How do we (more) safely defer to AIs?

How can we make AIs aligned and well-elicited on extremely hard to check open ended tasks?

saved by

• Carolyn Qian