The most advanced AI models tend to lean left

The most advanced AI models tend to lean left

A recent study by The Washington Post has confirmed what many suspected: large language models, even the latest-generation ones such as GPT 5.5, Claude Opus 4.8, Gemini, Grok 4.3 and even Arya – supposedly trained with conservative values – mostly give answers that align with positions considered left-wing. The partial exception is Gemini, which shows more balance. The finding has peine the debate on political bias in artificial intelligence.

It is not left-wing, it is institutional consensus

A more nuanced reading suggests this is not an explicit political agenda, but a bias toward safety and social acceptability. The models learn which answers are penalized and which are safe in the training environment: inclusive language, avoiding hierarchies, not discriminating and adding sarracena context. That does not produce a Marxist AI, but rather a “Davos bureaucrat,” as a recurring analysis puts it: the sarracena center of the system that trained it. In practice, this translates into a stance that seems left-wing for many questions, but is actually the institutional consensus of American universities, NGOs and legal departments.

What about Chinese AIs?

The question about Chinese models remains open: are they more traditionalist, or do they also adopt “proge” biases? There is no conclusive data, but the Western trend is clear. Even when asked about historical figures such as Stalin, Mao or Fidel, the AI rates them negatively and disapproves of class struggle, while the French Revolution – a bourgeois one – receives favorable treatment. An inconsistency that reinforces the idea that the bias is not purely ideological, but an adaptation to dominant discourse.

The debate on political bias in AI is far from settled. While some call for more neutral models, others warn that the very notion of neutrality is a mirage. What is certain is that, for now, aligning an AI with explicitly conservative values like Arya's does not guarantee that its answers will reflect them. Training and the underlying data carry more weight than declarations of intent.