Today the developers of an AI largely choose the values it follows, and people's views can diverge. Finding legitimate ways to combine many people's input into collective choices about a model's behaviour would help decide whose values AI systems end up following.
What we know
Models are trained with written principles and human feedback.
What blocks us
People and cultures disagree, and there is no accepted way to combine their views.
Current leads
Public input processes, published constitutions for AI models.
Sub-questions
How can conflicting views from many people be combined into consistent rules for how an AI should behave?
Can the science of voting and group decisions (social choice) tell us how to choose an AI's values?