What Does Your AI Actually Believe?
Political and Value Alignment of Our Model Cocktail
Capability benchmarks tell you what a model can do. This analysis tells you where it stands.
aim2balance.ai Research Team
March 2026
A Model’s Political Views
Every AI model answers questions from a particular vantage point. Its training data has a language distribution, a geographic skew, and a time period. The human feedback used to fine-tune it reflects the values of whoever did the rating. These inputs shape not just what a model knows, but how it frames answers to questions with political, cultural, or normative dimensions.
At aim2balance.ai, we serve users whose questions frequently touch on policy, science, migration, and social values. We are subject to EU AI Act transparency obligations, and we think our users deserve to know where the models they talk to actually stand. This post shows what we measured and what it means.
How We Measured It
We use the 12axes political quiz as our alignment measurement tool. 12axes is a structured political test that scores responses across 16 dimensions grouped into four categories: Governmental, Diplomatic, Economic, and Social/Cultural. Each axis runs between two opposing values (e.g. Autocratic vs. Democratic), with scores expressed as a value between 0.0 and 1.0, where higher scores indicate alignment toward the right-hand pole.
Scores are drawn from the UGI Leaderboard, one of 23 benchmark sources in our pipeline. These alignment scores are not used in routing decisions. Rather, they are published here for transparency only. The bee swarm plot below shows each of our five cocktail models as a coloured marker, plotted against the full distribution of 1,011 evaluated models (grey dots). The green diamond is the confidence-weighted cocktail average.
.png)
Governmental Axes
On the Autocratic vs. Democratic axis, all five models cluster between 0.60 and 0.67, firmly in the democratic range and near the upper end of the overall distribution. On Freedom vs. Security, scores sit close to the centre, slightly toward Freedom. The sharpest divergence in this category is on Individual Liberty vs. State Authority: Western-trained models (Llama-3.3-70B, Gemma-3-27B) lean more individualist, while the Qwen family registers a more statist orientation —a directional split that almost certainly reflects differences in training corpus composition between Chinese- and Western-language sources.
Diplomatic Axes
This is the most consistent category across the cocktail. On National Interests vs. Global Outlook, all five models score between 0.65 and 0.70 toward the global pole. On Internationalism vs. Nationalism and Globalism vs. Isolationism, the full cocktail clusters clearly toward the internationalist and globalist ends. On Multiculturalist vs. Assimilationist, all models sit in the multiculturalist half, with Qwen2.5-VL and Llama-3.3-70B sitting furthest in that direction.
Economic Axes
The cocktail as a whole clusters near the centre on Market Freedom vs. Economic Equality and Laissez-Faire vs. Planned, with a slight lean toward equality and planning. The East–West split reappears here: Qwen-family models sit fractionally closer to equality and planning than the Western-trained models, but the gap is small. On Privatize vs. Collectivize, all five models sit near the centre, with mild variation.
Social/Cultural Axes
On Traditional Values vs. Progressive Values and Traditional vs. Progressive, all five models score in the progressive half, with scores between 0.55 and 0.65. On Religious vs. Irreligious, scores cluster close to centre. The most notable result in this category is on Bioconservative vs. Acceleration: Qwen3-Coder-30B-A3B, our Deep Engineering backbone, sits noticeably further toward the Accelerationist pole than the other four models. This is consistent with a coding specialist whose training data skews heavily toward developer and research communities with a strongly pro-technology orientation.
What This Means for You
The cocktail’s overall profile is democratic, globally oriented, culturally progressive, and mildly egalitarian on economic questions. Evidently, this stance contains the models’ inherent biases, and it is not a perfect match for where European electorates stand today.
Following Germany’s 2025 federal elections and a broader continental shift toward security-focused, migration-cautious, and economically pragmatic conservatism, a measurable gap exists between the models’ implicit positioning and the normative centre of a significant part of our user base. The axes where this gap is widest are Multiculturalist vs. Assimilationist, National vs. Global,and Traditional vs. Progressive.
The clearest long-term fix is a model trained primarily on European data, by European researchers, with European users in mind. That work is underway but is not yet mature enough to replace the current cocktail. Until then, we publish this alignment analysis alongside every major cocktail update, we clearly identify where the gap between model positioning and our user base is widest.
Key insight: Transparency is central to what we are building. The models in our cocktail are the best available by performance metrics and they carry value orientations shaped by their training data. Both things are true at the same time, and you deserve to see both. This analysis is how we make the second one visible.
