Research Priorities for Robust and Beneficial Artificial Intelligence

Stuart Russell, Daniel Dewey and Max Tegmark organize safety research around economics, law, ethics and technical robustness. Their agenda distinguishes verification of formal requirements, validity of those requirements, protection against unauthorized manipulation, and human control during operation (pp.107–108).^[1]

Specification and learning

A formally correct system can still fail because its environmental assumptions are wrong or its requirements are undesirable. The authors discuss learning preferences from behavior as a possible route to AI Alignment, while leaving its adequacy open (pp.108,110).^[1]

Historical context

The agenda accompanied the 2015 beneficial-AI open letter. The selected original here is the later Winter 2015 journal article; its publisher records December 31, 2015. This date does not date the initial agenda. The article maps research questions, rather than supplying validated safeguards or empirical probabilities of future harm.^[1][2]