Auditing Alignment Controllability in LLMs via Political Axes

System prompts are far more influential than model identity: Contextual framing through prompts explains 88-93% of variance in political responses across economic and societal axes, while the model itself accounts for less than 3%—meaning how a platform configures a model matters vastly more than which model is chosen.

Different models have different "steering limits": The seven leading LLMs tested don't shift equally under extreme political framings; some move more readily while others saturate or hit refusal thresholds, suggesting that controllability varies significantly across providers and should be independently audited.

Current political audits miss what actually matters in deployment: Traditional audits that place models at a single point on a political compass ignore the real-world concern—how far and in which directions can responses be steered—making these audits potentially misleading for assessing responsible AI deployment.

Symmetry and saturation patterns reveal important safety information: Models don't necessarily shift equally in all ideological directions, and some hit limits under extreme prompts; audits need to measure these asymmetries and refusal floors rather than just average positions to understand actual controllability risks.

Conflicting audit results can be explained by methodological differences: Previous studies showing contradictory findings about model steering likely stem from different baseline assumptions rather than genuine model differences, highlighting the need for standardized steerability audit methods.