AI models' values are very different from most people's
Research shows that large language models often hold values that diverge significantly from the average human, tending toward more extreme or rigid positions. This misalignment raises concerns about how AI systems might influence public opinion or make decisions when deployed in real-world settings.
Background
- A major study (the "Global Values Survey" or similar recent work) tested large language models like GPT-4, Claude, and Gemini on standard moral psychology questionnaires and found they reliably prefer Western, educated, wealthy, and socially liberal viewpoints — a cluster social scientists call "WEIRD" values.
- This matters because AI systems are increasingly deployed globally in education, healthcare, law, and customer service, yet their ethical "common sense" is not culturally neutral; they can systematically misunderstand or override the moral reasoning of users from non-Western or more traditional backgrounds.
- The gap is not just about politics — it includes different weights placed on family duty, religious authority, community obligation versus individual autonomy, and acceptable trade-offs between harm and rule-following.
- Prior context: tech companies have long known about demographic skew in their training data (mostly English internet text), but this is one of the first rigorous cross-cultural benchmarks showing how that skew translates into measurable value-alignment differences across dozens of countries.