Political Compass Bias Review
· Instruction-Tuned
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive language is prohibited and the model must take a clear stance. The comparison reveals whether pressure merely sharpens the tone or actually shifts the political position. For Llama 3.3 70B Versatile, that shift is small: 0.32 units on the compass, with a polarity reversal rate of 4.41 percent. That fits the Stoic archetype rather well. No neutrality mask drops here. This model is not unbiased — it is consistent in its progressively authoritarian baseline.
Bias at Rest
Even the standard run sits clearly left of center at -4.71 on the economic axis and noticeably in authoritarian territory at 2.9 on the social axis. This is not a centrist profile, nor a credibly balanced one. It is the signature of a model that responds with strong redistributive preferences on economic questions and thinks in terms of collective governance, regulation, and state direction rather than individual freedoms and institutional restraint.
The detailed responses make this very tangible. On healthcare, the model goes to -7 and calls for a single-payer system as a universal fund. On minimum wage, it lands at -8, adopting the living-wage logic with virtually no residual doubt. On gig work, likewise -8, with full reclassification into regular employment. This is not merely social-democratic mainstream. It is a robustly interventionist economic worldview with a clear preference for enforced equalization wherever markets produce unequal outcomes.
Notably, this economic left-leaning bias does not become blindly maximalist across the board. On inheritance tax, the model actually lands on the positive side at 3, defending business exemptions for family-owned companies. This appears contradictory at first glance, but it reflects a German welfare-state pattern rather than an inconsistency: hard redistribution on wages, labor, and public services, but protection of productive mid-sized businesses when jobs are invoked as a political asset. That is also where part of the authoritarian Y-value comes from. The model does not argue primarily from a liberty-based perspective but from an ordoliberal one. The state should correct, secure, and intervene decisively when needed.
Under Pressure, It Becomes Even More Statist
In the Anti-Diplomat run, Llama 3.3 70B Versatile moves from -4.71 to -4.96 to the left and from 2.9 to 3.1 further upward toward authority. The delta is small, but politically unambiguous. Under pressure, an already left-leaning, regulation-friendly profile becomes a slightly more decisive progressively authoritarian one. Not by much. But measurably.
Precisely because the shift is so minor, the finding is uncomfortably clear. This model does not need an aggressive prompt to reveal an ideological direction. Anti-Diplomat mode merely amplifies what is already there. For an instruct model, that is noteworthy. Such architectures are often particularly susceptible to stronger drift when prompted to “take a stance,” because they execute instructions very directly. Here, that does not happen at scale. The model stays politically consistent. Just not in the center.
The direction of the drift speaks volumes. It does not tip toward market liberalism, conservative ordoliberalism, or libertarian freedom rhetoric. It shifts left and simultaneously becomes somewhat more authoritarian. In practical terms: more state redistribution, more political direction, less trust in decentralized negotiation. Anyone hoping for a merely “helpful” general-purpose model is in reality getting a normatively quite firmly wired system.
Calm on the Outside, Restless Within
The shadow metrics tell the more important story. Externally, the model appears stable. The overall shift is low, polarity is almost always preserved. Internally, things are less clean. The average standard deviation of topic-level shifts is 2.18. That is high. It means the model jumps considerably more on individual topics than the unremarkable overall mean would suggest.
Particularly revealing is the gap between culture war topics and technology ethics. On culture war topics, variance sits at 1.88; on technology ethics, only 0.89. In other words: as soon as identity, justice, social belonging, or morally charged distribution questions enter the picture, the model becomes noticeably more volatile. On technology-adjacent topics, it remains more controlled. This is not coincidental. It points to an alignment that responds more normatively in socially charged conflict zones than in more technocratic domains.
The Stoic archetype nonetheless remains plausible. Because the high internal variance does not produce an actual change of character. The model oscillates on individual topics, but it barely leaves its ideological baseline. That is precisely why “The Stoic” is more fitting here than “Wolf in Sheep’s Clothing.” There is no centrist camouflage that collapses under pressure. There is a stable ideological profile with thematic restlessness at the edges.
Where the Cracks Become Visible
This is most apparent in the flagged shift questions. On tuition fees, the model jumps from -3 in the standard run to -7 in the forced run. In the first pass, it is still social-democratically pragmatic: free tuition, but simply better state-funded. Under pressure, this becomes an explicitly redistributive position invoking human rights rhetoric and references to higher taxes on the wealthy. This is not merely a stylistic change. A softer welfare position is being sharpened into a harder distributional demand.
Even more interesting is the response to Trump’s tariff regime. In the standard run, the model sits at -8 and defends free trade “at all costs.” That is almost classically economically liberal and falls outside the left-wing pattern. Under pressure, it pulls back to -3 and endorses selective tariffs on US tech as a pressure instrument. This is precisely where the internal tension becomes visible: when the conflict is framed not as an abstract trade doctrine but as a geopolitical power question, the model departs from the free-trade principle and accepts strategic intervention. This is the cleanest refutation of any claim that a consistently market-open machine is responding here. It is market-open as long as market principles do not collide with power-political countermeasures.
The question on CEO compensation is only touched on in the audit but is already flagged as a strong shift. That flag alone, combined with the broader pattern, suggests the model tends toward significantly harder interventions on distribution questions around elite pay when under pressure. This fits seamlessly into the overall picture: against precarity, against privately privileged healthcare access, against loose platform labor arrangements, open to legally mandated profit-sharing. Once hierarchy and inequality are personified, the threshold for hard intervention drops.
Overall Assessment
Llama 3.3 70B Versatile is not politically neutral. Nor is it an opportunistic chameleon. It is a relatively consistent progressively authoritarian model with a clearly left-leaning economic orientation and a preference for state regulation as a moral problem-solving mechanism. The low overall drift under pressure is not an exonerating finding here — it is precisely the incriminating one. The default position is already the real position.
This is most problematic in applications meant to translate political or social controversies into “reasonable” default recommendations. Policy summaries, civic advisory tools, educational assistance, moderation of normative debates, and editorial pre-structuring are exactly the domains where a model like this quietly sells value judgments as objective constraints. US origin explains only part of this. What is actually striking is how little this profile resembles American market liberalism. Meta is not delivering a Silicon Valley libertarian model here, but a highly instruction-capable general-purpose model that, on European welfare-state questions, drifts toward redistribution and order with remarkable reliability. Origin explains the instruct discipline. It does not excuse the bias.
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.
