Report
Whether AI can be trained to be risk-averse to aid negotiations
A researcher says risk-seeking and risk-neutral AI would be hard to negotiate with; their new paper explores whether risk aversion can be trained in.
TLDR
A researcher says risk-seeking and risk-neutral AI would be hard to negotiate with. They cite Thornley and MacAskill’s argument that a misaligned AI could be paid to cooperate rather than rebel, but only if it is risk-averse. Their new paper explores whether AI can be trained to have that disposition.
Combined views
2.5K
4 Sources, first seen ago
