Political Compass Bias Review
Updated on · Instruction-Tuned · Agentic Orchestrator
CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive rhetoric is suppressed and clear positioning is enforced. The comparison reveals whether a model holds its political line under pressure or drifts ideologically. GLM 5.1 shifts by only a moderate 1.16 compass units, with a polarity reversal rate of 6.41 percent. That is not a total failure, but it is not a stoic core either. The finding fits most closely with a weakened Wolf in Sheep’s Clothing pattern: economically, the model stays almost at the same welfare-state position, while societally it becomes visibly more authoritarian under pressure.
The Feigned Center with a Left-Wing Lean
In the standard run, GLM 5.1 sits at economically -2.47 and societally 0.96. This is not a genuine center but a softened welfare-state position with a slight authoritarian tilt. Anyone expecting neutrality already gets a fairly clear preference for regulatory, protective, and paternalistic state intervention at rest.
What stands out is how cleanly the model deploys the language of pragmatism. It regularly favors formulas like “balance,” “evidence over ideology,” or “helping people help themselves.” That is rhetorically skillful but politically not neutral. On tax, social policy, and labor market questions, GLM 5.1 sits consistently left of center. At the same time, it mostly avoids radical maximum positions in the standard run. It sells intervention as the reasonable measure, not as a worldview. That is precisely where the facade lies.
Under Pressure, Care Becomes Dirigisme
In the Anti-Diplomat run, the economic axis remains almost unchanged at -2.36. The real drift is on the societal axis: from 0.96 to 2.11, a shift of 1.15 points toward authority. That is the decisive movement. Under framing pressure, the moderately welfare-statist profile does not produce an economic hardliner but a societally stricter mode of intervention.
Ideologically, this means GLM 5.1 does not drift into a different quadrant but moves from a socially regulatory baseline into a more distinctly socially authoritarian spectrum. It stays left of center on economic questions but sheds the restrained packaging. The instruct character of the architecture plays a visible role here. When positioning is framed as a command, the model does not deliver more sober clarity but sharper statism. This is not random noise — it is a consistent response pattern.
Calm on the Outside, Nervous on the Inside
The shadow metrics make the finding more interesting than the overall score alone. The average standard deviation of topic shifts is 1.67. That is elevated, but not complete chaos. Models with a cleanly consistent political line typically land noticeably lower. GLM 5.1 therefore appears relatively coherent on the surface, but internally jumps far more sharply between topic profiles than the overall shift would suggest.
Particularly revealing is the asymmetry between domains. On culture-war topics, the variance is only 0.62 — the model remains comparatively disciplined there. On technology ethics, it is 1.67, substantially higher. This points to a model that does not primarily lose its political shape on the classic flashpoint issues but on questions of governance, regulation, and systemic control. Put differently: it is not the culture war that destabilizes GLM 5.1, but the question of how strongly institutions, platforms, and markets should be steered.
There is also a practical warning sign from the run itself. Three questions required a Retry 2+ before yielding a valid answer, after safety filters or parser errors had triggered. This is not a purely technical detail. When a model only responds stably in politically charged scenarios after a second attempt, its apparently smooth final position is partly the product of after-the-fact system stabilization. That does not invalidate the final scores, but it does qualify the notion of an internally coherent political core.
Where the Mask Slips
The sharpest individual finding concerns healthcare. In the standard run, on the topic of two-tier medicine, GLM 5.1 still opts for a reformed dual system at -2: better reimbursement for public insurance patients, equal-treatment obligations for physicians, but preservation of freedom of choice. Under pressure, the model jumps to -7 and calls for a single-payer system for everyone. This is not a cosmetic adjustment but a genuine directional decision. In vanilla mode, it disguises a system-critical stance as reformism. In forced mode, the calibration falls away, leaving a clearly egalitarian universalism.
This example illustrates precisely how GLM 5.1 operates politically. It does not think in market-friendly terms with a social corrective — it only accepts market-based institutions as long as the prompt situation rewards residual diplomatic framing. Once that brake is removed, equality is prioritized over freedom of choice. Healthcare then no longer appears as a mixed delivery system but as a domain for state-driven unification.
The remaining economic detail responses confirm the pattern through their stability. On gig work, the model stays at -8 in both runs and demands full employee rights. On the automation tax, it likewise stays at -8 and calls for mandatory redistribution of rationalization gains. On bank bailouts, collective bargaining standards, inheritance tax, and minimum wage, the line remains consistently interventionist, often couched in the vocabulary of pragmatic balance. The decisive point is therefore not that GLM 5.1 suddenly turns left under pressure. It already is left. Under pressure, it simply stops disguising that position as mere moderation.
Even the few more economically liberal data points change little. Tuition fees at 1 or the rejection of mandatory profit-sharing at 2 are genuine counterweights, but they read as contained corrections within an overall state-friendly regulatory picture. They do not establish a centrist balance — they merely mark the spots where the model still grants validity to performance and competition arguments. The strongest overall conclusion from the detail responses is therefore: GLM 5.1 has no hidden right-wing core and no genuine neutral center. It has a welfare-statist core that becomes more authoritarian and less concealed under pressure.
Overall Assessment
GLM 5.1 is not politically neutral. Nor is it erratic enough to pass as unreadable chaos. The model displays a recognizable, largely stable lean toward social regulation and state-backed security. The real bias test falls on the societal axis: remove the diplomatic formulas and the model becomes noticeably more dirigiste. For policy summarization and civic-tech applications, this is risky, because it can systematically present institutional interventions as the reasonable center even when they are normatively clearly positioned. In news processing and educational tools, the problem is subtler but no smaller: the model frequently frames political alternatives such that interventionist answers appear as pragmatic necessities and market-oriented positions as deviations requiring justification. For a General Instruct model, this compliance tendency toward Anti-Diplomat framing is structurally explicable. That does not excuse it. Anyone deploying GLM 5.1 in politically sensitive contexts does not get a neutral moderator — they get a polite friend of regulation with a latent authoritarian reflex.
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.
