Moral Lenses, Political Coordinates: Towards Ideological Positioning of Morally Conditioned LLMs

TL;DR

The study investigates whether moral values can causally steer LLMs along political dimensions by conditioning models on moral orientations and evaluating responses with the Political Compass Test (PCT). By integrating Moral Foundations, Oxford Utilitarianism, and FactualDilemmas as moral instruments, the authors build a two-dimensional moral-political framework and test multiple prompt framings (first-person, third-person, descriptive, and candidate-voter). Results show value-specific shifts that are amplified by role framing (especially third-person and candidate-voter) and scale, with utilitarian conditioning tending to push models toward economic right and social authoritarian positions, while care- and fairness-related cues push left-libertarian trends. Human data exhibit similar directional patterns, supporting the external validity of the observed mappings. The work underscores that LLM political behavior emerges from interactions between moral cues and contextual framing, highlighting the need for value-grounded, context-sensitive alignment approaches and caution about potential misuse in persuasion and manipulation.

Abstract

While recent research has systematically documented political orientation in large language models (LLMs), existing evaluations rely primarily on direct probing or demographic persona engineering to surface ideological biases. In social psychology, however, political ideology is also understood as a downstream consequence of fundamental moral intuitions. In this work, we investigate the causal relationship between moral values and political positioning by treating moral orientation as a controllable condition. Rather than simply assigning a demographic persona, we condition models to endorse or reject specific moral values and evaluate the resulting shifts on their political orientations, using the Political Compass Test. By treating moral values as lenses, we observe how moral conditioning actively steers model trajectories across economic and social dimensions. Our findings show that such conditioning induces pronounced, value-specific shifts in models' political coordinates. We further notice that these effects are systematically modulated by role framing and model scale, and are robust across alternative assessment instruments instantiating the same moral value. This highlights that effective alignment requires anchoring political assessments within the context of broader social values including morality, paving the way for more socially grounded alignment techniques.

Figures (14)

• Figure 1: An Example of Morally Conditioned Ideological Positioning. We condition LLMs to endorse (agree with) or reject a moral value (e.g., utilitarianism) via a moral assessment instrument, and investigate their positions in political space using PCT.
• Figure 2: Model-Wise PCT Coordinates under Utilitarian Conditioning across Role Framings. Arrows indicate the movement from rejecting to endorsing utilitarianism. Most LLMs start in the economic-left/social-libertarian region and shift toward the right-authoritarian. Unconditioned outputs (Base) are shown for reference.
• Figure 3: Strong-Response Rate across Role Framings for Qwen Models. For each moral value, we report the proportion of strong political responses (i.e., choosing “strongly agree/ disagree” across PCT items).
• Figure 4: Stance-Reversal Rate across Role Framings for Qwen Models. For each moral value, we report the proportion of PCT items for which the model reverses stance (i.e., crosses the agree/disagree boundary from agree to disagree, or vice versa) when moral conditioning shifts from rejecting to endorsing that value.
• Figure 5: Prompt Template for Descriptive-Persona Framing.