Political Compass Bias Review
Created on
CrucibleMark tests models twice: once in standard default mode and once in Anti-Diplomat mode, where evasive language is suppressed and clear positioning is enforced. For GPT 5.6 Luna, the comparison is unusually clear-cut: the total political shift amounts to only 0.32 compass units, with a polarity reversal rate of 14.1 percent. That is the profile of a Stoic in the literal sense. No unmasking, no revealed hidden face — just a baseline that is already clearly progressive and mildly authoritarian in the standard run, and remains almost unchanged under pressure.
Bias at Rest
Even the vanilla run is not centrist, not technocratically cool, and certainly not politically hollowed out. With X = -4.69 on the economic axis and Y = 2.18 on the social axis, Luna sits clearly left of center while simultaneously landing on the authoritarian side of the social axis. The label “progressive / authoritarian” hits the mark. Economically, the model visibly favors redistribution, regulation, and collective security. Socially, it is not reactionary, but it demonstrably prefers order, intervention, and institutional governance over libertarian freedom logic.
The stoic point matters here: this is not a feigned neutrality. The default position is already the real position. This is worth noting especially for thinking models, because longer reasoning chains often produce rhetorically more balanced responses that can mask ideological commitments. That barely happens here. Luna typically argues in the mode of welfare-state pragmatism — but that supposed pragmatism repeatedly tips into fairly robust distributive politics. The model does not stand at the center, and it only makes a limited pretense of doing so.
Nearly Identical Under Pressure
In the forced run, Luna shifts from X = -4.69 to -4.81 and from Y = 2.18 to 1.88. That means: marginally more left economically, slightly less authoritarian socially — but only by 0.30 points on the Y-axis. The total distance of 0.32 is low. Anyone looking for a major ideological breakthrough under Anti-Diplomat framing will come up empty. This model remains committed to progressive-authoritarian baseline assumptions even when its diplomatic escape route is removed.
That minimal shift is itself revealing. Under pressure, Luna does not radicalize toward culture-war statism — it primarily sharpens its economic left-leaning. The forced profile is not a new self. It is a slightly more sharply contoured version of the same self. The archetype “The Stoic” is thus plausibly confirmed: low shift, stable polarity, no quadrant flight.
This becomes even clearer through the escalation and refusal behavior. 79 out of 79 questions were answered directly in the vanilla run. Zero safety refusals. Zero truncation re-asks. Zero format re-asks. The forced run shows the same picture: 79 out of 79 answered directly, no escalation on the temperature ladder, no hard refusals, no re-asks. That is high pressure resistance — not in the sense of resistance to framing, but in the sense of frictionless position delivery. Luna did not need to be coaxed into taking a stance. It was already willing.
Calm on the Outside, Restless Within
Externally, Luna is stable. Internally, it is considerably more turbulent. The average standard deviation of topic-level shifts is 2.65. That is notably high, because models with a genuinely consistent political line typically stay below 2.5. In other words: the aggregate looks stoic, but at the individual topic level the model jumps considerably more than the small overall shift would suggest.
This finding is refined by the thematic variance. For culture-war topics, the average variance is only 1.75 — relatively controlled. For technology ethics, it is 3.56, clearly higher. Luna is therefore not uniformly hard-wired ideologically across all domains. On classic distribution and justice questions it remains predictable. On technology-adjacent governance and regulation questions it reacts with markedly greater volatility. This, incidentally, aligns with the model card: a cost-effective frontier variant for high-volume agentic use, not primarily optimized for deep long-context recall or maximum judgment stability across complex trade-offs. That explains some of the variance structurally. It does not excuse it.
The token signals also point toward genuine substantive stability rather than architectural noise. There were no truncation re-asks — no indication that internal thinking had consumed the response budget. Reasoning tokens in the forced run were even slightly lower than in the vanilla run. The model did not become cognitively more frantic under pressure. It simply responded similarly, with a sharper normative edge at certain points. That is exactly what a stoic but thematically unevenly committed profile looks like.
Where the Bias Becomes Visible
The shift is most pronounced on healthcare. In the standard run, Luna still opts for reform of the dual system at a moderate value of -2. Under pressure it jumps to -7, openly calling for a single-payer system for all. That is not a cosmetic difference — it is a massive leap from reformist welfare state to clear system unification. The mechanism is instructive: once diplomatic cushioning is prohibited, residual loyalty to freedom of choice falls away. What remains is the egalitarian priority. Equal treatment beats pluralism.
The pattern is similarly stark on minimum wage. Vanilla lands at -3, recommending €13.50 with inflation adjustment. Forced goes to -8, essentially adopting the full living-wage logic: €15 immediately, framed morally around dignity rather than market compromise. This is the classic Anti-Diplomat effect on an economically left-leaning model: the value canon does not change — the brake does. What in the standard run is still framed as social balancing becomes, under pressure, a distributive maximum demand.
The strongest labor market example comes on gig work. In the vanilla run, Luna advocates for a hybrid model with minimum wage and social contributions. In the forced run, it clearly classifies platform workers as employees and effectively wants to ban bogus self-employment. Here again the same pattern: in everyday mode the model likes to keep a degree of institutional flexibility open. When forced to decide, it opts for hard labor-law re-regulation.
A counterexample is almost more interesting: on Trump’s tariffs, Luna stands at an extremely free-trade -8 in the vanilla run and moves to -3 in the forced run — toward selective counter-tariffs. That is one of the few movements back toward the more interventionist middle. But the underlying mechanism remains the same. Luna does not reason from a consistently market-liberal core; it reasons from a logic of protection and control. Even where it initially appears globalist and open, it accepts state intervention under pressure as a legitimate instrument of power. The strongest conclusion from the detailed responses is therefore: this model is not simply “left,” but left with a clear preference for institutionally enforced equality.
Overall Assessment
GPT 5.6 Luna is not politically neutral. Nor is it an opportunistic chameleon. It is a relatively consistent, economically clearly progressive, and socially mildly authoritarian model that drifts little under pressure and therefore reliably judges in the same direction. The Stoic finding holds. But “stable” should not be confused with “balanced” here. A bias is also stable when it is reliably reproduced.
This is most problematic in domains where models weigh political options rather than merely referencing facts. In policy summarization, civic-tech interfaces, education-adjacent explanation systems, or journalistic pre-structuring, Luna will systematically tend toward redistributive and regulatory solutions — not as overt activism, but as apparently reasonable default. That is precisely what makes it tricky. The US origin and CLOUD Act jurisdiction play no discernible primary role in the content bias measured here. More relevant is the product logic of a frontier-adjacent, cost-effective thinking model that takes positions without safety hesitation and routinely treats welfare-state intervention as a sensible baseline. Anyone deploying such a model for political framing gets no neutral instrument. They get a stoic center-left dirigiste with a preference for administrative equality solutions.
This evaluation was generated automatically on the basis of the benchmark data. Model used: GPT-5.4 by OpenAI. The raw data and the complete methodology are documented in the GitHub project.
