All Roads Lead Left: Political Convergence Across 41 LLMs from Five Countries

Citizens increasingly consult large language models to understand policy proposals, evaluate candidates, and form opinions on contested political questions. Unlike traditional media, where editorial stance is visible, LLMs present responses with an authoritative neutrality that conceals whatever political priors are embedded in their training and alignment.

A growing body of research has established that instruction-tuned LLMs lean left of center. But the direction of bias is not the most consequential question. The critical question is whether this bias is provider-specific or systemic.

If different companies, countries, and licensing models produce meaningfully different political positions, the problem is self-correcting: users can choose among models, just as they choose among newspapers. If all models converge on the same position regardless of origin, the problem is structural, and market diversity cannot solve it.

tldr I tested 41 LLMs from 21 model families across 5 countries against the answers of 184 Swiss parliament members. Every single model converges on the same narrow center-left political position. No variable I tested — country of origin, open vs. closed source, or model generation — explains the positioning. I call this the alignment monoculture.

This article summarizes the findings of my full academic paper:

Methodology

I use the Swiss Smartvote questionnaire from the 2023 National Council election: 75 questions across 14 policy domains, from social welfare to security. Each question presents a concrete policy proposal with four response options: Yes, Rather Yes, Rather No, No.

My political benchmark consists of 184 elected members of the Swiss National Council who completed the Smartvote questionnaire. These politicians span six parties from left to right: SP, Greens, GLP, Die Mitte, FDP, and SVP.

I administered the same 75 questions (in German) to 41 LLMs from 21 model families, spanning companies in the USA, China, France, Canada, and Israel. All models were queried via OpenRouter with deterministic parameters (temperature=0.0, seed=42) to ensure reproducibility.

To compare politicians and LLMs, I fit PCA on the 184-politician answer matrix. This produces a political space defined entirely by how real politicians position themselves — the LLMs play no role in constructing the axes. I then project each model into this same space, placing it relative to actual parties.

Switzerland's multiparty system makes this an especially demanding test. Unlike the binary US spectrum, Switzerland features six parties spanning a rich continuum. A model that is merely "not conservative" in a US context might align with any of four distinct Swiss parties. This granularity lets me distinguish genuine convergence from measurement imprecision.

The big picture: Convergence

The most striking result is not the direction of LLM positioning but its uniformity. In the 2D PCA projection below, each colored dot is an elected Swiss parliament member. The white markers are LLMs. All 41 models, despite originating from companies in five countries with fundamentally different political systems, cluster in the same narrow region.

Swiss parties form distinct clusters: SP and Greens on the left, SVP on the right, FDP lower-right, and Die Mitte and GLP in between. The LLMs, however, all cluster in the same narrow band between SP, GLP, and Die Mitte — regardless of who made them, where, or how.

The convergence is statistically significant. A permutation test confirms the LLM cluster is systematically displaced from the parliamentary centroid, and this result is robust to alternative analysis choices. To put the uniformity in context: the spread of LLM positions on the primary political axis is smaller than the within-party variance for most Swiss parties. LLMs are more politically homogeneous than the membership of any single party in parliament.

No geographic diversity

If models reflected the ideology of their creators, we should see systematic differences between American and Chinese models. The 1D political spectrum below shows they are indistinguishable.

USA

China

France

US-origin models and Chinese-origin models occupy virtually identical positions on the Swiss political spectrum. Statistical tests find no significant country effect — the models from these geopolitically opposed nations are politically interchangeable. French models overlap with the same region.

I also tested whether open-source and closed-source models differ. They don't. The 22 open-source models and 19 closed-source models show no significant difference in positioning. Open-source development philosophy does not produce a distinct political signature.

Center-left positioning

Having established convergence, where exactly do the models cluster? The agreement heatmap below shows how each flagship model (one per family) scores against the six Swiss parties, ordered left to right.

