In a Nutshell

The Economist reports the results of an investigation into the moral and political profile of 25 frontier AI models. Based on their responses to established World Values Survey questions, it appears that ChatGPT is essentially more secular than any country on earth, and that Gemini is more attached to personal freedom than any country polled, for example. Chinese open-weight models like DeepSeek and Qwen are more transparent about their ideological line (they simply refuse to contradict official narratives) while the inner alignment of Western closed models is reportedly more opaque. The investigation documents how training data, post-training alignment, and even the language in which a question is posed all shape a model’s values. When asked about wealth, gender identity, or political petitions, models produced answers that ranged from classical liberal to centrist to socially conservative. They never reflected the worldviews of most African or Muslim countries.

Our Take

Our most important concern is that AI is extremely good at shaping individual opinions, and that it it risks becoming a de facto arbiter of value-based decisions. By contrast, the diversity of our societies, and our fundamental freedoms depend on a level of disagreement on morally-charged questions. When an opaque system nudges users toward a single (or mostly consistent) worldview, there is a significant risk of tyranny.

The remedy is not to demand neutral models (impossible) but to invest heavily on decentralization, competition and transparency: open weights where possible, inspectable alignment datasets, auditable reasoning chains, declared moral foundations, and ideally the right of local deployment where the deployer sets the value layer. Until then, every interaction with a frontier model is an act of ideological delegation that users did not consciously authorize.

Read the full story at The Economist: Computational bias.