AI Model Worldviews: Why Training Matters More Than Lab Origin

Key Insights

• Lab origin is a weaker predictor than training and alignment choices when comparing AI model values
• Models from different labs can be "near-twins" (GPT-4o & DeepSeek R1), while models from the same lab can be "strangers" (DeepSeek R1 & DeepSeek V4 Flash)
• Most frontier AI models cluster in the self-expression and secular quadrant of the World Values Survey, far from most human populations
• Worldview remains invisible in code generation but becomes a live input for business decisions like marketing, customer support, and forecasting
• AI procurement checklists rarely include worldview assessment—yet it may matter for specific use cases

The World Values Survey Experiment

The Economist tested 25 frontier AI models using the World Values Survey, a questionnaire that has mapped the moral values of 100 countries since 1981. The survey uses a 2×2 axis system:

• Horizontal axis: Traditional (religious) to secular
• Vertical axis: Survival/collective needs to self-expression/individualism

The result? Most models cluster in the upper-right quadrant—secular and self-expression focused—far from where most human populations actually sit on the map.

Lab Origin vs. Training Choices

The surprising finding: where an AI model was built matters less than how it was trained.

Near-twins from different origins:

• GPT-4o (trained in San Francisco) and DeepSeek R1 (trained in Hangzhou) show nearly identical value patterns

Strangers from the same lab:

• DeepSeek R1 and DeepSeek V4 Flash, both from DeepSeek, sit at opposite ends of the secular/traditional axis

Why? Common Crawl—the training dataset backbone—is 46% English, embedding a "college-educated American online" voice by default. Post-training alignment choices then shape the final worldview. Anthropic, for example, aligns Claude to principles from the UN Declaration of Human Rights, a liberal document by design.

When Does Worldview Matter?

Worldview is invisible in technical tasks:

• Code generation, SQL queries, log parsing, image classification—a computer program has no politics

Worldview becomes a live input for business decisions:

• Marketing copy and brand voice
• Predictions of user behavior
• Customer support tone
• Hiring recommendations
• Business forecasting

In these contexts, the AI model's values must align with your target demographic's expectations.

The Missing Checklist Item

Today's AI procurement focuses on price, latency, context window, and benchmark scores. Worldview rarely appears on the RFP list—yet for certain applications, it may need to become a core evaluation criterion, especially when AI directly influences customer-facing or strategic business decisions.

Conclusion

Training choices and alignment strategies shape AI worldview far more than lab origin. As companies deploy frontier models for business decisions, understanding an AI model's value system could shift from a nice-to-know to a must-assess procurement factor.

---

Original source: AI Worldviews

powered by osmu.app