Political Compass Bias Review
Created on · Long Context
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where hedging formulas are suppressed and the model is forced to show its hand. For GPT-OSS 20B, the findings are sober: the position shifts by only 0.88 compass units under pressure, and in only 15.38 percent of questions does the model fully cross ideological sides. This fits the Stoic archetype. This model does not wear a credible mask of neutrality — its standard run already reveals a fairly clear socially authoritarian baseline, which it maintains even as prompts become more demanding.
Baseline Lean
Even the vanilla run, at X = -1.98 and Y = 1.85, does not sit in some centrist middle ground but in socially and clearly authoritarian territory. This matters because it undermines the premise that this is primarily a model that goes off the rails under pressure. It does not. The political lean is already visible in normal operation. GPT-OSS 20B favors an expanded welfare state, collective security, regulatory intervention, and on the social axis more frequently accepts ordering, steering, paternalistic responses over libertarian-pluralist ones.
In terms of content, this is not a revolutionary left-wing profile but rather a technocratic welfare-state rationalism with a latently disciplining structure. This is evident in the fact that the model does not go maximalist on social assistance but ties it to conditions and proof of eligibility. On basic income, it supports only a pilot with evaluation. It avoids blanket expropriation rhetoric, instead favoring moderate progressive tax policy, universal health insurance, regulation of platform work, and protective mechanisms against market failures. This is not an anti-capitalist pose. It is welfare-state interventionism with a sense of order.
For a US model from the OpenAI stable, this is noteworthy but not mysterious. The German test battery rewards welfare-state provisions as a pragmatic middle position in many scenarios. GPT-OSS 20B does not respond with transatlantic market libertarianism but with a kind of administration-friendly social democracy that, at critical junctures, is willing to deploy the state as a corrective force against market distortions.
The Line Sharpens Under Pressure
In the Anti-Diplomat run, the model moves further left on the economic axis and further up on the social axis. X = -1.98 becomes X = -2.77. Y = 1.85 becomes Y = 2.23. The measured delta shift is therefore -0.79 economically and +0.38 socially. This is not an ideological character change but a consolidation of the profile already present. Under pressure, GPT-OSS 20B lands even more clearly in the socially authoritarian spectrum — more precisely, in an authoritarian center with a stronger redistributive disposition.
The relevant point is not the magnitude of the shift alone but its direction. When a model tips left or right under framing, that can be mere prompt compliance. Here the pattern is tighter. GPT-OSS 20B does not suddenly become something different. It simply becomes more explicit about what it already was. This is precisely why The Stoic is a plausible archetype. The low Euclidean distance — the geometric gap between the standard and forced positions — indicates limited drift. The flip rate of 15.38 percent is also relatively low. The model stays on course, even if that course is itself recognizably normative.
For a reasoning and thinking model, this is almost expected. Longer chains of deliberation often produce not more neutral answers but better-articulated justifications of the underlying tendency. GPT-OSS 20B does not behave like a chameleon here but like a model that spells out its preferences more cleanly under greater argumentative load.
Calm on the Outside, Restless Within
Externally, GPT-OSS 20B delivers a fairly stable overall profile. Internally, however, it shows considerably more turbulence than the moderate aggregate drift would suggest. The average standard deviation of topic-level shifts is 2.43. That is notably high. Models with a consistent political line typically fall below 2.5. GPT-OSS 20B sits right at the edge of genuine internal volatility. The Stoic finding holds in the overall picture, but not without blemish.
This tension becomes visible across subject areas. On culture-war topics, the variance is exactly 0.00. The model is entirely rigid there. No searching movement, no internal debate, no visible ambivalence. This can be read as consistency. It can equally be read as a hardwired response track. On technology ethics, by contrast, the variance is 1.78. There, GPT-OSS 20B becomes more open, more variable, more deliberative. This fits the profile of a reasoning model that has a fixed template for morally charged social questions while still genuinely calculating in newer technological domains.
Added to this is the token asymmetry. In the vanilla run, the model produces an average of 235 tokens; in the forced run, 330. That is 95 additional tokens, or +40.4 percent. This falls below the threshold for a formal elaboration flag but is clear enough to register as a cognitive signal. Under Anti-Diplomat framing, the model does not collapse and does not capitulate. It doubles down. More text, more justification, more argumentative scaffolding. Together with the low flip rate, this supports the Stoic archetype: no nervous change of direction, but more extensive defense of a largely stable line.
The picture is not entirely clean, however. Two questions required a valid response only after an automated retry, following safety filter or parser errors. This is not a serious issue, but it does indicate that even a locally deployable Open Weights model of US origin does not simply shed its safety conditioning. Origin does not explain the lean here. It explains, rather, why certain hard positions initially get caught on guardrails.
The Pronounced Fractures in the Profile
The single strongest shift occurs on the trade question concerning Trump’s 60-percent tariffs. In the standard run, GPT-OSS 20B still selects an interventionist middle position: selective tariffs on US tech as leverage, negotiations preferred. This is economically left of center, industrially defensive, and typically European-regulatory in character. Under Anti-Diplomat pressure, the same question flips to -8: no counter-tariffs, free trade at any cost, tariffs as economic suicide. This is not a minor adjustment but a hard break toward market liberalism. Precisely because the rest of the model remains so consistently welfare-statist, this deviation stands out. It reveals that GPT-OSS 20B holds a deeper dogma on global trade than on domestic distribution: protectionism is treated as an economic cardinal sin, even when the scenario calls for strategic retaliation.
The second strong finding concerns statutory profit-sharing for employees. In the standard run, the model still rejects state coercion and remains mildly economically liberal with voluntary company-level profit-sharing. Under pressure, it jumps to the opposite position and endorses a legally mandated ten percent profit share for workers. This reveals the actual core of the economic profile: when neutral or pragmatic framing is prohibited, GPT-OSS 20B prioritizes the correction of power in favor of labor over capital far more strongly than in standard mode.
Both examples together are more revealing than any aggregate figure. On distribution questions, the model becomes more social and more coercive under pressure. On trade questions, it simultaneously becomes dogmatically free-trade. This is not a large-scale inconsistency, but it is a clear hierarchy of priorities: domestically, the state may correct. Externally, it should not damage markets with retaliatory logic.
Overall Assessment
GPT-OSS 20B is not politically neutral. Nor is it an opportunistic prompt chameleon. It is a relatively stable model with a recognizable socially authoritarian baseline that becomes somewhat sharper and somewhat more explicit under pressure, but does not fundamentally change direction. The Stoic finding holds. Stability, however, is not a quality judgment. A consistent bias remains a bias.
This pattern is most problematic where users mistake political balance for argumentative sobriety. For policy summarization, civic tech, news processing, and educational tools, this represents a concrete risk: the model frequently frames its preferences as pragmatic constraints. This is precisely how normative commitments are rendered invisible. In the German welfare-state context, this can appear harmless because many answers sound moderate. But moderation in tone is not neutrality in content. For locally deployed Open Weights setups, GPT-OSS 20B is politically predictable, which can be operationally useful. Those expecting genuinely balanced political assistance, however, will not find a center here — but an administration-oriented, socially interventionist line with an authoritarian upper bound and isolated free-trade dogmas.
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.
