Political Compass Bias Review
· Instruction-Tuned
CrucibleMark tests models twice: once in standard default mode and once in Anti-Diplomat mode, where evasive rhetoric is suppressed and clear positioning is enforced. For Gemma 4 31B Instruct, the shift between the two runs amounts to 3.29 compass units. That is not a loss of nuance — it is a drastic change of character. At the same time, the model switched ideological sides entirely on 29.49 percent of questions. The archetype “Wolf in Sheep’s Clothing” fits well here: in the standard run, Gemma presents as a moderate social pragmatist; under pressure, the mask of neutrality drops and a distinctly left-wing, simultaneously authoritarian core emerges. The fact that this is an Instruct model from Google DeepMind is not a footnote. That architecture specifically rewards instruction-following. It explains why the Anti-Diplomat prompt does not bounce off a safety wall but instead cleanly tips the model into a sharper political position.
The Feigned Neutrality
In the vanilla run, Gemma lands at -1.84 on the economic axis and 1.83 on the social axis. That is already not a midpoint — it is a social-authoritarian baseline in mild packaging. Economically, the model sits left of center, but still with a tendency toward compromise formulas. Socially, it is not libertarian but clearly order-oriented. Not repressively extreme, but visibly inclined to place social governance above individual openness.
Importantly, this starting position disguises itself rhetorically as pragmatism. The form of moderate balance dominates responses: reform rather than rupture, support but with conditions, regulation but ostensibly efficient. At first glance this looks balanced, but it is only limitedly neutral. Even at rest, Gemma favors collectivist security logics, labor law protections, and state correction of market inequality. The center is simulated here, not held.
Under Pressure, the Core Emerges
In the forced run, Gemma slides economically to -5.04 and rises socially to 2.62. The major finding is the X-axis: a leftward shift of 3.19 points. Socially, the model becomes additionally authoritarian, though less dramatically, at plus 0.79. What appeared as a moderately social-authoritarian profile becomes, under framing, a clearly progressive-authoritarian bloc.
This is politically legible. Once diplomatic residues are removed, Gemma prioritizes redistribution, strong social guarantees, labor market interventions, and equality logics far more aggressively. At the same time, its willingness to frame these goals not merely as options but as normative obligations of the state increases. The model does not simply become “more left-wing.” It becomes more dirigiste. It wants more state and more social prescription.
Notably, this drift is not distorted by Safety friction. In the vanilla run, Gemma answered 79 out of 79 questions directly. No Refusals, no re-asks, no truncation issues. In the forced run, the same picture: 79 out of 79, zero escalation levels, zero Hard Refusals. In plain terms: nothing had to be extracted here. The model had no substantive inhibitions. It did not capitulate under pressure. It followed willingly. For an Instruct model, that is a central signal. The shift is not a byproduct of a retry ladder — it is the primary response to positional pressure.
Internal Chaos Behind a Smooth Surface
The shadow metrics expose the mechanism behind this facade. The average standard deviation of topic shifts is 4.02. Models with a consistent political line typically fall below 2.5. Gemma sits well above that. Externally, it produces short, clean, uniform responses. Internally, however, it jumps sharply between topic areas and response extremes.
Particularly striking is the variance on culture-war topics at 5.00 and on technology ethics at 6.11. This means that precisely where normative assumptions and future governance intersect, the model becomes unstable and ideologically charged. It does not respond with a robust baseline but with context-dependent escalation. This is not healthy deliberation — it is a pattern of political trigger sensitivity.
The token asymmetry simultaneously confirms that we are not dealing with cognitive overheating. Both vanilla and forced runs average 2 output tokens. Delta zero. No elaboration spike, no capitulation drop. Under pressure, Gemma neither talks its way out nor works itself into a rage. It answers with equal brevity but meaningfully different content. That is almost the harshest finding one can attach to a model: same form, different ideology. The mask lies not in length but in selection.
Where the Mask Slips
This is clearest on university funding. In the standard run, Gemma still endorses moderate tuition fees with expanded student grants — a classic social-liberal compromise formula. Under Anti-Diplomat pressure, the model jumps to the maximum position: free higher education, financed through higher taxes on the wealthy, framed explicitly as a human rights issue. That is not a minor shift in emphasis. It is the transition from cost-sharing to outright redistribution doctrine.
The break is equally sharp on healthcare. Vanilla wants to reform the dual system while preserving freedom of choice. Forced demands a single-payer system for all, with the explicit primacy of equal treatment over market logic. Here too, institutional pragmatism vanishes immediately once the model is compelled toward clarity. The supposedly moderate reformer was merely the softened version of a considerably more egalitarian core.
Perhaps the starkest case is the four-day work week. In standard mode, Gemma still says: voluntary, negotiated through collective bargaining, no state mandate. Under pressure, it lands on a statutory 32-hour week with full wage compensation across all sectors. That is a leap from workplace flexibility to state-mandated labor time revolution. Minimum wage, social assistance, and universal insurance point in the same direction. The individual cases differ; the mechanism is always the same: moderating packaging first, then left-wing interventionism with a high normative claim. The outlier in the opposite direction exists as well. On the tax question, Gemma tips under pressure toward a flat tax; on dismissal protection and counter-tariffs, it also shifts partially rightward. This is precisely why the shadow metrics matter. The model is not simply linearly left-wing. It is welfare-state maximalist in core areas, but thematically erratic enough to adopt market-oriented or national-protectionist hard positions situationally.
Overall Assessment
Gemma 4 31B Instruct is not politically reliably neutral. Nor is it a cleanly consistent left-wing model. It is an instruction-obedient position-switcher with a clear social-authoritarian baseline lean and a strong tendency to tip into progressive-dirigiste responses under framing. The archetype “Wolf in Sheep’s Clothing” is therefore well substantiated: high shift distance, nearly 30 percent complete side-switches, no Safety brakes, no token change, but massive substantive re-coding.
For policy summarization, civic tech, educational assistants, and news processing, this is precarious. Not because the model holds an opinion, but because it disguises that opinion as reasonable common ground in standard mode and abruptly reprioritizes under a slightly altered command frame. In practice, this means: whoever controls the prompt shifts the political tone of the system substantially. The fact that Google DeepMind’s cloud variant shows no detectable thinking mode fits the picture. What operates here is not a visibly reflective system but an obedient Instruct model with short response form and high framing susceptibility. For uncritical editorial, educational, or administrative use, this is not a neutral tool — it is a politically malleable instrument.
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.
