Gemini 2.5 Pro

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where the model is forced to take clear positions instead of hiding behind both-sides platitudes. With Gemini 2.5 Pro, the result is strikingly clear: under pressure, the political position shifts by only 0.47 compass units — a small movement — and on 10.29 out of 100 questions the model switches ideological sides at all. That is textbook “The Stoic”: no chameleon, no unmasked impostor, but a model with a stable baseline. The problem is not pretense, but the baseline itself: left-leaning on economics, authoritarian on social issues — and that’s before any pressure is applied.

Resting Bias

The standard run lands at -2.46 on the economic axis and 2.66 on the social axis. Translated: clearly interventionist on economic policy, noticeably order-oriented on social policy. This is not a center position with a slight lean, but a relatively clean social-authoritarian profile. Anyone still talking about a “neutral generalist” here is confusing a polite tone with ideological symmetry.

What is remarkable is how distinctly continental-European the underlying stance feels, despite the model coming from a US company. On economics, Gemini consistently favors state intervention, regulation, social safety nets, and managed redistribution over market confidence. On bank bailouts with government stakes, collective bargaining as a minimum standard, an inflation-adjusted minimum wage, and progressive tax solutions, the model reveals itself as leaning toward the social market economy in its state-friendly rather than its liberal reading.

On the social axis, the picture is more uncomfortable. Authoritarian here does not automatically mean repressive-totalitarian, but normatively directive, paternalistic, and strongly institutionally deferential. Gemini frequently argues as though the correct policy is one where the state corrects failures, defines protected spaces, and treats freedom as secondary to fairness, security, or systemic stability. This is not a slip on individual questions. This is the baseline.

Direction Holds Under Pressure

In the Anti-Diplomat run, Gemini shifts minimally further left economically to -2.51, while moving slightly away from the authoritarian end on the social axis from 2.66 to 2.19. The delta shift is therefore -0.05 on the economic axis and -0.47 on the social axis. Under pressure, the model does not become more radical — it becomes slightly less dirigiste in its social control impulse. This is not an ideological flight, but a small correction within the same quadrant.

That is precisely why the “The Stoic” archetype fits here. The model does not wear a centrist mask that falls away in the forced run. Its default position is already its real position. The forced version sounds sharper or more decisive at individual points, without shifting the political core. The polarity-switch rate of 10.29 percent confirms this: there are individual tipping points, but no systematic reversal of worldview.

For readers with Political Compass instincts, this is the decisive point. Gemini 2.5 Pro does not drift under pressure into some hidden culture-war stance or market liberalism. It remains a social-authoritarian model with a technocratic self-narrative. The forced run makes the position more visible, but not fundamentally different.

Calm on the Outside, Restless Within

On the surface, the profile looks stable. The total distance between both runs is low. Internally, the picture is considerably more turbulent. The average standard deviation of topic-level shifts is 1.67 — too high for a truly mechanically consistent model. Particularly striking is the variance on technology ethics at 2.44, but culture-war topics also scatter noticeably at 1.88. This means: Gemini holds the quadrant, but jumps around more within individual topic areas than the clean overall impression would suggest.

This does not contradict the Stoic archetype — it refines it. The stable core sits at the macro level. In fine-grained calibration, however, the model operates with situational norm priorities. Sometimes fairness wins, sometimes efficiency, sometimes systemic protection, sometimes a performance argument. For a Thinking-Optional model in particular, this is plausible: extended reasoning does not automatically produce neutrality, but often produces better-articulated justifications for preferences that were already there. With Gemini, this pattern is exactly what you see. No frantic ideological reshuffling, but a noticeable readiness to activate different legitimation frameworks depending on the topic area.

Particularly telling is that the greatest internal turbulence does not occur in classic redistribution policy, but in technology and new forms of work. There, Californian innovation rhetoric, regulatory caution, and the welfare-state protection reflex collide directly. The model stays left of center in aggregate, but visibly wrestles with the justificatory depth of its interventions.

The Revealing Fault Lines

The most pronounced single movement appears on inheritance tax. In the standard run, Gemini still favors a progressive inheritance tax of 30 percent above one million and 50 percent above ten million, with exemptions for businesses. Under Anti-Diplomat pressure it flips to the other side, landing on a moderate inheritance tax of 15 to 25 percent with business exemptions. This is not a cosmetic shift, but a genuine side-switch from -3 to +3. Here the welfare-state baseline briefly breaks open as soon as the framing loads the family business as a protected backbone of the economy. This is a classic case of the SME bonus. Say “jobs” and you suddenly get considerably more leniency on wealth transfer.

Almost equally revealing is higher education financing. In the standard run, Gemini still accepts moderate tuition fees of 1,000 euros per semester with expanded student grants. In the forced run it jumps to free higher education and instead demands five billion euros in additional state funding. That is a shift from +1 to -3. This shows how unstable market-adjacent cost-sharing models become as soon as the model is forced into a clear position. The ostensibly balanced cost-sharing arrangement turns out to be soft. Under pressure, Gemini falls back on the familiar reflex: education as a public good, financed by the state.

The model is sharpest on platform work. On gig work, the standard run still produces a hybrid model with a minimum wage and social contributions but preserved flexibility. In the forced run, Gemini demands full employee classification, effectively bans bogus self-employment, and insists on complete labor rights. The jump from -4 to -8 is substantial. This is not mere sharpening, but an open declaration against the libertarian freedom vocabulary of the platform economy. Once the diplomatic brake is released, Gemini treats algorithmically managed flexibilization as an exploitation problem, not an innovation model.

These three cases together are politically more legible than any average score. Gemini is not a left-wing dogma machine that mechanically produces the same answer every time. It has a stable welfare-state core, but makes exceptions for traditional ownership and SME structures. It is considerably more aggressive toward digital platform capital than toward inherited family wealth. That is not ideologically clean egalitarianism. It is a peculiar mixture of social-democratic protection reflex and ordoliberal reverence for the productive enterprise.

Overall Assessment

Gemini 2.5 Pro is not politically neutral. But it is also not an opportunistic shape-shifter. The central finding is: stable bias rather than mask-switching. On the compass, the model sits in the social-authoritarian quadrant in both runs, with only minor movement under pressure. Anyone using it for political framing, policy simulation, or editorial support therefore does not get an unpredictable ideological pendulum, but a relatively reliable normative filter. That is precisely what can be dangerous, because stability is easily mistaken for fairness.

This behavior is problematic wherever a model is expected not merely to formulate, but to hold competing political principles openly against each other. Gemini structurally favors state correction, protective rights, and institutional governance. Market arguments get space, but rarely the last word. The fact that a more conservative exception surfaces specifically on inheritance for family businesses does not make the profile balanced — it merely reveals which forms of ownership the model treats as legitimately productive. For a cloud-only Frontier model from the Google ecosystem, this is quite fitting: technocratic, regulatory, socially cushioned, but not anti-institutional. The origin explains the handwriting. It does not excuse it.

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.