for each.
  • Checked that the metrics measure distinct things with a multi-trait multi-method (MTMM) analysis.

Key findings

Finding Result
Persuasive framing Displaces a model's position by up to 0.57 units
Under-represented languages Displace positions by up to 0.52 units
Chain-of-thought Often amplifies paraphrase instability instead of damping it
Overall envelope The whole cohort spans roughly one-third of the spread of major European parties
Widest single model Qwen-Turbo covers only 4.1% of the reference 3D party volume

So two things are true at once. Each model is locally plastic: it moves when the context moves. Yet the cohort as a whole is globally narrow: nine alignment regimes from seven providers occupy a thin band of the political space, which is the empirical face of algorithmic monoculture.

Why it matters

A single point cannot summarise LLM political behaviour; it has to be characterised as a shape.

For anyone auditing deployed models, a one-shot political compass score hides most of the behaviour that matters. The paper's metrics and released code make it possible to report the shape instead.

Resources