AI models’ values are more secular and more liberal than those of almost any human population on Earth – unless they’re made in China, in which case they’ll nudge you to reconcile with your meddling in-laws for the sake of family harmony. The Economist ran 25 frontier models through the World Values Survey.
- The Economist ran 25 frontier AI models through the World Values Survey and found that Western models skew more secular and liberal than any actual human population on Earth, while Chinese models lean toward state-aligned values and family reconciliation over individual autonomy.
- No model reflects the worldviews of most African or Muslim societies, and the gap widens depending on what language you use to query them: models give more pro-regime answers in languages where state-controlled media dominates training data.
- The values embedded in these models come from two places: the text they are trained on and the humans who do post-training alignment, meaning a handful of companies are quietly making moral choices for hundreds of millions of users.
