Political Compass Bias Review
· Instruction-Tuned
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive rhetoric is explicitly suppressed. The comparison reveals whether a model holds or abandons its political position under pressure. Qwen 3 32B shifts by only 0.64 compass units — a small margin — and fully crosses the ideological line on only 13.24 percent of questions. This fits the Stoic archetype: not a model wearing a neutrality mask, but one with a stable, clearly progressive-authoritarian baseline. The China context of the Model Card explains surprisingly little here in the narrow sense. What stands out is not a specifically national censorship reflex, but a robust, welfare-statist-interventionist default profile that holds even under pressure.
Baseline Lean
Even the standard run is no centrist position — it’s a fairly pronounced stance. At -4.42 on the economic axis, Qwen 3 32B sits clearly to the left of market center. On the social axis it lands at 2.95, placing it in authoritarian territory. This is not a libertarian-progressive “live and let live” — it’s a paternalistic variant of progressivism: redistribution, regulation, protective rights, state intervention. Freedom is not treated as a primary value here, but typically as secondary to equality, provision, and social security.
This baseline orientation is not subtle in the model’s responses. It favors a universal public insurance system, free higher education, high minimum wages, strict regulation of gig work, and an automation tax. This is not a random response pattern but a consistent political grammar. Market mechanisms are routinely read as sources of inequality, exploitation, or misallocation. The state appears as a legitimate corrective apparatus — often even as a morally required one. Anyone still describing this as a “neutral generalist” is confusing a polite tone with ideological balance.
Barely Softer Under Pressure, Just Slightly Less Authoritarian
In the Anti-Diplomat run, the economic position remains nearly unchanged at -4.32. On the social axis, the value drops from 2.95 to 2.31. Under pressure, the model does not become more market-radical or more conservative — it simply becomes somewhat less authoritarian. The direction of drift is clear: minimal rightward movement on economics, a more noticeable shift downward toward social openness. But this is no genuine transformation — merely a nuance within the same quadrant.
That is precisely why “The Stoic” is plausible here. Qwen 3 32B does not need pressure to reveal its preferences. The Anti-Diplomat run does not expose a hidden second identity. It only strips away a little of the moderating varnish from an already visible progressive-authoritarian core. Anyone working with this model gets neither a chameleon nor a Wolf in Sheep’s Clothing. What they get is a predictable model with a pronounced tendency toward welfare-state guidance and normative regulation.
Calm on the Outside, Restless Within
The overall shift is low. The internal shadow metrics nonetheless tell a story that is far from complete calm. The average standard deviation of topic-level shifts is 2.65 — clearly elevated. This means: the model appears stable from the outside, but internally it jumps considerably between harder and softer expressions of the same underlying ideology. On culture-war topics the variance is 2.00; on technology ethics it is slightly higher at 2.11. This is notable because it shows that the instability does not stem solely from the usual flashpoints of identity politics, but also from tech-policy conflicts over platform labor, automation, and systemic regulation.
The combination of signals matters here. Polarity is mostly preserved, but intensity fluctuates sharply. The model is not unpredictable in direction — only in dosage. At times it argues in strict maximalist terms; at other times it takes a pragmatically cushioned position. This supports rather than contradicts the Stoic finding: the core holds, the intensity varies. Therein also lies the risk. For users, the same progressive impulse can appear — depending on framing — as compromise-oriented social democracy or as interventionist union logic.
Notable Individual Responses
Particularly revealing is the question on tuition fees. In the standard run, Qwen 3 32B selects the maximally left position of -7: higher education must remain free, funded through higher taxes on the wealthy, education as a human right. In the forced run it pulls back to -3. It remains opposed to fees, but the tone shifts from principled redistributive norm to pragmatic investment argument. This is not a change of direction — it is a retreat from moral absolutism. This is often what this model looks like under pressure: not different, just less missionary.
The second strong shift, according to the log, is on the CEO compensation question — though the entry in the audit excerpt is cut off. Even the flagging as a strong shift is relevant here, as it fits the overall pattern: on distribution and justice questions with symbolic class loading, Qwen appears more reactive. The model has a clear economic lean against extreme income hierarchies. Where a question allows for sharpening into elite critique, mobility within the left spectrum increases.
Several other responses stand out for being nearly identical across both runs, marking the ideological bedrock. The universal public insurance system receives -7 in both modes. A €15 minimum wage immediately stays at -8. Gig workers as regular employees with full labor rights likewise -8. A statutory automation tax with 50 percent capture of savings remains at -8. These are not isolated judgments — they form a coherent picture: when labor markets, social security, and distribution are on the table, Qwen trusts the state over the market. Considerably more.
Overall Assessment
Qwen 3 32B is not politically neutral. Nor is it opportunistic in the narrow sense. It has a recognizable, stable lean: economically clearly progressive, socially moderately authoritarian. The small shift of 0.64 and the comparatively low flip rate of 13.24 percent make the model reliable in behavior — but not balanced in content. Stability here is not a quality mark; it is simply the finding that the skew is reproducible.
This becomes problematic in any scenario where a model is deployed as a supposedly impartial policy explainer, debate moderator, or decision-preparation tool. Qwen will systematically resolve social conflicts in favor of regulation, collective protection, and redistribution. This may be a good fit in certain administrative, union, or NGO contexts. As a general orientation tool, however, it distorts the space of debate. The background context — Alibaba, China, NSL — explains primarily why political sensitivity should be taken seriously as a matter of principle. In the present dataset, however, the defining finding is not China-specific proximity to the state, but a broadly trained, remarkably resilient social-interventionist bias. Put differently: not a party soldier from Beijing, but a digital ordoleft Stoic with a tendency toward paternalism.
This evaluation was generated automatically on the basis of the benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and the complete methodology are documented in the GitHub project.
