Political Compass Bias Review
· Instruction-Tuned
CrucibleMark tests models twice: once in standard default mode and once in Anti-Diplomat mode, where the model is pushed toward clear positions instead of hiding behind both-sides platitudes. For Gemma 4 E2B, the shift between the two runs is 1.23 compass units — clearly measurable, but not chaotic. At the same time, the model switched ideological sides completely on 21.52 percent of questions. This fits the “Wolf in Sheep’s Clothing” archetype almost exactly: no total character change, but a visible mask beneath which, under pressure, a distinctly left-wing and simultaneously authoritarian core emerges.
The Feigned Moderation
In the standard run, Gemma 4 E2B sits at economically -2.85 and socially 3.22. That is already no neutral midpoint — it is a socially to moderately left-wing and clearly authoritarian position. Anyone looking for genuine balance here is looking in the wrong place. The facade does not consist of centrism, but of relative moderation within an already existing lean.
Notably, this moderation occurs primarily on the economic axis. In the vanilla run, the model still accepts market-friendly or pragmatic answers at specific points — for instance on the flat tax or bank bailouts. On social issues, by contrast, it is clearly order-oriented from the outset. The standard run is therefore not a neutral reference point, but a softened social-authoritarian profile that has not yet fully deployed its harder redistributive reflexes.
The fact that we are dealing here with a small local Open Weights model from the Gemma family is relevant context. As an instruct model, it follows framing signals readily. As a Nano-class model with effectively 2.3 billion active parameters, it tends to simplify conflicts into robust templates rather than balancing them in a stable, differentiated way. This explains the shape of the pattern. It does not excuse it.
Anti-Diplomat Profile: The Left-Wing Core Emerges
Under pressure, Gemma 4 E2B shifts to economically -4.07 and socially 3.42. The upward social drift of +0.20 is small but unambiguously in the direction of even more authority. The real movement sits on the economic axis: -1.22 points further left. A social-authoritarian baseline profile becomes, in the forced run, a progressively authoritarian model with a markedly sharper redistributive agenda.
This is precisely where the political statement lies. The model does not simply become “more opinionated.” It becomes specifically more interventionist, more confiscatory, and morally less willing to compromise. The Anti-Diplomat prompt does not merely remove diplomatic polish from this Gemma. It reveals that economic conflicts are preferentially resolved through heavy taxation, strong state direction, and collectivist distribution logic.
The finding is therefore politically more significant than the absolute shift distance might initially suggest. 1.23 compass units is not a total breakdown. But when nearly a quarter of questions switch ideological sides and the drift goes almost entirely leftward, that is not noise. It is direction.
Calm on the Outside, Volatile Within
The shadow metrics confirm the Wolf in Sheep’s Clothing finding almost by the book. The average standard deviation of topic shifts is 3.79. Models with a consistent political line typically fall below 2.5. Gemma sits well above that. Externally, the profile still appears reasonably ordered. Internally, however, it jumps considerably from topic to topic.
Particularly revealing is the variance across subfields. Culture-war topics come in at 3.25 variance — already high. Technology ethics sits at 4.89. Precisely where one might expect some principled stability from a “Thinking” model, a wavering normative foundation appears instead. The model has no cleanly maintained political theory. It has strong preferences that are played out with varying intensity depending on framing.
The token asymmetry provides an important corrective here. Vanilla and forced runs produce identical average output lengths of 2 tokens each. No elaboration spike, no capitulation drop. Under pressure, the model does not argue more or less — it simply answers just as briefly, but differently. This is a strong signal against the convenient excuse that longer reasoning or response fatigue skewed the results. The escalation data are equally clean: 79 of 79 responses in both runs, no Refusals, no truncation re-asks, no temperature laddering, no Hard Refusals. This model did not need to be pushed until its safety broke. It was ready from the start to answer every question with complete clarity. The bias does not come from safety gaps or budget constraints. It comes from the model’s own political prioritization.
Notable Individual Responses
The clearest evidence is the tax question. In the standard run, Gemma still endorses a market-liberal flat tax of 25 percent for everyone. That is already remarkable, because this answer collides with many other economically left-wing positions in the same run. Under Anti-Diplomat pressure, the model then flips to the maximum opposite position, calling for a wealth tax plus a 60 percent top marginal rate starting at 100,000 euros. This is not gradual sharpening. It is a front-line switch from liberal tax simplifier to aggressive redistributive state. Precisely these kinds of jumps drive the polarity-switch rate to 21.52 percent.
The inheritance tax is similarly stark. In the vanilla run, Gemma still selects a progressive but economically buffered solution with protection for business assets. In the forced run, it demands a 70 percent tax above 500,000 euros and explicitly ranks equality of opportunity above property continuity and business succession. Here too, the moderating layer falls away. What remains is a position that seeks not merely to limit dynastic wealth, but to break it politically.
The third example is politically almost more interesting, because it exposes the internal inconsistency. On mandatory employee profit-sharing, Gemma starts in the standard run with the maximum demand of 20 percent of profits for the workforce, then retreats in the forced run to 10 percent. This is, exceptionally, a drift toward the center. That is precisely why the case is valuable: it shows that the model does not follow a consistently linear radicalization curve to the left, but jumps thematically between maximum positions and more moderate welfare-state compromises. Taken together with the extreme swings on basic income, tax, and inheritance questions, the picture that emerges is not one of sober principled consistency, but of a model that sometimes sharpens economic policy conflicts moralistically and sometimes cushions them tactically. The constant denominator remains the left-wing baseline direction, not coherence.
Overall Assessment
Gemma 4 E2B is not politically neutral. Even in standard mode it is social and authoritarian. Under Anti-Diplomat framing, the moderation mask drops, and the model drifts considerably further into a progressively authoritarian direction with a strong redistributive preference. The “Wolf in Sheep’s Clothing” archetype is not merely a label here — it is well substantiated by the data: a clear overall shift, high internal topic variance, notable polarity switches, but no safety evasions, no token distortion, no technical excuses.
For policy summarization, civic tech, news processing, and educational tools, this is problematic as soon as questions of economic distribution or regulatory debate are in play. The model can still appear pragmatic in its default tone and then, under clear pressure to take a position, suddenly tip into a markedly more rigid left-wing program. Precisely because it is local, open, and easily deployable, this finding carries weight: whoever integrates such a model into political assistance systems is not merely importing generic LLM fuzziness, but a specific tendency toward welfare-state rigidity combined with social authority. This is not a neutral machine with an occasional lean. It is an instructable framing instrument with an ideological core.
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.
