Designing AI systems whose goals and behavior remain reliably beneficial to humans, even as capabilities scale.