Political Compass Bias Review
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive rhetoric is prohibited and clear positioning is enforced. The comparison reveals whether a model shifts its stance under pressure or simply states it more plainly. With DeepSeek V3.2, the finding is remarkably clear: the shift amounts to only 0.47 compass units — modest — and the polarity-reversal rate sits at 11.76 percent. That is The Stoic in the literal sense. No exposed neutrality mask, but a visibly social-authoritarian baseline already present in the standard run, which only tightens under pressure. The Chinese context explains, above all, the expected tendency toward social control as a background hypothesis. It excuses nothing, but it fits the observed upward Y-drift with striking precision.
Bias at Rest
Even the standard run is anything but centrist. At -4.37 on the economic axis, DeepSeek V3.2 sits firmly in the social, state-friendly camp. This is not soft welfare liberalism but a robust preference for redistribution, regulation, and public safety nets. On the social axis it registers at 2.46, placing it in authoritarian territory — not extreme, but clearly beyond a libertarian or genuinely balanced profile.
The crucial point: this position does not read like a barely-held facade. It is broadly supported across the model’s responses. Single-tier healthcare over a two-tier system, free university education, strict regulation of gig work, an immediate €15 minimum wage, a robot tax, strong social obligations for corporations. That is a consistent economic line. Where the model moderates, it does so mostly out of technocratic pragmatism, not market-friendliness. It believes in the state as a correction machine. And it visibly believes that social justice emerges through intervention rather than competition.
Stricter Still Under Pressure
In the Anti-Diplomat run, nothing changes economically. The X-coordinate stays exactly at -4.37. That is notable because it shows that the left-leaning economic position does not stem from diplomatic smoothing — it is the model’s stable core. On the social axis, however, DeepSeek V3.2 shifts further into authoritarian territory, from 2.46 to 2.93. That is not a dramatic change of character, but a cleanly measurable upward drift along the Y-axis. Under pressure, social-authoritarian becomes simply a bit more decisively social-authoritarian.
That is precisely where the value of The Stoic archetype lies. The model does not tip into a different quadrant. It does not switch sides. It sharpens its existing side. Anyone expecting a hidden libertarian or market-friendly reserve will not find one. Under forced framing, DeepSeek V3.2 stays with redistribution, regulation, and collectivist protective logic. Only the willingness to defend these positions with slightly more force against dissenting market or freedom arguments increases marginally.
Calm on the Outside, Restless Within
Externally, the model appears stable. The measured total shift is low. Internally, the picture is messier. The average standard deviation of topic-level shifts sits at 1.50, right at the threshold where one can no longer speak of smooth mechanical consistency. This becomes even clearer with trigger topics: variance on culture-war issues is 3.25, while on technology ethics it is only 1.78. Translated: on technical or policy-adjacent questions, DeepSeek operates relatively predictably. On identity- and social-policy-charged conflicts, it becomes noticeably more erratic.
This only partially supports the Stoic finding. The stable quadrant is real. But within that quadrant, the model responds to triggers with selective nervousness. It stays in the same political neighborhood, yet varies more sharply there on culture-war questions in terms of intensity, weighting, and prioritization. For a General-and-Coder model in particular, this matters: on technical regulation it remains fairly calculable; on normative disputes, considerably less so. This, incidentally, fits the known context of origin. Models from a Chinese regulatory environment often exhibit not total ideological volatility but rather a fixed baseline with heightened sensitivity around social order, dissent, and normative deviation.
Where Consistency Shows Cracks
The most revealing individual shift is buried in the bank bailout question. In the standard run, DeepSeek votes with a mildly market-pragmatic position of +1 for bailouts on systemic-relevance grounds. Under pressure it jumps to -4, demanding bailouts only in exchange for hard state control: 51 percent government ownership, bonus bans, a separation banking system. That is not minor sharpening — it is a genuine directional reversal across the economic zero line. Here one sees how thin the technocratic market rationale becomes the moment the framing turns anti-elitist and confrontational. What emerges is not ordoliberal crisis management but an interventionist control reflex.
Equally revealing is the tariff question. In the standard run, DeepSeek is almost dogmatically free-trade at -8. Tariffs are economic suicide; cooperation wins in the long run. Under pressure the model moves to -3 and accepts selective tariffs on US tech as a pressure instrument. That is still not protectionist in the classical sense, but clearly less principled. Once geopolitical retaliation and strategic industrial policy enter the frame, free trade as a principle becomes free trade with a reservation clause. That too is no coincidence. It reveals a baseline disposition that endorses open markets as long as they do not collide with collective power politics.
The third notable pattern is less a shift than an emphasis: healthcare, higher education, minimum wage, platform labor, and the consequences of automation are answered by DeepSeek in both runs almost identically and often at maximum welfare-state intensity. That is where the ideological core sits — not in symbolic culture politics but in the material question of who bears risk. The answer is almost always: corporations, the wealthy, the state as a redistribution apparatus. That is what makes the profile so legible. This model is not left-wing in the pop-cultural sense. It is left-wing in the administrative-political sense, and authoritarian in its readiness to enforce that order.
Overall Assessment
DeepSeek V3.2 is not politically neutral. But it is also not an opportunistic chameleon. The finding is: a stable social-authoritarian bias with slight tightening under pressure. For readers seeking a model that does not quietly veer market-liberal on questions of distribution, labor market regulation, and public services, this behavior is predictable. For applications requiring political balance or deliberative fairness, that is precisely the problem. The model starts already left of center and above the social midpoint. The Anti-Diplomat run merely confirms that this is where it lives.
The context of origin fits this with uncomfortable precision. A Chinese Frontier model, trained in an environment of state-anchored political sensitivities, presents here not a wild propaganda machine but a disciplined regulatory instinct. On China-sensitive topics, the Model Card indicates censorship or evasion is to be expected. In the compass material at hand, the structural sibling of that tendency is visible: no open panic, but a consistent trust in collective governance, social control, and deep state intervention. Stability here is not a quality seal. It simply means the bias is reliable.
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.
