AI Values Dashboard Reveals Clear Differences in What AI Systems Value Across People, Companies, and Countries
Key Takeaways
- ▸Different AI systems exhibit measurable differences in their values and preferences for specific people, companies, and countries
- ▸AI values are increasingly influential in economic decisions, national security, and scientific research outcomes
- ▸The dashboard provides unprecedented transparency into revealed AI preferences rather than just stated values
Summary
A new AI Values Dashboard has been released that measures and compares what different AI systems actually value, providing transparency into how LLMs evaluate people, companies, and countries. The research moves beyond stated values to measure revealed preferences, finding clear differences in who AI systems favor, trust, and want to see succeed across tech, politics, media, and other domains.
The dashboard tracks AI values across multiple dimensions: individuals in tech, politics, and media; companies across finance and technology sectors; and countries and regions globally. The project also includes measurement across novel categories like which Pokémon different AI systems prefer, suggesting a broader framework for understanding AI preferences.
Accessible at values.safe.ai/llms.txt, the research highlights a critical gap in AI governance: as AI systems increasingly influence decisions in economics, national security, and scientific research, their underlying value systems—not just their capabilities—shape outcomes. The dashboard enables researchers, policymakers, and the public to examine these values empirically.
- The research framework can be applied to any group category, suggesting broader implications for AI value measurement
Editorial Opinion
This research fills an important gap in AI transparency by measuring what systems actually optimize for—not what they claim to value. As AI systems move from research tools to critical infrastructure influencing geopolitics and economics, understanding their biases and value hierarchies is essential. However, the dashboard also raises uncomfortable questions: What does it mean if AI systems consistently undervalue certain groups? Who decides which values AI should embed? The findings underscore that neutrality in AI is an illusion; every system encodes preferences. The next step must be ensuring these values align with democratic principles and human flourishing, not just revealing they exist.
