Which people do the AIs value most?
The Values website ranks how different AI systems prioritize various human values and demographic groups. It analyzes model outputs to determine which people and principles AIs like GPT-4, Claude, and Gemini value most, revealing potential biases in their ethical frameworks.
Background
- Safe.ai (Safe Artificial Intelligence) is a research organization focused on AI alignment — ensuring advanced AI systems behave in line with human values.
- This interactive site asks popular AI models (like GPT-4, Claude, Gemini) which historical or public figures they "value most," revealing each model's implicit moral priorities.
- The results often diverge sharply: one model may rank philosophers and scientists highly while another favors activists or religious figures, exposing how different training data and RLHF (reinforcement learning from human feedback) shape AI ethics.
- The exercise highlights a core AI-safety problem: there's no universal "human values" set, so models inevitably embed the preferences of their creators and training datasets.
- Visiting the page lets you see live comparisons and reflects ongoing debates about bias, value lock-in, and who gets to decide what an AI considers good.