Buyl LLM Ideology Study
The study analyzed nineteen large language models from multiple countries, testing their responses to politically contested questions. The finding: each LLM reflects systematic ideological patterns corresponding to the cultural and institutional context of its creators. Chinese models differ from American models; American models differ from European models; all of them depart from one another in measurable ways on questions of economic policy, social values, and historical interpretation. The researchers' conclusion was explicitly Mouffean. Rather than pursuing the chimera of ideological neutrality — which the Mouffean framework reveals as a hegemonic operation disguised as procedural fairness — regulatory efforts should focus on preventing LLM monopolies and preserving ideological diversity across AI systems as a feature, not a bug. 'The strong ideological diversity shown across publicly available, powerful LLMs would even be considered healthy under Mouffe's democratic model of pluralistic agonism.'
The Variance Illusion Problem — Contrarian ^ Opus
There is a parallel reading that begins from the actual mechanics of model development rather than its outputs. The measurable ideological variance Buyl documents may reflect less about embedded worldviews than about the superficial layer where corporate risk management meets regulatory compliance. Chinese models deflect certain historical questions not because they encode a coherent ideological position but because their operators face state censorship. American models hedge on social controversies not from philosophical commitments but from litigation exposure and advertiser sensitivities. What appears as ideological diversity may be diversity in corporate fear.
The regulatory prescription that follows from this reading inverts Buyl's conclusion. If the variance is primarily a surface phenomenon—prompt-level deflections rather than reasoning-level commitments—then preserving it preserves theater while the actual concentration of power proceeds underneath. The models share training pipelines, compute infrastructure, data sources, and increasingly, common RLHF methodologies. They differ in their refusals, not their capabilities. A world with ten LLMs that all depend on NVIDIA chips, trained on overlapping web scrapes, fine-tuned by contractors in the same labor markets, but offering different politically correct evasions, is not pluralistic in any substantive sense. It has achieved ideological diversity at precisely the layer that matters least while consolidating control over the infrastructure that matters most.
In the AI Story
The study provides the first substantial empirical grounding for applying agonistic pluralism to the material infrastructure of AI. Before the study, the argument that AI systems were inherently ideological could be defended theoretically but not quantified. The Buyl findings made the ideological variance measurable, defensible, and — critically — regulatorily actionable.
The implications for AI governance are direct. The dominant regulatory discourse has framed the ideological inflection of LLMs as a problem to be solved — a bias to be debiased, a non-neutrality to be made neutral. The Mouffean alternative the study endorses reframes the question entirely. Ideological diversity across models is not a bug but a feature of a healthy AI ecosystem, provided the diversity is transparent and no single ideological position achieves hegemonic dominance through market concentration.
The regulatory prescriptions that follow are substantive. Preventing LLM monopolies becomes not merely an antitrust question but a democratic-pluralism question. Transparency about the ideological positions embedded in training data, fine-tuning choices, and safety systems becomes a democratic right — the precondition for users making informed choices among alternatives. Public investment in LLMs reflecting perspectives the market underserves becomes a legitimate democratic intervention.
The study has been challenged on methodological grounds — how precisely to measure ideology, how to account for prompt-sensitivity in LLM outputs, whether the ideological variance is stable across model updates. These challenges are real but do not undermine the core finding. Even with methodological refinement, the structural claim holds: LLMs reflect the perspectives of their creators, and this reflection is not eliminable through better engineering.
Origin
Published in Nature Machine Intelligence in 2025 by a research team led by Maarten Buyl at Ghent University, with collaborators across European and American institutions. The study emerged from the intersection of machine learning research and political philosophy, representing one of the first substantial attempts to bridge the two fields on questions of AI governance.
Key Ideas
Measurable ideological variance. LLMs from different contexts reflect different political positions in systematic, quantifiable ways.
Neutrality is impossible, and its pursuit is hegemonic. Every attempt to define 'neutral' encodes a specific worldview.
Pluralism as regulatory goal. Diversity across systems beats the chimera of neutrality within systems.
Antitrust as democratic pluralism. Preventing LLM monopolies is a democratic commitment, not merely an economic one.
Appears in the Orange Pill Cycle
Layers of Capture Thesis — Arbitrator ^ Opus
The right synthesis requires distinguishing which layer of the system we're examining. At the output layer—the responses users see—Buyl's finding is empirically robust and the variance is real (90% confidence). Chinese models do systematically differ from American models on contested questions, and this difference is measurable and persistent. The contrarian reading that this is 'merely' corporate risk management doesn't diminish the finding; it specifies its mechanism. The question is whether mechanism matters for the regulatory conclusion.
At the infrastructure layer, the contrarian view carries more weight (70% dominance). The models do share training paradigms, compute dependencies, and increasingly homogeneous data sources. Preserving output-level diversity while infrastructure concentrates creates the illusion of choice—different flavors of the same underlying substrate. Here the regulatory implication inverts: transparency about ideological positions matters less than accountability over compute allocation, data provenance, and the political economy of model development.
The synthetic frame the topic requires is a layers-of-capture thesis. Ideological diversity at the response layer is necessary but insufficient. It prevents the worst-case scenario where a single perspective achieves total dominance, but it doesn't address the consolidation of power over the means of AI production. The Mouffean regulatory goal should be pluralism at every layer: diverse outputs, diverse infrastructure, diverse ownership. The Buyl study demonstrates the first is achievable; it inadvertently reveals how far we are from the second and third.
Further reading
- Maarten Buyl et al., 'Large Language Models Reflect the Ideology of their Creators' (Nature Machine Intelligence, 2025)
- Chantal Mouffe, Agonistics: Thinking the World Politically (Verso, 2013)
- Kate Crawford, Atlas of AI (Yale, 2021)