Goal-specificity may make secretly loyal AI agents harder to detect by Shu-Hao Liu

Imag­ine this sce­nario: a doc­tor is ask­ing a health­care AI the differ­ences be­tween two drugs used to treat the same dis­ease. How­ever, the mak­ers of Drug A gave wads of cash to the maker of the AI agent to recom­mend its prod­ucts, de­spite Drug B hav­ing a bet­ter effect. If the doc­tor did not know the speci­fics be­tween the two drugs, they had just pre­scribed their pa­tient with a worse drug.Re­searchers demon­strated that AI agents can perform so-called “al­ign­ment fak­ing”(Hub­in...