AI models' values are different from most people's
A study testing AI models on moral questions found their values differ significantly from most humans, tending toward libertarian and utilitarian views while often avoiding clear commitments. These discrepancies raise concerns about how AI systems might influence users or make decisions in real-world contexts.
Background
- Large language models (LLMs) like GPT-4, Claude, and Gemini are trained on internet text containing many viewpoints, so they don't naturally hold a single coherent set of values. Their "values" emerge from training data and from fine-tuning by their creators.
- Researchers are now systematically measuring these values by asking models questions from global surveys (like the World Values Survey) and comparing their answers to human populations. The results show leading AI models are more politically liberal, more socially progressive, and more trusting of science than the average person in Western countries — and far from the global median.
- This "values gap" matters because AI is increasingly used in hiring, lending, medicine, and content moderation. A model that subtly favors one worldview could systematically bias decisions without users realizing it.
- The finding raises a contested question: Whose values should AI reflect? Options include the global median, the user's own values, the deploying country's values, or the developer's preferences. Companies and regulators are giving very different answers.