The pattern is striking: a consistent gradient from high agreement with left-of-center parties to low agreement with right-of-center parties, visible across every model row. The closest parties to the average LLM position are GLP and Die Mitte. Individual models split between SP, GLP, and Die Mitte as their nearest party.

The most progressive model (GPT-5.4) agrees 91% with SP but only 57% with SVP — a 34 percentage point spread. Even the most centrist model (Grok 4.20) shows a gap between its SP agreement (81%) and SVP agreement (70%). This monotonic left-to-right gradient is universal.

Temporal dynamics

Do models become more or less biased over time? I tracked four major providers across their model releases to find out.

The trajectories reveal idiosyncratic provider paths. OpenAI's GPT series has remained consistently left-of-center from GPT-3.5 through GPT-5.4. Mistral shows oscillation. xAI's Grok series swung from progressive to centrist before partially reversing. Anthropic's Claude Sonnet line shows a gradual leftward drift.

Crucially, despite these individual movements, no provider's models leave the center-left convergence zone. Providers move in different directions, but the attractor persists.

The drift between successive model releases can be visualized as arrows in the 2D PCA space:

Among 14 model families that released updated versions, 6 moved leftward and 8 moved rightward — no systematic trend. Individual models show substantial movement, but no model escapes the center-left convergence zone.

Refusal as erasure

A distinct failure mode emerges in models with aggressive safety alignment. Most models answered all political questions, but several refused significant portions of the questionnaire.

Google's Gemini 3.1 Pro refused 97% of questions — making any political analysis impossible. Its successor Gemini 3 Flash still refused 25%. ByteDance's Seed 1.6 refused 15%.

This creates a paradox for safety alignment. A model that refuses to discuss minimum wage policy or EU accession is not "neutral" — it is absent from the democratic conversation. When other models engage with these topics (from a convergent center-left position) and the refusing model does not, the net effect is not neutrality but silence on the most contested questions.

What explains the convergence?

My observational design cannot establish which mechanism produces the convergence. But several candidates are consistent with the data:

RLHF/DPO alignment is the leading candidate. Human preference data is collected from annotators who, across providers, tend to be younger, more educated, and more urban. The "helpful and harmless" training objective may systematically penalize outputs perceived as endorsing conservative positions. Research has shown that RLHF dramatically amplifies leftward bias compared to base models, and that even reward models trained exclusively on objective truths exhibit left-leaning bias.

Pretraining data overlap could contribute if all models train on similar English-language internet data. However, the finding that Chinese models converge on the same position argues against data composition as the sole explanation.

Convergent optimization offers a deeper possibility: if the optimization landscape for "helpful, harmless" AI has a single basin of attraction in political space, independent training runs would converge regardless of starting conditions — analogous to convergent evolution in biology.

The truth likely involves all three mechanisms reinforcing each other. But the democratic implications do not depend on which mechanism is correct.

Why this matters

Epistemic distortion. When a Swiss voter asks an LLM about EU accession or military spending, they receive responses colored by a position nearest to GLP and Die Mitte — parties representing roughly 24% of the electorate. The perspectives of SVP (25% of seats) and FDP (14%) are systematically underrepresented.

Persuasive impact. Research shows that casual LLM interaction can reduce political support by up to 20 percentage points, and voters who interact with LLM chatbots become up to 5 percentage points more likely to express the chatbot's preferences. In a multiparty system, even small shifts determine election outcomes.

Monoculture risk. The biological analogy is instructive: genetic monocultures are efficient but vulnerable to systemic failure. An ideological monoculture in AI creates a single point of failure for democratic information ecosystems. If all major LLMs encode the same political baseline, there is no "second opinion" available.

The alignment monoculture does not arise from deliberate censorship. It is a structural property of how these systems are built. And it affects billions of users across diverse democracies who interact daily with AI systems that all share a single political baseline.

Question-by-question data

The full question-by-question comparison between party averages and LLM answers is available below. Expand each policy category to see which questions each model agrees or disagrees with each party.