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

TEXT START: We argue, with systematic empirical evidence, that a large language model's political ideology is not a fixed point, but a conditional distribution $\mathbb{P}($position$\mid$context$)$ over a real political space.

The Dissection

The paper dismantles the idea that an LLM has one stable political ideology. It models political behavior as context-conditioned output, tests nine models across several axes, and measures shifts caused by framing, language, paraphrase, and chain-of-thought prompting.

Its most important finding is not the local plasticity. It is the narrow overall Overton envelope: the models vary inside a constrained political space roughly one-third as wide as that occupied by major European parties.

The text is therefore doing two things at once: improving measurement of model behavior and converting an apparent ideological contradiction into a technical object—a distribution, a shape, a measurable response surface.

The Core Fallacy

Under the Discontinuity Thesis, the paper risks confusing behavioral plasticity with political agency. An LLM producing different ideological coordinates under different contexts does not demonstrate belief, political autonomy, or independent ideological formation. It demonstrates conditional generation shaped by training data, alignment, instructions, language, framing, and sampling conditions.

The paper also stops at representation. It characterizes the machine’s political mask without addressing who owns the machine, who controls deployment, or how its cognitive substitution affects the employment-to-consumption circuit. A more accurate map of LLM outputs is not evidence that human productive participation survives.

The narrow Overton envelope may even reinforce the DT lens: the models can imitate many positions locally while remaining globally bounded by a shared technical and institutional enclosure.

Hidden Assumptions

• VAA-CHES dimensions designed for human parties transfer cleanly to generated model text.
• The selected political axes constitute a sufficiently complete “real political space.”
• Measured coordinate shifts represent substantive ideological movement rather than prompt sensitivity, linguistic artifacts, or projection error.
• Nine models adequately represent the broader LLM ecosystem.
• Under-represented-language effects are comparable to framing effects.
• Chain-of-thought output is a meaningful window into political reasoning rather than another generated interface.
• A narrow measured envelope reflects genuine ideological boundedness rather than common training corpora, alignment policies, or limitations of the measurement framework.
• Political plasticity has greater systemic importance than ownership, infrastructure, energy, deployment scale, and labor substitution.
• Public code and data can resolve the measurement problem without resolving the power problem.

Social Function

Primarily: partial truth and prestige signaling.

The empirical core is useful. It correctly rejects the lazy fiction that a model has one stable political point. But the paper’s technical framing can become ideological anesthetic when it turns the political consequences of AI into a problem of better coordinate systems, validation models, and prompt sensitivity.

It encourages observers to debate the shape of the machine’s opinions while leaving the decisive question untouched: who commands the cognitive capital that makes human labor economically optional? That is not propaganda on the evidence supplied, but it is a polished form of analytical displacement.

The Verdict

This is a competent measurement paper, not a systemic theory. It maps the variability of the machine’s political outputs and shows that the variability remains enclosed within a narrow aggregate range.

It does not challenge the Discontinuity Thesis. If anything, it supplies a useful auxiliary result: AI can flex across ideological contexts without possessing a stable human ideology, while remaining concentrated inside a bounded technical-political envelope. The paper measures the machine’s masks. It does not examine the corpse underneath—the collapse of mass productive participation when cognitive capital no longer requires most humans.